Summary

For a long time, quality in health care was managed by looking backwards: an event happens, it is counted, it is reported, and only then is action taken. Predictive quality aims to reverse this sequence. It estimates a patient’s risk of falling, developing a pressure injury, being readmitted or deteriorating clinically before the event occurs, and turns that estimate into a preventive quality decision.

This chapter is written to take a reader with no prior knowledge of the concept, step by step, to an advanced level. It first introduces the basic building blocks (quality, risk and prediction), and then examines the history of the concept, its definitional families, how it differs from neighbouring concepts and its types. The final sections connect the concept to the present: a map of the 2025 and 2026 literature, the attributes of a good application, the risks involved and the implications for Türkiye.

The evidence base of the chapter is a clean dataset of 257 records, obtained from a search in the Web of Science Core Collection and screened record by record. To describe the current landscape, the 83 records of this dataset published in 2025 and 2026 were analysed separately. Classic sources of the field were also used to trace the roots and methodological foundations of the concept.

Learning Objectives

After completing this chapter, the reader will be able to:

1. Distinguish between the concepts of quality, risk and prediction, and explain that predictive quality stands at the intersection of the three.
2. Summarise the historical development of the concept from Nightingale to the age of artificial intelligence.
3. Compare the three definitional families of predictive quality (outcome-, measurement- and system-oriented).
4. Distinguish predictive quality from neighbouring concepts such as predictive analytics, clinical prediction models, risk adjustment and early warning systems.
5. Classify predictive quality applications by function, level, time horizon and method.
6. Recognise the maturity levels that range from simple risk scales to the learning health system.
7. Evaluate the criteria of a good application (discrimination, calibration, clinical utility, external validation and fairness).

Concept Map of the Chapter

Figure 1. Concept map of the chapter

Figure 1. Concept map of the chapter. The left column shows the roots and boundaries of the concept, the right column its applied dimensions; each box points to the relevant section.

Contents

The chapter moves from the simplest concepts to the most sophisticated applications. Each section adds a layer on top of the previous one.

1. Introduction: Managing quality through the windscreen, not the rear-view mirror
2. Building blocks: Quality, risk and prediction
   1. Quality of care and patient safety
   1. Risk and probability
   1. The difference between explaining and predicting
   1. What is a prediction model?
3. A brief history: From Nightingale to TRIPOD+AI
4. Defining predictive quality
   1. The double-meaning trap: “prediction of quality” and “quality of prediction”
   1. Three definitional families: outcome-, measurement- and system-oriented definitions
   1. The working definition of this chapter and its conceptual schema
5. Distinguishing predictive quality from neighbouring concepts
6. Types of predictive quality
   1. Five types by function
   1. Types by Donabedian dimension, level, time horizon and method
7. The predictive quality cycle: How does a model do its work?
8. From simple to sophisticated: Six maturity levels
9. Projection to the present: Mapping the 2025 and 2026 literature
10. Attributes of a good predictive quality application
11. Risks, limitations and ethics
12. Implications for Türkiye and conclusion
13. Glossary and discussion questions
14. References

1. Introduction: Managing Quality Through the Windscreen, Not the Rear-View Mirror

Imagine driving a car while looking only in the rear-view mirror. You would see the road behind you very clearly, but you would learn about the bend ahead only once you were already in it. For many years, quality management in health care organisations worked in this way. Falls, infections, medication errors and readmissions were recorded after they happened, turned into indicators months later, and only then did anyone talk about improvement.

The cost of this backward-looking approach is high. The US Institute of Medicine’s report To Err Is Human placed patient safety at the centre of health policy by showing that preventable medical errors cause tens of thousands of deaths in hospitals every year (Kohn et al., 2000). Crossing the Quality Chasm, published a year later, defined high-quality care through six aims: care that is safe, effective, patient-centred, timely, efficient and equitable (Institute of Medicine [IOM], 2001). None of these aims can be achieved merely by counting the past; preventing harm requires anticipating harm.

Predictive quality is the name for the effort to manage quality through the windscreen. Its central question is simple: In which patient, unit or hospital is a quality or safety problem likely to arise in the near future, and what should we do now with that information?

Box 1. One ward, two mornings

Picture two different mornings on the same internal medicine ward. On the first morning, the nurse in charge fills in the incident report for an 82-year-old patient who fell the night before. On the second morning, an alert from the electronic record has flagged the same patient as being at “high risk of falling”: age, a sedative medication, frequent night-time trips to the toilet and yesterday’s gait assessment have been brought together. The nurse lowers the bed, puts the call button in the patient’s hand and arranges overnight supervision. On the first morning quality was measured; on the second morning quality was produced. The essence of predictive quality is the move from the first morning to the second.

An early and instructive example showing that this move can work in the real world comes from a large urban public teaching hospital in the United States. Researchers automatically calculated the 30-day readmission risk of patients with heart failure every day from electronic records and delivered an intensive transitional care programme to the patients at highest risk. The readmission rate fell from 26.2% to 21.2%, while no significant change was observed in the control conditions (Amarasingham et al., 2013). The model by itself fixed nothing here. What improved quality was the decision to direct scarce resources to the patients who needed them most. This distinction will come up again and again throughout the chapter: prediction is a tool; quality is the result of the action that the tool triggers.

Predictive quality is a rapidly growing field. In the clean dataset built for this chapter, records before 2020 are limited to single-digit annual counts, whereas 47 studies were published in 2025 and 36 in the first ten months of 2026. This rapid growth is an opportunity, but it also creates confusion: the same words are used for different things. The primary aim of this chapter is therefore to clarify the content of the concept.

2. Building Blocks: Quality, Risk and Prediction

Predictive quality stands at the intersection of three concepts: quality, risk and prediction. It is hard to grasp the whole without understanding each of the three on its own. We therefore start from the very basics.

2.1. Quality of Care and Patient Safety

Quality of care, in its simplest form, is the degree to which the services provided produce the desired health outcomes. Avedis Donabedian, the founding thinker of this field, proposed examining quality through three components: structure, process and outcome (Donabedian, 1966/2005, 1988). Structure refers to staff, equipment and physical conditions. Process is what is done during care. Outcome is the change that occurs in the patient’s health status.

Patient safety is the most fundamental dimension of quality and refers to preventing avoidable harm caused by health care itself. Harm related to health care is often called an adverse event: a fall in hospital, a pressure injury, a healthcare-associated infection, a medication error or a postoperative complication.

The numerical measures used to monitor quality are called quality indicators. “Falls per 1,000 patient days” and “30-day unplanned readmission rate” are examples of such indicators. Those that are particularly sensitive to the quality of nursing care are called nursing-sensitive quality indicators; falls and pressure injuries are the best-known examples.

2.2. Risk and Probability

Risk is the probability that an undesired outcome will occur. A patient carrying “an 18% risk of falling” means that, of every 100 similar patients, about 18 are expected to fall within a given period. This answer is not a prophecy; it is a statement of probability. It does not say that this particular patient will or will not fall.

Risk has two faces. At the individual level, risk helps to answer the question what should I do for this patient? At the organisational level, it is used to answer the question are this hospital’s outcomes better or worse than expected given the risk of its patients? Predictive quality makes use of both faces.

2.3. The Difference Between Explaining and Predicting

A distinction that researchers often confuse is the one between explanatory modelling and predictive modelling. An explanatory model asks why? and tests a causal hypothesis. A predictive model asks what will happen? and aims to produce as accurate an estimate as possible for new observations (Shmueli, 2010).

This difference matters in practice, because a good predictor is not always a cause. In a study published in 1995, a nurse’s record that the patient “required total assistance for bathing” emerged as the strongest single predictor of in-hospital death in patients with stroke and pneumonia (Davis et al., 1995). Help with bathing does not cause death, but it is a strong signal of the patient’s overall frailty. A prediction model makes use of this signal. A quality manager, however, must not fall into the following trap: changing the predictor (for example, not filling in the record) does not change the outcome.

2.4. What Is a Prediction Model?

A prediction model is a mathematical tool that combines several characteristics to estimate the probability that a particular outcome will occur in an individual (Steyerberg, 2019). If the outcome is a condition that is present now, the model is called diagnostic; if it will occur in the future, it is called prognostic (Collins et al., 2015). Most models used in predictive quality are prognostic.

Every prediction model has the same anatomy:

Component	What it means	Fall risk example
Outcome	The event being predicted	A fall during the hospital stay
Predictors	The variables entering the model	History of falling, secondary diagnosis, ambulatory aid, intravenous access, gait, mental status
Target population	The patients for whom the model is valid	Adult inpatients
Time horizon	The period covered by the prediction	Throughout the stay
Output	A probability or a score	Total score
Threshold	The cut-off for taking action	Above a given score = “high risk”

The predictors in the table are the six items of the Morse Fall Scale, which is widely used in nursing (Morse et al., 1989). The scale works through simple addition, yet its logic is the same as that of today’s complex artificial intelligence models: combine characteristics, estimate the probability, act once the threshold is crossed.

A model’s life passes through three phases. In the development phase, the model is built on a dataset. In the validation phase, it is tested on other data: validation in data split from the same source is called internal validation, while validation in a different time, place or population is called external validation. In the updating phase, the model is adapted to new conditions (Steyerberg, 2019). As we will see, these three phases are the backbone of the reliability of predictive quality.

3. A Brief History: From Nightingale to TRIPOD+AI

Although the expression “predictive quality” is new, the idea behind it is the product of a century and a half of accumulated work. Reading this body of work in five eras makes it easier to understand why the concept has taken its present shape (Figure 2).

Figure 2. The five eras of predictive quality

Figure 2. The five eras of predictive quality and their main milestones. Eras are shown at equal intervals rather than to scale; the highlighted era represents the present.

3.1. The Birth of Measurement (1850s-1960s)

The story begins with the idea of comparing the outcomes of health care with numbers. In Notes on Hospitals, Florence Nightingale argued that hospital mortality statistics should be kept in a uniform and comparable way (Nightingale, 1863). This was one of the first steps in turning quality into a measurable object.

The second root comes from industry. Walter Shewhart’s statistical process control proposed calculating the “normal” range of variation of a process so that departures from that range could be caught early (Shewhart, 1931). Control limits are, in fact, a simple prediction: if the process does not change, future measurements will stay within this range. As we will see, this industrial root is also the source of the present-day manufacturing use of the term “predictive quality”.

3.2. Risk Adjustment and Severity Scores (1966-1999)

Donabedian’s structure-process-outcome framework showed that outcomes could be used as quality indicators, but only on one condition: differences in patients’ baseline risk must be taken into account (Donabedian, 1966/2005). A high mortality rate in a hospital that admits severely ill patients may be the result of its case-mix rather than of poor care.

This problem gave rise to the field of risk adjustment (Iezzoni, 2013). APACHE II, which scores illness severity in intensive care (Knaus et al., 1985), the Charlson index, which summarises comorbidity burden (Charlson et al., 1987), and the Elixhauser measures adapted to administrative data (Elixhauser et al., 1998) are the core tools of this era. All of them are prediction models: they calculate the “expected” outcome from a patient’s characteristics.

What this era means for quality thinking is that the gap between expected and observed came to be read as a quality signal. The oldest record in this chapter’s dataset does exactly that: a logistic prediction model was built for patients hospitalised with AIDS-associated pneumonia, showing that most of the up to threefold difference in mortality between two hospitals was explained by patients’ severity on admission (Mathews et al., 1989). Yet risk-adjusted comparisons were not flawless either. An analysis of Medicare hospital mortality data in the United States showed that some centres could look better than they were, because patients referred from distant areas formed a selected and healthier group (Ballard et al., 1994). In the same period, Berwick (1989) brought to health care an understanding of quality as continuous improvement rather than a hunt for culprits. In the late 1990s, early warning scores based on vital signs emerged for ward patients (Morgan et al., 1997, as cited in Royal College of Physicians [RCP], 2017).

3.3. The Safety Era and Report Cards (2000-2009)

The To Err Is Human and Crossing the Quality Chasm reports placed patient safety and quality at the top of the policy agenda (IOM, 2001; Kohn et al., 2000). Risk-adjusted “report cards” became widespread in this period. Testing five risk models for coronary artery bypass surgery in the same community hospital showed that the models could have similar discrimination but different calibration (Reed & Olenchock, 2003). A similar study stressed that interventional cardiology models developed in other centres could suffer calibration problems when applied to local patients (Matheny et al., 2005). In other words, a model that worked well in one place did not guarantee the same success elsewhere.

Prediction models were also used in this period to define new quality indicators. In Switzerland, risk-adjusted rates of “potentially avoidable reoperations” were computed from routine hospital data (Halfon et al., 2007). In Japan, prediction models for the hospital standardized mortality ratio were developed from national administrative records (Miyata et al., 2008). At the end of the decade, the finding that almost one in five Medicare patients (19.6%) were rehospitalised within 30 days of discharge moved readmission to the centre of quality policy (Jencks et al., 2009).

3.4. Electronic Records and Machine Learning (2010-2019)

The spread of electronic health records changed the rules of the game. Risk was no longer a number calculated months later; it could be calculated in real time while the patient was still in hospital. In patients with heart failure, an automated model running on record data, which also used social and economic markers, predicted readmission better than existing national models (Amarasingham et al., 2010). A simple readmission score for general medicine patients (Hasan et al., 2010) and the widely used LACE index (van Walraven et al., 2010) were published in the same year. Nevertheless, a systematic review a year later found that most readmission models performed poorly (Kansagara et al., 2011).

In the same period, the National Early Warning Score (NEWS) was standardised in the United Kingdom; its ability to discriminate cardiac arrest, unanticipated intensive care admission and death was demonstrated (Smith et al., 2013), and it was updated as NEWS2 in 2017 (RCP, 2017). On the reporting side, the TRIPOD statement defining how prediction model studies should be reported was published (Collins et al., 2015; Moons et al., 2015).

At the end of the decade, deep learning entered the stage. Models using the raw record data of two academic centres predicted in-hospital death, 30-day readmission and prolonged length of stay with higher accuracy than traditional models (Rajkomar et al., 2018). Two warnings came in the same years. The PROBAST tool made it possible to assess the risk of bias in prediction model studies systematically (Wolff et al., 2019). Obermeyer et al. (2019) showed that an algorithm affecting millions of patients produced systematic bias against Black patients because it predicted health care costs rather than illness.

3.5. Validation, Implementation and Governance (2020-2026)

The main question of the latest era is not “can a model be built?” but “does the model work safely in real life?” The example that best shows why this question is critical is the external validation of a commercial sepsis model used in hundreds of US hospitals. Across 38,455 hospitalisations, the model showed a low discrimination of 0.63, missing 67% of sepsis cases while generating alerts for 18% of all patients (Wong et al., 2021). The COVID-19 pandemic vividly demonstrated how changes in data over time (dataset shift) can break models (Duckworth et al., 2021; Finlayson et al., 2021).

On the other hand, tangible gains from well-designed applications were also documented. A programme that automatically predicted clinical deterioration in ward patients and relayed alerts to rapid response teams was associated with lower 30-day mortality after an alert across 19 hospitals (Escobar et al., 2020). Deploying a deep learning sepsis model in the emergency department was associated with a 17% relative reduction in in-hospital sepsis mortality and with improved compliance with the sepsis care bundle (Boussina et al., 2024). Organisations began to establish governance structures to evaluate and monitor artificial intelligence applications (Liao et al., 2022); the World Health Organization published guidance on the ethics and governance of artificial intelligence (World Health Organization [WHO], 2021). The TRIPOD statement was updated as TRIPOD+AI to cover machine learning (Collins et al., 2024).

In short, the concept has evolved from measuring to comparing with the expected, from there to seeing ahead, and finally to turning what is seen ahead into safe action.

4. Defining Predictive Quality

Clarifying the content of the concept is the most important task of this chapter. To do this, we first need to see where the confusion comes from, then classify the existing definitional approaches, and finally propose a working definition.

4.1. The Double-Meaning Trap: “Prediction of Quality” and “Quality of Prediction”

The expression “predictive quality” appears in the literature in at least four different senses. Only one of them is the subject of this chapter.

Meaning	Question asked	Within the scope of this chapter?
Prediction of quality	Can a quality or safety outcome be estimated in advance?	Yes, the core of the concept
Quality of prediction	How accurate, reliable and free of bias is a prediction model?	No, but an indispensable tool (see Section 10)
Predictive validity of an indicator	Does a quality indicator really reflect quality of care?	Partly, a problem for the measurement-oriented definition
Predictive quality in manufacturing	Can product quality be estimated from production process data?	No, the industrial meaning of the term

The most established literal use of the term is in fact in manufacturing. There, predictive quality means using data collected from the production process to estimate a product’s quality before it is inspected (Tercan & Meisen, 2022). When transferred to health care, the “product” is the outcome of care and the “process data” are the records of the patient and of the care provided. The analogy is useful, but one difference must not be forgotten: a patient is not a product, and the prediction shapes decisions made about that person.

In the health care literature, “predictive quality” is rarely used as a term. The concept is usually built by placing three components side by side: predictive analytics or a prediction model, quality or patient safety, and a health care setting. The dataset of this chapter was searched precisely at the intersection of these three components, and records that did not substantively carry all three were screened out. For example, records using the word “quality” only in the sense of quality of life or of the methodological quality of studies were excluded.

4.2. Three Definitional Families

When the uses in the literature are examined, predictive quality is seen to be organised around three definitional families. Each family answers a different question.

1. Outcome-oriented definition: “Who will be harmed?” In this approach, predictive quality is the estimation of a patient-level quality or safety outcome before the event occurs. A study that defines the pressure injury rate as “an important patient safety metric and an indicator of the quality of nursing care” and predicts it with machine learning is a typical example of this family (Song et al., 2021). Most fall, pressure injury, readmission, delirium and sepsis models belong to this family. The aim is prevention.

2. Measurement-oriented definition: “Who is better or worse than expected?” In this approach, the prediction model is used not to protect a patient but to evaluate an organisation fairly. The model calculates the expected outcome for each patient, and the organisation’s observed outcome is compared with this expectation. A study that monitored the lung cancer surgery outcomes of 148 hospitals in France with risk-adjusted rates and modelled the hospitals’ performance trajectories (Bernard et al., 2026), and a study that used risk-adjusted inpatient falls as indicators of system performance (Satoh et al., 2026), are current examples of this family. The aim is fair comparison and accountability.

3. System-oriented definition: “How does the organisation turn foresight into safe action?” This is the broadest definition. It views predictive quality not as a single model but as an organisational capability: collecting data, validating the model, embedding it in the workflow, monitoring results and governance. Studies on organisational artificial intelligence governance (Freeman et al., 2025; Liao et al., 2022) and a scoping review that maps the contribution of artificial intelligence to quality and patient safety against the World Health Organization’s quality dimensions (Xu et al., 2026) belong to this family. This definition is directly related to the idea of the learning health system, in which data are continuously turned into learning (Friedman et al., 2010). The aim is sustainable and safe transformation.

These three families are not rivals but nested rings. A fall model (outcome-oriented) can also be used to calculate risk-adjusted fall rates (measurement-oriented), and both uses live within the organisation’s monitoring and governance arrangements (system-oriented).

4.3. The Working Definition of This Chapter

In light of the discussion above, this chapter adopts the following working definition:

Predictive quality is the advance estimation, through validated prediction models that draw on available data, of outcomes, events or performance related to the quality of care and patient safety in health services, and the translation of these estimates into preventive, corrective or comparative quality decisions whose effects are monitored.

The definition contains four necessary components. If any one of them is missing, what is present is not predictive quality but a neighbouring concept:

1. Forward-looking estimation: There is a prediction model or a risk score. A dashboard that only describes the past is not enough.
2. A quality or safety target: What is estimated is related to a dimension of quality. Pure clinical prognosis that predicts only the course of a disease is not enough.
3. A link to decision and action: The estimate is tied to a prevention, targeting, comparison or resource decision.
4. An evaluation loop: The effects of the model and the action are monitored, and the model is updated when needed.

The relationship between these components and the definitional families is summarised as a conceptual schema in Figure 3.

Figure 3. Conceptual schema of predictive quality

Figure 3. Conceptual schema of predictive quality. Three inputs (prediction model, quality target and health care setting) form the core of the concept. The outcome- and measurement-oriented definitions live within the system-oriented definition; the meanings on the right are distinguished from the concept.

5. Distinguishing Predictive Quality from Neighbouring Concepts

Predictive quality sits within a family of concepts whose boundaries touch one another. Most of these concepts are a part, a tool or a neighbour of predictive quality; none is synonymous with it. The table below is designed to show these differences at a glance.

Neighbouring concept	Core question	Relationship to predictive quality	Distinguishing difference
Predictive analytics	What will happen in the future?	Umbrella concept; predictive quality is its application to quality and safety	Can also predict targets unrelated to quality, such as costs, hospital bankruptcy or medical waste volumes
Clinical prediction model	What is the probability of the outcome in this patient?	Main tool, raw material	Can be pure prognosis or diagnosis; a quality target and a link to action are not required
Risk adjustment	Has the comparison been purged of case-mix?	The technique of the measurement-oriented definitional family	Its aim is to compare organisations fairly, not to protect the individual
Early warning system	Is the patient deteriorating right now?	A real-time subtype of the outcome-oriented family	Time horizon of hours; usually based on vital signs
Clinical decision support system	What should the clinician be told at the right moment?	The delivery channel through which a prediction reaches the clinic	Can be rule-based and contain no prediction at all
Quality indicator	What happened in the past period?	Can be the target or the output of a prediction	Looks backwards; predictive quality looks forwards
Surveillance	Has an event occurred, and where is it occurring?	Neighbour; some surveillance algorithms are predictive	Detects an event; predictive quality anticipates it
Quality improvement	How can we make the process better?	Predictive quality supplies data to the improvement cycle	A methodological framework; it need not involve prediction
Prescriptive analytics	What should we do?	The next link after predictive quality	Turns a risk prediction into a recommendation for the optimal action
Operational forecasting	How many patients will arrive tomorrow, how many beds will be occupied?	Boundary case	Merely capacity planning unless a timeliness, crowding or safety frame is explicitly established

Some of the distinctions in the table are often blurred in everyday language. For example, a systematic review of machine learning studies on healthcare-associated infections reported that 63% of the included approaches were predictive and 37% were aimed at retrospective detection (Scardoni et al., 2020). Both are valuable, but only the former belongs to the core of predictive quality.

Operational forecasting is the finest boundary. Forecasting the next day’s patient volume in an emergency department is, on its own, a capacity issue. However, when the same forecast is used to adjust the number of physicians and to reduce patients’ post-evaluation length of stay, with an explicit recognition that crowding threatens the quality of care, it becomes a quality intervention in the timeliness dimension. The pioneering study conducted at Hacettepe University İhsan Doğramacı Children’s Hospital is a good example of this transition (Birbilen et al., 2026).

As a practical test for separating the concept from its neighbours, the following three questions can be asked in order:

1. Does it look forward? Is a model or score that estimates a future state at the centre of the study? If not, the topic is quality measurement or surveillance.
2. Is what is estimated related to quality? Is the outcome a quality or safety indicator, or does the model explicitly serve a quality purpose? If not, the topic is clinical prognosis or operational forecasting.
3. Is it tied to a decision? Does the prediction feed a prevention, targeting, comparison or resource decision? If not, the topic is still only a model development study; it carries the potential of predictive quality but has not realised it.

6. Types of Predictive Quality

Predictive quality is not a single application but a family of applications. The most useful way to classify this family is to look at what the prediction is used for. This section first defines five types by function and then summarises the other classification axes.

6.1. Five Types by Function

Type 1. Prevention and early warning. This is the most common type. The model flags the risk of an adverse event before it occurs and triggers preventive care. Studies that use tree-based machine learning methods to identify the most important risk factors for inpatient falls (Lindberg et al., 2020) and studies that compare an artificial intelligence-based model with traditional scales such as the Morse scale (Lo et al., 2025) belong to this type. The prospective external validation of a CE-certified tool that calculates the risk of postoperative delirium, carried out within a quality improvement initiative, is also instructive: the tool showed good discrimination (AUROC 0.77) and generally accurate calibration (Reeve et al., 2025). In Western Australia, a framework was built that uses administrative data to predict hospital-acquired complications such as delirium, aspiration pneumonia, pneumonia and urinary tract infection on a daily basis (Mukhtar et al., 2025).

Type 2. Risk stratification and targeting. Here the aim is not to do everything for every patient but to direct scarce resources to the patients who will benefit most. After an artificial intelligence-based decision support tool was deployed at a regional hospital in the United States, the readmission rate fell from 11.4% to 8.1%; once the decline in control hospitals was taken into account, the relative reduction was 25% (Romero-Brufau et al., 2020). A drug-related problem risk score developed in Thailand was designed to direct clinical pharmacists’ limited time to high-risk patients (Deawjaroen et al., 2025). In this type, the value of the model depends less on its accuracy than on the effectiveness of the intervention it triggers.

Type 3. Risk-adjusted measurement and benchmarking. This is the application of the measurement-oriented definitional family. The model makes it possible to compare the performance of hospitals or units after purging it of case-mix (Bernard et al., 2026; Satoh et al., 2026). This type has its own pitfalls. An analysis using data from 125 intensive care units in Australia and New Zealand showed that risk-adjusted length-of-stay rankings changed according to the model and indicator chosen, and that there was no single “best” model (Moran et al., 2023). A multilevel analysis using data from 53 hospitals in the Netherlands found that, after adjusting for case-mix, only a very small part of the variation in readmissions (between 0.48% and 2.70% depending on the diagnosis group) could be attributed to the hospital level (Hekkert et al., 2018). These findings are a reminder that risk-adjusted rankings must be interpreted with care.

Type 4. Surveillance and signal detection. In this type, the model catches signals pointing to quality problems within large and scattered data. A model that distinguished complaints involving genuine patient harm from ten years of patient complaints in one hospital showed a discrimination of 0.904 in the test set (Luo et al., 2025). Approaches that prioritise quality gaps from free-text comments in national patient experience surveys (Ojo et al., 2024) and pharmacovigilance models that predict adverse drug reactions in hospitalised patients (Dsouza et al., 2025) are close to this type. One interesting example audits the quality measurement system itself: a study showing inflation of self-reported measures in nursing home star ratings in the United States used a prediction model to identify 6% to 8.5% of nursing homes as likely “inflators” (Han et al., 2018).

Type 5. Flow and timeliness. This type serves the IOM’s aim of timely care. In one emergency department, almost half of 177,665 patients (48.2%) experienced prolonged waiting times, and this could be predicted with interpretable machine learning models (Wang et al., 2025). A study of 172,104 emergency patients in three hospitals in the Netherlands found that machine learning models performed similarly to logistic regression in predicting the need for hospital admission (De Hond et al., 2021). This last finding carries an important lesson: a more complex model does not always mean better quality.

6.2. Other Classification Axes

Alongside function, five further axes help describe what an application is. Each application has a position on every axis; using these axes together makes it easier to see whether two studies are genuinely comparable.

Axis	Categories	Example
Donabedian dimension	Structure (e.g., nurse staffing needs), process (e.g., compliance with a care bundle), outcome (e.g., falls, death)	A sepsis model affecting both bundle compliance and mortality (Boussina et al., 2024)
Level of analysis	Micro: patient; meso: unit or hospital; macro: health system	Hospital-level performance trajectories (Bernard et al., 2026)
Time horizon	Minutes to hours (early warning), length of stay (falls, pressure injuries), weeks to months (readmission), years (fall risk in the community)	One-year fall risk in older adults (Ye et al., 2020)
Method	Point-based scale, regression and nomogram, machine learning, deep learning and language models	Predicting postoperative death from free-text anaesthesia notes with a large language model (Chen et al., 2025)
Data source	Administrative data, structured record data, nursing records, free text, sensors and images	Predicting readmission with nursing data (Oh et al., 2025)

7. The Predictive Quality Cycle: How Does a Model Do Its Work?

A prediction model does not improve quality on its own. What improves quality is the cycle in which the model sits. This cycle can be summarised in eight steps (Figure 4), and the chain is only as strong as its weakest link.

Figure 4. The predictive quality cycle

Figure 4. The predictive quality cycle. The eight steps form a loop; monitoring findings return to a new quality question.

1. Define the quality question and the decision. “We want to reduce falls” is not enough. The right question is: Which preventive bundle will be applied to which patients, at what moment, and what information would change that decision? A model is built to improve an existing decision.
2. Prepare the data. Which data are actually available at the moment of decision? A diagnosis code that is finalised only at discharge cannot be used in a model that will run on admission.
3. Develop the model. A simple score, regression or machine learning? The choice is made according to the structure of the problem and the data, not out of curiosity about complexity. Current evidence shows that, for many quality questions, a well-built regression model performs close to machine learning (De Hond et al., 2021; Mukhtar et al., 2025).
4. Validate. Internal validation comes first, followed by external and temporal validation whenever possible. Given that only 23% of the studies in a medication safety review included external validation, this step is clearly one of the weakest links in the field (Dsouza et al., 2025).
5. Run a silent pilot. The model is run in real time without being shown to clinicians, and its outputs are monitored. A model that predicts which emergency patients need an electrocardiogram was run silently for two months; across 18,335 encounters, its “in vitro” and “in vivo” decisions agreed in 95.5% of cases, and coding errors were found and corrected during this process (Bunney et al., 2026).
6. Embed in the workflow and link to action. The prediction must reach the right person, at the right time, with a concrete recommendation for action. This step usually takes place through a clinical decision support system. The threshold determines how many alerts will be generated. That is why the question “how many patients must be evaluated to find one true case?” (number needed to evaluate) is critical. In a sepsis study across nine hospitals, this number was 6.1 in emergency departments and 7.5 on wards (Schnetler et al., 2025). Too many alerts lead to alert fatigue and cause clinicians to ignore them (Ancker et al., 2017).
7. Evaluate the effect through quality outcomes. What ultimately matters is not the model’s discrimination but the change in quality outcomes such as falls, readmissions or deaths (Amarasingham et al., 2013; Boussina et al., 2024; Escobar et al., 2020).
8. Monitor and update. As the patient profile, recording habits and clinical practice change, the model’s performance changes too. This phenomenon is called data drift or, more generally, dataset shift (Finlayson et al., 2021). During COVID-19, the performance of an emergency admission model fell from 0.856 to 0.826, and explainable machine learning was used to monitor this drift (Duckworth et al., 2021). Monitoring findings return to the first step, and the cycle begins again.

The point at which the cycle most often breaks is the sixth step, the link between prediction and action. Most models in the literature stop at the third and fourth steps. In a meta-research study following 109 clinical prediction models, only 17% were subsequently externally validated, and only one reached an impact assessment (Arshi et al., 2025). This picture shows that the real tests of predictive quality begin after model development.

8. From Simple to Sophisticated: Six Maturity Levels

Predictive quality is a spectrum that stretches from a nurse’s intuition to organisational systems that continuously audit themselves. It is useful to think of this spectrum as a six-step ladder (Figure 5). Each step does not replace the previous one; it adds a new capability to it.

Figure 5. Six maturity levels of predictive quality

Figure 5. Six maturity levels of predictive quality. Each step adds a new capability on top of the previous one; higher steps carry more possibilities and more responsibility.

Level 0. Clinical intuition. An experienced nurse’s feeling that “this patient may fall tonight” is the oldest form of prediction. It is valuable, but it varies from person to person, is not recorded and cannot be measured.

Level 1. Simple risk assessment scales. Intuition becomes written and scorable. The Morse scale for falls (Morse et al., 1989), the Braden scale for pressure injuries (Bergstrom et al., 1987) and NEWS2 for clinical deterioration (RCP, 2017) are the classics of this level. They are transparent and easy to learn. Their weaknesses are fixed weights and their reliance, in most cases, on a single point in time. Capability added: standardisation.

Level 2. Statistical prediction models. Logistic regression, nomograms and risk adjustment models make up this level. The model’s weights are learned from data, and its performance is reported in terms of discrimination and calibration. A score that uses six predictors to calculate the risk of drug-related problems in hospitalised patients (Deawjaroen et al., 2025) and a LASSO model that distinguishes harm in patient complaints (Luo et al., 2025) are current examples. Capability added: probability and calibration.

Level 3. Data-intensive machine learning and explainability. Hundreds of variables from electronic records are processed with tree-based algorithms, deep learning and language models (Rajkomar et al., 2018). Hybrid models that combine structured data with clinical notes (Pandey et al., 2025) and transformer architectures (Pham et al., 2025) form the frontier of this level. The price of complexity is a loss of transparency; explainable artificial intelligence methods, especially SHAP values (Shapley additive explanations), are therefore widely used to show which factor pushes the prediction in which direction (Wang et al., 2025). Capability added: capturing complex patterns.

Level 4. Integrated, real-time, action-linked systems. The model no longer lives in a journal article but inside the clinic. Risk is calculated automatically every day or every few minutes and is tied to a defined intervention protocol (Amarasingham et al., 2013; Escobar et al., 2020). A decision support framework that embeds prediction models in nursing workflows for patients with heart failure in home health care (Chae et al., 2025) illustrates the design questions of this level well. At this level, it also becomes important for the model to know its own uncertainty: a triage model that can say “I don’t know” in cases where it is unsure left the decision to the clinician in 34.5% of cases at a 90% confidence threshold, while raising its sensitivity to 0.88 and its specificity to 0.99 in the remaining cases (Abdulai et al., 2025). Capability added: timely action.

Level 5. Learning and governed system. At the most sophisticated level, the organisation continuously monitors its models, catches drift, audits fairness, and updates or retires a model when necessary (Duckworth et al., 2021; Liao et al., 2022). A study showing that a single sepsis model performed differently in emergency departments and on wards across nine hospitals, and that models trained separately for emergency and ward patients performed better, explains why the assumption that “one model fits all” must be abandoned (Schnetler et al., 2025). The “Learn-Predict-Monitor-Detect-Correct” framework proposed for intensive care also expresses this cyclical logic at the conceptual level (Boussi Rahmouni et al., 2025). This level is the counterpart, in predictive quality, of the ideal of the learning health system (Friedman et al., 2010). Capability added: self-correction.

The ladder is not a “the higher the better” ranking. A well-implemented Level 1 scale for managing fall risk on a general ward can produce more quality than an unmonitored Level 3 model that nobody owns. The right level is chosen according to the severity of the problem, the maturity of the data and the organisation’s monitoring capacity. Higher steps offer more possibilities but also carry more responsibility.

9. Projection to the Present: Mapping the 2025 and 2026 Literature

To see where the concept stands today, the 83 records of the clean dataset published in 2025 and 2026 (71 research articles, 10 reviews and 2 early access articles) were analysed separately. The essence of the analysis is this: the field is moving from the question “can we build a model?” to the question “does the model hold up in real life, and does it help to measure quality?”

Box 2. How was this map drawn?

The title, abstract and keywords of each publication were scanned with a controlled vocabulary of 37 terms, and the concepts contained in each publication were identified. A network was built among the 37 concepts appearing in at least four publications, based on how often they co-occurred in the same publication, and the concepts were clustered with the Louvain algorithm. The modularity of the network is low (Q = 0.17); the clusters should therefore be read not as sharply bounded boxes but as overlapping emphases. The year 2026 is partial, as of October.

9.1. The Most Frequent Concepts

The methodological language of the current literature is machine learning (46 publications), but the language of quality is equally strong: patient safety appears in 44 publications and quality of care in 31. Reporting model performance through discrimination and calibration (44 publications) has now become standard. Implementation and workflow integration (39 publications), external or temporal validation (22 publications) and explainable artificial intelligence (22 publications) are other frequent concepts.

9.2. Three Thematic Clusters

The concept co-occurrence network shows that the current literature is organised around three thematic clusters (Figure 6).

Figure 6. Concept co-occurrence network of predictive quality, 2025-2026 (n = 83). Node size shows the number of publications containing the concept, line thickness shows co-occurrence frequency, and colour shows the Louvain cluster. Source: Author’s analysis of Web of Science data.

Cluster 1. Safe and validated artificial intelligence. At its centre are patient safety, artificial intelligence and implementation. External validation, clinical decision support, bias and fairness, and governance and ethics meet in this cluster. In terms of target outcomes, sepsis and clinical deterioration, healthcare-associated infections and medication safety stand out; intensive care is the dominant setting. This cluster is the current face of the system-oriented definitional family (El Arab et al., 2025; Freeman et al., 2025; Xu et al., 2026).

Cluster 2. Flow and outcome indicators with data-intensive machine learning. At its centre are machine learning, electronic records and quality of care. Deep learning, language models and explainable artificial intelligence form the methodological side, while readmission, length of stay and emergency department flow form the outcome side (Guan et al., 2026; Oh et al., 2025; Pandey et al., 2025; Wang et al., 2025). Risk adjustment and benchmarking also sit in this cluster. In other words, it is the area where big data meet quality indicators.

Cluster 3. Care-sensitive and perioperative events with classical risk models. At its centre are calibration, logistic regression and nomograms. Nursing data, the perioperative period, quality improvement and quality indicators gather around this methodological core. Falls, delirium, surgical complications and older patients are typical targets (Reeve et al., 2025; Tang et al., 2026). This cluster is the classical, nursing-centred branch of the outcome-oriented definitional family.

9.3. The Shift from 2025 to 2026

When the shares of concepts within each year are compared, a clear change of direction can be seen (Figure 7).

Figure 7. Shares of publications in 2025 (n = 47) and 2026 (n = 36) for concepts whose share changed by at least 7 percentage points. Asterisks mark concepts with p < 0.05 in Fisher’s exact test. Source: Author’s analysis of Web of Science data.

Three changes are statistically significant (Fisher’s exact test, p < 0.05). The share of publications containing the concept of quality indicators and performance rose from 4.3% to 27.8%, and the share containing external or temporal validation rose from 17.0% to 38.9%. The share of the machine learning emphasis fell from 68.1% to 38.9%. Although they do not reach the threshold of significance, implementation and workflow integration (from 40.4% to 55.6%), time-series forecasting (from 10.6% to 25.0%) and falls (from 4.3% to 13.9%) show upward trends, while deep learning, electronic records and emergency department flow show downward trends.

This picture can be read as three trends:

1. The era of validation. The appeal of algorithmic novelty is giving way to testing whether a model also works at other times and in other places. Comparisons showing no difference between a single-hospital model and a national model given a sufficient sample (Cogill et al., 2025) are an example of this trend.
2. A return to measurement. Prediction models are increasingly used not only to protect patients but also to monitor organisational performance and to forecast quality indicators over time (Bernard et al., 2026; Satoh et al., 2026; Tang et al., 2026). In other words, the outcome-oriented and measurement-oriented definitions are converging again.
3. Settling into the clinic and the organisation. Studies on silent pilots, implementation frameworks and governance are increasing (Agius et al., 2026; Bunney et al., 2026). This trend points to a movement towards the fourth and fifth steps of the maturity ladder.

Because the year is not yet complete and publication counts are small, these trends should be considered preliminary. Even so, their direction is consistent with the line drawn in the chapter’s history: first measure, then predict, then use the prediction safely.

10. Attributes of a Good Predictive Quality Application

At this point we need to return to the second meaning distinguished at the beginning of the chapter, namely the quality of prediction. Quality cannot be produced with a poor-quality prediction. A manager, clinician or researcher evaluating a predictive quality application should question the eight attributes below.

Attribute	Question asked	How it is shown	Key source
Discrimination	Can the model separate patients who will experience the event from those who will not?	Area under the ROC curve (AUC) or C statistic	Steyerberg (2019)
Calibration	When the model says 20%, do roughly 20 out of 100 patients actually experience the event?	Calibration plot, calibration slope and intercept	Van Calster et al. (2019)
Clinical utility	Is using the model better than intervening in everyone or in no one?	Decision curve analysis, net benefit	Vickers and Elkin (2006)
Generalisability	Does the model also work at another time, hospital or population?	External and temporal validation	Collins et al. (2024)
Fairness	Does the model work similarly across groups (sex, ethnicity, socioeconomic status)?	Subgroup performance, effect of social determinants	Obermeyer et al. (2019)
Explainability and uncertainty	Can the clinician see the reasons for the prediction and how confident the model is?	Explanation methods such as SHAP, confidence intervals, an “I don’t know” option	Abdulai et al. (2025)
Transparent reporting and low risk of bias	Has the study been reported in a way that others can assess?	TRIPOD+AI checklist, PROBAST assessment	Collins et al. (2024); Wolff et al. (2019)
Impact and sustainability	Do quality outcomes improve when the model is used, and is the improvement sustained?	Before-after or randomised evaluation, drift monitoring	Kelly et al. (2019)

The first two attributes in the table are the ones most often confused. A model can rank patients correctly by risk (high discrimination) while systematically overestimating absolute risks (poor calibration). In that case, the same threshold declares more patients “high risk” than necessary, resources are wasted and alert fatigue sets in. Van Calster et al. (2019) therefore describe calibration as the “Achilles heel” of predictive analytics. The slight overestimation of risk in high-risk patients by the postoperative delirium tool (Reeve et al., 2025) and the calibration problems of cardiology models from other centres when applied to local patients (Matheny et al., 2005) are concrete examples.

Box 3. The shadow of a high AUC

A model may look brilliant in a paper with an AUC of 0.93. But the meaning of this number for quality cannot be known without asking the following questions: Was the model tested in another hospital? Was its calibration reported? How many alerts does it generate to catch one true case? Which intervention is it tied to, and is that intervention effective? In practice, for many quality questions, the difference between a well-built logistic regression and complex machine learning models is small (De Hond et al., 2021). The real difference emerges in the quality of the cycle around the model.

In measurement-oriented applications, two further attributes should be added. The first is ranking stability: if the ranking of hospitals changes depending on the chosen model, decisions based on rankings should be made with caution (Moran et al., 2023). The second is the validity of the indicator: one should ask how much of the variation in an indicator actually comes from the hospital (Hekkert et al., 2018).

11. Risks, Limitations and Ethics

Although predictive quality is designed to improve quality, it can damage quality when it is badly built or badly used. This section examines the six risks most frequently encountered.

1. Predicting the wrong target (proxy bias). A model may predict not what we want to measure but a proxy variable that resembles it and is easy to measure. Obermeyer et al. (2019) showed that an algorithm that predicted health care costs instead of health needs failed to recognise that Black patients with the same risk score were sicker than White patients. Correcting this bias would raise the share of Black patients receiving additional help from 17.7% to 46.5%. The quality manager’s first question should always be: What does this model actually predict?

2. Inequity. Average performance can hide problems in subgroups. At an academic centre in New York, adding social determinants of health to a readmission model did not improve prediction across all patients, but it did produce significant improvements in subgroups such as Medicaid patients, patients aged 65 or older and patients with obesity (Zhang et al., 2020). In an application designed to prioritise patients with language barriers for interpreter services, stakeholders voiced concerns about transparency, privacy, stigmatisation and alert fatigue (Barwise et al., 2024). Fairness is not an ornament added to the model afterwards; it is a quality dimension that must be considered from the start of the design (IOM, 2001).

3. Gaming the indicator. When a measure becomes a target, there is a tendency to improve the measure rather than what it measures. Lessons drawn from the experience of the English National Health Service set out the unintended consequences of performance measurement in detail (Mannion & Braithwaite, 2012). The inflation of self-reported measures in nursing home ratings (Han et al., 2018) shows that this risk can spill over into predictive quality: risk-adjusted indicators can be made to look good by coding risk factors more heavily.

4. Alert fatigue and automation bias. A large number of low-value alerts leads clinicians to ignore alerts (Ancker et al., 2017). The commercial sepsis model that generated alerts for 18% of patients while missing most cases (Wong et al., 2021) is the most striking example. The opposite is also a risk: clinicians may trust the model more than they should and push their own clinical judgement into the background.

5. Decay over time. A model that works well today may break down tomorrow. Patient profiles, coding habits, treatment protocols or an epidemic change the structure of the data (Duckworth et al., 2021; Finlayson et al., 2021). Predictive quality also has its own paradox: as the model succeeds and events are prevented in high-risk patients, the risk-outcome relationship the model has learned weakens, and the model may appear to perform worse. Performance monitoring should therefore be interpreted with the interventions applied taken into account.

6. Gaps in accountability, privacy and governance. Who is responsible when a model makes a wrong prediction? Which data are used, for what purpose and with whose permission? The World Health Organization’s guidance places principles such as protecting autonomy, transparency, responsibility and inclusiveness at the foundation of the use of artificial intelligence in health (WHO, 2021). To put these principles into practice at the organisational level, multidisciplinary governance committees, evaluation criteria and continuous monitoring arrangements are recommended (Freeman et al., 2025; Liao et al., 2022). The barriers to turning artificial intelligence into clinical impact are as much organisational and human as they are technical (Kelly et al., 2019).

These risks are not a reason to abandon predictive quality, but an agenda for moving it up the maturity ladder. Risk itself can be predicted and managed; predictive quality should approach its own risks with the logic of predictive quality.

12. Implications for Türkiye and Conclusion

12.1. Implications for Türkiye

Türkiye has an infrastructure that is well suited to developing predictive quality. In hospitals, quality management has acquired an institutional structure within the framework of the Health Quality Standards (Sağlıkta Kalite Standartları, SKS), with patient safety at its centre (Republic of Türkiye Ministry of Health, General Directorate of Health Services [SHGM], 2020). The national personal health record system e-Nabız brings citizens’ health data together in a single place (Republic of Türkiye Ministry of Health, n.d.). In nursing practice, structured risk assessments for risks such as falls and pressure injuries have become routine. In other words, the first steps of the maturity ladder are largely in place.

Studies from Türkiye are still few in the international literature, but there are notable examples. A systematic review of artificial intelligence prediction models based on electronic records in intensive care, examined from a risk management perspective, identified mortality, sepsis, pressure ulcers and acute kidney injury as the main prediction areas (Çayırtepe & Şenel, 2022). A prospective pilot study conducted at Hacettepe University combined deep learning demand forecasting with physician scheduling and reported a consistent but modest reduction in post-evaluation length of stay that was not statistically conclusive (Birbilen et al., 2026). This honest reporting is a good example of how steps aimed at closing the “prediction-action gap” in predictive quality should be tested.

In light of this picture, six recommendations can be made for health care organisations and policy makers:

1. Test existing scales on local data. The discrimination and calibration of the fall and pressure injury scales in use should be measured in the organisation’s own patients. This is the cheapest way to move from Level 1 to Level 2.
2. Report quality indicators with risk adjustment. In comparisons between organisations, case-mix-adjusted rates and funnel plots should be used instead of crude rates, and rankings should be interpreted with caution (Bernard et al., 2026; Moran et al., 2023).
3. Treat nursing data as a strategic asset. Nursing assessments are among the strongest inputs of prediction models (Davis et al., 1995; Oh et al., 2025; Song et al., 2021). Recording these data in a structured and complete way is a quality investment.
4. Run a silent pilot for every model. Every model, whether commercial or developed in-house, should be run silently on local data before being shown to clinicians (Bunney et al., 2026; Wong et al., 2021).
5. Bring the quality unit and the IT department to the same table. An organisational governance committee for artificial intelligence applications should be established, with quality, patient safety, clinical, data science, legal and ethics representation (Liao et al., 2022). Compliance with the Turkish Personal Data Protection Law No. 6698 is a natural part of this structure.
6. Add prediction literacy to training. Health managers, quality professionals and nurses should be able to read concepts such as AUC, calibration, thresholds and data drift. Contributing to this literacy is one of the main aims of this chapter.

12.2. Conclusion

Predictive quality turns the direction of quality management in health care from the past to the future. In this chapter, the concept was defined as the advance estimation, through validated prediction models, of outcomes, events or performance related to the quality of care and patient safety, and the translation of these estimates into monitored quality decisions. The definition has four necessary components: forward-looking estimation, a quality target, a link to action and an evaluation loop. These components distinguish the concept from neighbours such as predictive analytics, clinical prognosis, risk adjustment and operational forecasting.

The history of the concept follows a line from Nightingale’s comparable statistics to risk adjustment, from there to machine learning running on electronic records, and today to validation, implementation and governance. The 2025 and 2026 literature shows that this line continues: interest in external validation and quality indicators is rising markedly, while the emphasis on algorithms alone is declining.

The main message the chapter wishes to leave with the reader is simple: prediction is not quality itself, but a lantern on the road to quality. A lantern is useful only when it is pointed in the right direction, when its settings are checked regularly, and when we actually walk to where its light shows the way.

13. Glossary and Discussion Questions

13.1. Glossary

Term	Turkish equivalent	Short definition
Adverse event	İstenmeyen olay	Harm caused by health care rather than by the disease itself
Alert fatigue	Uyarı yorgunluğu	Reduced clinician responsiveness to alerts because of their large number
Calibration	Kalibrasyon	Agreement between predicted probabilities and observed frequencies
Clinical decision support system	Klinik karar destek sistemi	A system that offers the clinician information or recommendations at the moment of decision
Clinical prediction model	Klinik tahmin modeli	A tool that combines several characteristics to estimate an individual’s outcome probability
Dataset shift	Veri seti kayması	A change over time in the structure of the data on which a model runs, degrading its performance
Decision curve analysis	Karar eğrisi analizi	Calculation of the net clinical benefit a model provides across different thresholds
Discrimination	Ayırt edicilik	A model’s ability to rank patients who will and will not experience an event; usually measured with the AUC
Early warning system	Erken uyarı sistemi	A tool, usually based on vital signs, that flags clinical deterioration hours in advance
Explainable artificial intelligence	Açıklanabilir yapay zekâ	Methods that show the reasons behind a model’s prediction in a way people can understand
External validation	Dış doğrulama	Testing a model in a time, place or population different from the data it was developed on
Learning health system	Öğrenen sağlık sistemi	A system in which routine care data are continuously turned into knowledge and improvement
Patient safety	Hasta güvenliği	Prevention of avoidable harm caused by health care
Predictive analytics	Tahmine dayalı analitik	A set of methods that predict future events from past data
Predictive quality	Prediktif kalite	Advance estimation of quality and safety outcomes with validated models, translated into monitored quality decisions
Quality indicator	Kalite göstergesi	A numerical measure used to monitor quality
Quality of care	Bakım kalitesi	The degree to which the services provided produce the desired health outcomes
Risk adjustment	Risk düzeltme	Purging comparisons of differences in case-mix
Silent pilot	Sessiz deneme	Testing a model in real time without showing it to clinicians
Threshold	Eşik değeri	The level of predicted risk at which action is taken

13.2. Discussion Questions

1. Explain the difference between “quality of prediction” and “prediction of quality” with an example from your own organisation.
2. Place a risk assessment scale used in your organisation on the maturity ladder. What would it take to move up one step?
3. If a readmission model shows high discrimination but poor calibration, for which management decisions can it be used, and for which can it not?
4. Under what conditions is forecasting patient volume in an emergency department merely operational planning, and under what conditions does it count as predictive quality?
5. Why might a successful prevention programme make its own prediction model appear to perform poorly? How would you manage this problem?
6. In what ways can risk-adjusted quality indicators be gamed? What audit mechanisms would you propose to prevent this?
7. If an artificial intelligence governance committee were to be set up in a hospital, who should its members be, and what criteria should it look for before a model is deployed?

References

Abdulai, A. S. B., Storm, J., & Ehrlich, M. (2025). I don’t know: An uncertainty-aware machine learning model for predicting patient disposition at emergency department triage. International Journal of Medical Informatics, 201, Article 105957. https://doi.org/10.1016/j.ijmedinf.2025.105957

Agius, S., Cassar, V., Magri, C., Khan, W., & Topham, L. (2026). Position paper: Advocating for a structured methodology in developing data-driven predictive models for healthcare: Evidence from a large-scale national study. Health and Technology, 16(1), 137–143. https://doi.org/10.1007/s12553-025-01010-5

Amarasingham, R., Moore, B. J., Tabak, Y. P., Drazner, M. H., Clark, C. A., Zhang, S., Reed, W. G., Swanson, T. S., Ma, Y., & Halm, E. A. (2010). An automated model to identify heart failure patients at risk for 30-day readmission or death using electronic medical record data. Medical Care, 48(11), 981–988. https://doi.org/10.1097/MLR.0b013e3181ef60d9

Amarasingham, R., Patel, P. C., Toto, K., Nelson, L. L., Swanson, T. S., Moore, B. J., Xie, B., Zhang, S., Alvarez, K. S., Ma, Y., Drazner, M. H., Kollipara, U., & Halm, E. A. (2013). Allocating scarce resources in real-time to reduce heart failure readmissions: A prospective, controlled study. BMJ Quality & Safety, 22(12), 998–1005. https://doi.org/10.1136/bmjqs-2013-001901

Ancker, J. S., Edwards, A., Nosal, S., Hauser, D., Mauer, E., & Kaushal, R. (2017). Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system. BMC Medical Informatics and Decision Making, 17(1), Article 36. https://doi.org/10.1186/s12911-017-0430-8

Arshi, B., Cowley, L. E., Rijnhart, E., Reeve, K., Smits, L. J., & Wynants, L. (2025). External validation, impact assessment and clinical utilization of clinical prediction models: A prospective cohort study. Journal of Clinical Epidemiology, 186, Article 111902. https://doi.org/10.1016/j.jclinepi.2025.111902

Ballard, D. J., Bryant, S. C., O’Brien, P. C., Smith, D. W., Pine, M. B., & Cortese, D. A. (1994). Referral selection bias in the Medicare hospital mortality prediction model: Are centers of referral for Medicare beneficiaries necessarily centers of excellence. Health Services Research, 28(6), 771–784.

Barwise, A. K., Curtis, S., Diedrich, D. A., & Pickering, B. W. (2024). Using artificial intelligence to promote equitable care for inpatients with language barriers and complex medical needs: Clinical stakeholder perspectives. Journal of the American Medical Informatics Association, 31(3), 611–621. https://doi.org/10.1093/jamia/ocad224

Bergstrom, N., Braden, B. J., Laguzza, A., & Holman, V. (1987). The Braden Scale for predicting pressure sore risk. Nursing Research, 36(4), 205–210.

Bernard, A., Cottenet, J., Tubert-Bitter, P., & Quantin, C. (2026). Measuring performance trajectories in lung cancer surgery: A longitudinal study using the French national hospital database from 2020 to 2024. BMJ Health & Care Informatics, 33(1), Article e101986. https://doi.org/10.1136/bmjhci-2025-101986

Berwick, D. M. (1989). Continuous improvement as an ideal in health care. The New England Journal of Medicine, 320(1), 53–56. https://doi.org/10.1056/NEJM198901053200110

Birbilen, A. Z., Akbasli, I. T., & Teksam, O. (2026). Forecast-driven dynamic physician staffing in a pediatric emergency department: A prospective quasi-experimental pilot study. Journal of Medical Systems, 50(1), Article 137. https://doi.org/10.1007/s10916-026-02464-4

Boussi Rahmouni, H., Hassine, N. B. E. H., Chouchen, M., Ceylan, H. I., Muntean, R. I., Bragazzi, N. L., & Dergaa, I. (2025). Healthcare 5.0-driven clinical intelligence: The Learn-Predict-Monitor-Detect-Correct framework for systematic artificial intelligence integration in critical care. Healthcare, 13(20), Article 2553. https://doi.org/10.3390/healthcare13202553

Boussina, A., Shashikumar, S. P., Malhotra, A., Owens, R. L., El-Kareh, R., Longhurst, C. A., Quintero, K., Donahue, A., Chan, T. C., Nemati, S., & Wardi, G. (2024). Impact of a deep learning sepsis prediction model on quality of care and survival. npj Digital Medicine, 7(1), Article 14. https://doi.org/10.1038/s41746-023-00986-6

Bunney, G., Miller, K., Graber-Naidich, A., Kabeer, R., Bloos, S. M., Wessels, A. J., Pasao, M. A., Rizvi, M., Brown, I. P., & Yiadom, M. Y. A. B. (2026). In vitro to in vivo translation of artificial intelligence for clinical use: Screening for acute coronary syndrome to identify ST-elevation myocardial infarction. Journal of the American Medical Informatics Association, 33(1), 7–14. https://doi.org/10.1093/jamia/ocaf101

Çayırtepe, Z., & Şenel, A. C. (2022). Risk management in intensive care units with artificial intelligence technologies: Systematic review of prediction models using electronic health records. Journal of Basic and Clinical Health Sciences, 6(3), 958–976. https://doi.org/10.30621/jbachs.993798

Chae, S., Davoudi, A., Song, J., Evans, L., Bowles, K. H., McDonald, M. V., Barron, Y., Min, S. H., Oh, S., Scharp, D., Xu, Z., & Topaz, M. (2025). Developing a clinical decision support framework for integrating predictive models into routine nursing practices in home health care for patients with heart failure. Journal of Nursing Scholarship, 57(1), 165–177. https://doi.org/10.1111/jnu.13030

Charlson, M. E., Pompei, P., Ales, K. L., & MacKenzie, C. R. (1987). A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. Journal of Chronic Diseases, 40(5), 373–383. https://doi.org/10.1016/0021-9681(87)90171-8

Chen, Y.-H., Ruan, S.-J., & Chen, P.-F. (2025). Predicting 30-day postoperative mortality and American Society of Anesthesiologists physical status using retrieval-augmented large language models: Development and validation study. Journal of Medical Internet Research, 27, Article e75052. https://doi.org/10.2196/75052

Cogill, S., Heberer, K., Kaushal, A., Fang, D., & Lee, J. (2025). Comparing single-hospital and national models to predict 30-day inpatient mortality. Journal of General Internal Medicine, 40(4), 803–810. https://doi.org/10.1007/s11606-024-09315-3

Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., Ghassemi, M., Liu, X., Reitsma, J. B., van Smeden, M., Boulesteix, A.-L., Camaradou, J. C., Celi, L. A., Denaxas, S., Denniston, A. K., Glocker, B., Golub, R. M., Harvey, H., Heinze, G., . . . Logullo, P. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, Article e078378. https://doi.org/10.1136/bmj-2023-078378

Collins, G. S., Reitsma, J. B., Altman, D. G., & Moons, K. G. M. (2015). Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): The TRIPOD statement. BMJ, 350, Article g7594. https://doi.org/10.1136/bmj.g7594

Davis, R. B., Iezzoni, L. I., Phillips, R. S., Reiley, P., Coffman, G. A., & Safran, C. (1995). Predicting in-hospital mortality: The importance of functional status information. Medical Care, 33(9), 906–921. https://doi.org/10.1097/00005650-199509000-00003

De Hond, A., Raven, W., Schinkelshoek, L., Gaakeer, M., Ter Avest, E., Sir, O., Lameijer, H., Hessels, R. A., Reijnen, R., De Jonge, E., Steyerberg, E., Nickel, C. H., & De Groot, B. (2021). Machine learning for developing a prediction model of hospital admission of emergency department patients: Hype or hope? International Journal of Medical Informatics, 152, Article 104496. https://doi.org/10.1016/j.ijmedinf.2021.104496

Deawjaroen, K., Sillabutra, J., Poolsup, N., Stewart, D., & Suksomboon, N. (2025). Development and validation of a clinical score to identify hospitalised patients at high risk of drug-related problems. Journal of Pharmaceutical Policy and Practice, 18(1), Article 2557876. https://doi.org/10.1080/20523211.2025.2557876

Donabedian, A. (1988). The quality of care: How can it be assessed? JAMA, 260(12), 1743–1748. https://doi.org/10.1001/jama.260.12.1743

Donabedian, A. (2005). Evaluating the quality of medical care. The Milbank Quarterly, 83(4), 691–729. https://doi.org/10.1111/j.1468-0009.2005.00397.x (Original work published 1966)

Dsouza, V. S., Leyens, L., Kurian, J. R., Brand, A., & Brand, H. (2025). Artificial intelligence (AI) in pharmacovigilance: A systematic review on predicting adverse drug reactions (ADR) in hospitalized patients. Research in Social & Administrative Pharmacy, 21(6), 453–462. https://doi.org/10.1016/j.sapharm.2025.02.008

Duckworth, C., Chmiel, F. P., Burns, D. K., Zlatev, Z. D., White, N. M., Daniels, T. W. V., Kiuber, M., & Boniface, M. J. (2021). Using explainable machine learning to characterise data drift and detect emergent health risks for emergency department admissions during COVID-19. Scientific Reports, 11(1), Article 23017. https://doi.org/10.1038/s41598-021-02481-y

El Arab, R. A., Almoosa, Z., Alkhunaizi, M., Abuadas, F. H., & Somerville, J. (2025). Artificial intelligence in hospital infection prevention: An integrative review. Frontiers in Public Health, 13, Article 1547450. https://doi.org/10.3389/fpubh.2025.1547450

Elixhauser, A., Steiner, C., Harris, D. R., & Coffey, R. M. (1998). Comorbidity measures for use with administrative data. Medical Care, 36(1), 8–27. https://doi.org/10.1097/00005650-199801000-00004

Escobar, G. J., Liu, V. X., Schuler, A., Lawson, B., Greene, J. D., & Kipnis, P. (2020). Automated identification of adults at risk for in-hospital clinical deterioration. The New England Journal of Medicine, 383(20), 1951–1960. https://doi.org/10.1056/NEJMsa2001090

Finlayson, S. G., Subbaswamy, A., Singh, K., Bowers, J., Kupke, A., Zittrain, J., Kohane, I. S., & Saria, S. (2021). The clinician and dataset shift in artificial intelligence. The New England Journal of Medicine, 385(3), 283–286. https://doi.org/10.1056/NEJMc2104626

Freeman, S., Wang, A., Saraf, S., Potts, E., McKimm, A., Coiera, E., & Magrabi, F. (2025). Developing an AI governance framework for safe and responsible AI in health care organizations: Protocol for a multimethod study. JMIR Research Protocols, 14, Article e75702. https://doi.org/10.2196/75702

Friedman, C. P., Wong, A. K., & Blumenthal, D. (2010). Achieving a nationwide learning health system. Science Translational Medicine, 2(57), Article 57cm29. https://doi.org/10.1126/scitranslmed.3001456

Guan, Y., Chen, X., Zhang, T., Zhao, L., Meng, R., Tang, F., Tan, Z., & Huang, X. (2026). Development and validation of dual-time-window readmission prediction models under China’s medical insurance multi-payment reform. Frontiers in Public Health, 14, Article 1909323. https://doi.org/10.3389/fpubh.2026.1909323

Halfon, P., Eggli, Y., Matter, M., Kallay, C., van Melle, G., & Burnand, B. (2007). Risk-adjusted rates for potentially avoidable reoperations were computed from routine hospital data. Journal of Clinical Epidemiology, 60(1), 56–67. https://doi.org/10.1016/j.jclinepi.2006.03.013

Han, X., Yaraghi, N., & Gopal, R. (2018). Winning at all costs: Analysis of inflation in nursing homes’ rating system. Production and Operations Management, 27(2), 215–233. https://doi.org/10.1111/poms.12804

Hasan, O., Meltzer, D. O., Shaykevich, S. A., Bell, C. M., Kaboli, P. J., Auerbach, A. D., Wetterneck, T. B., Arora, V. M., Zhang, J., & Schnipper, J. L. (2010). Hospital readmission in general medicine patients: A prediction model. Journal of General Internal Medicine, 25(3), 211–219. https://doi.org/10.1007/s11606-009-1196-1

Hekkert, K., Kool, R. B., Rake, E., Cihangir, S., Borghans, I., Atsma, F., & Westert, G. (2018). To what degree can variations in readmission rates be explained on the level of the hospital? A multilevel study using a large Dutch database. BMC Health Services Research, 18, Article 999. https://doi.org/10.1186/s12913-018-3761-y

Iezzoni, L. I. (Ed.). (2013). Risk adjustment for measuring health care outcomes (4th ed.). Health Administration Press.

Institute of Medicine. (2001). Crossing the quality chasm: A new health system for the 21st century. National Academies Press. https://doi.org/10.17226/10027

Jencks, S. F., Williams, M. V., & Coleman, E. A. (2009). Rehospitalizations among patients in the Medicare fee-for-service program. The New England Journal of Medicine, 360(14), 1418–1428. https://doi.org/10.1056/NEJMsa0803563

Kansagara, D., Englander, H., Salanitro, A., Kagen, D., Theobald, C., Freeman, M., & Kripalani, S. (2011). Risk prediction models for hospital readmission: A systematic review. JAMA, 306(15), 1688–1698. https://doi.org/10.1001/jama.2011.1515

Kelly, C. J., Karthikesalingam, A., Suleyman, M., Corrado, G., & King, D. (2019). Key challenges for delivering clinical impact with artificial intelligence. BMC Medicine, 17(1), Article 195. https://doi.org/10.1186/s12916-019-1426-2

Knaus, W. A., Draper, E. A., Wagner, D. P., & Zimmerman, J. E. (1985). APACHE II: A severity of disease classification system. Critical Care Medicine, 13(10), 818–829. https://doi.org/10.1097/00003246-198510000-00009

Kohn, L. T., Corrigan, J. M., & Donaldson, M. S. (Eds.). (2000). To err is human: Building a safer health system. National Academies Press. https://doi.org/10.17226/9728

Liao, F., Adelaine, S., Afshar, M., & Patterson, B. W. (2022). Governance of clinical AI applications to facilitate safe and equitable deployment in a large health system: Key elements and early successes. Frontiers in Digital Health, 4, Article 931439. https://doi.org/10.3389/fdgth.2022.931439

Lindberg, D. S., Prosperi, M., Bjarnadottir, R. I., Thomas, J., Crane, M., Chen, Z., Shear, K., Solberg, L. M., Snigurska, U. A., Wu, Y., Xia, Y., & Lucero, R. J. (2020). Identification of important factors in an inpatient fall risk prediction model to improve the quality of care using EHR and electronic administrative data: A machine-learning approach. International Journal of Medical Informatics, 143, Article 104272. https://doi.org/10.1016/j.ijmedinf.2020.104272

Lo, C.-L., Liu, C.-E., Chang, H.-Y., & Wu, C.-H. (2025). AI-driven fall risk prediction in inpatients: Development, validation, and comparative evaluation. Nursing Practice Today, 12(2), 141–159. https://doi.org/10.18502/npt.v12i2.18337

Luo, S., Chen, X., Wen, X., Yao, B., Wang, C., Li, Q., Wang, W., Li, L., & Zhang, Y. (2025). The hidden dangers in routine medical complaints: Uncovering patient harm. BMC Health Services Research, 25(1), Article 957. https://doi.org/10.1186/s12913-025-13105-w

Mannion, R., & Braithwaite, J. (2012). Unintended consequences of performance measurement in healthcare: 20 salutary lessons from the English National Health Service. Internal Medicine Journal, 42(5), 569–574. https://doi.org/10.1111/j.1445-5994.2012.02766.x

Matheny, M., Ohno-Machado, L., & Resnic, F. (2005). Discrimination and calibration of mortality risk prediction models in interventional cardiology. Journal of Biomedical Informatics, 38(5), 367–375. https://doi.org/10.1016/j.jbi.2005.02.007

Mathews, W. C., Ferdon, E., Bennett, C., & Kanouse, D. (1989). Evaluating institutional performance in AIDS-associated Pneumocystis carinii pneumonia: A risk adjustment approach. Journal of Clinical Epidemiology, 42(5), 421–425. https://doi.org/10.1016/0895-4356(89)90131-5

Miyata, H., Hashimoto, H., Horiguchi, H., Matsuda, S., Motomura, N., & Takamoto, S. (2008). Performance of in-hospital mortality prediction models for acute hospitalization: Hospital standardized mortality ratio in Japan. BMC Health Services Research, 8, Article 229. https://doi.org/10.1186/1472-6963-8-229

Moons, K. G. M., Altman, D. G., Reitsma, J. B., Ioannidis, J. P. A., Macaskill, P., Steyerberg, E. W., Vickers, A. J., Ransohoff, D. F., & Collins, G. S. (2015). Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): Explanation and elaboration. Annals of Internal Medicine, 162(1), W1–W73. https://doi.org/10.7326/M14-0698

Moran, J. L., Duke, G. J., Santamaria, J. D., & Linden, A. (2023). Modelling of intensive care unit (ICU) length of stay as a quality measure: A problematic exercise. BMC Medical Research Methodology, 23(1), Article 207. https://doi.org/10.1186/s12874-023-02028-x

Morse, J. M., Morse, R. M., & Tylko, S. J. (1989). Development of a scale to identify the fall-prone patient. Canadian Journal on Aging / La Revue canadienne du vieillissement, 8(4), 366–377. https://doi.org/10.1017/S0714980800008576

Mukhtar, S. A., McFadden, B. R., Islam, M. T., Zhang, Q. Y., Alvandi, E., Blatchford, P., Maybury, S., Blakey, J., Yeoh, P., & McMullen, B. C. (2025). Predictive analytics for early detection of hospital-acquired complications: An artificial intelligence approach. Health Information Management Journal, 54(2), 109–120. https://doi.org/10.1177/18333583241256048

Nightingale, F. (1863). Notes on hospitals (3rd ed., enlarged and for the most part re-written). Longman, Green, Longman, Roberts, and Green.

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. https://doi.org/10.1126/science.aax2342

Oh, E. G., Oh, S., Cho, S., & Moon, M. (2025). Predicting readmission among high-risk discharged patients using a machine learning model with nursing data: Retrospective study. JMIR Medical Informatics, 13, Article e56671. https://doi.org/10.2196/56671

Ojo, A., Rizun, N., Walsh, G., Mashinchi, M. I., Venosa, M., & Rao, M. N. (2024). Prioritising national healthcare service issues from free text feedback: A computational text analysis & predictive modelling approach. Decision Support Systems, 181, Article 114215. https://doi.org/10.1016/j.dss.2024.114215

Pandey, S. R., Tile, J. D., & Oghaz, M. M. D. (2025). Predicting 30-day hospital readmissions using ClinicalT5 with structured and unstructured electronic health records. PLOS ONE, 20(9), Article e0328848. https://doi.org/10.1371/journal.pone.0328848

Pham, M.-K., Mai, T. T., Crane, M., Brennan, R., Ward, M. E., Geary, U., Byrne, D., O’Connell, B., Bergin, C., Creagh, D., McDonald, N., & Bezbradica, M. (2025). Explainable AI for infection prevention and control: Modeling CPE acquisition and patient outcomes in an Irish hospital with transformers. BMC Medical Informatics and Decision Making, 25(1), Article 391. https://doi.org/10.1186/s12911-025-03214-1

Rajkomar, A., Oren, E., Chen, K., Dai, A. M., Hajaj, N., Hardt, M., Liu, P. J., Liu, X., Marcus, J., Sun, M., Sundberg, P., Yee, H., Zhang, K., Zhang, Y., Flores, G., Duggan, G. E., Irvine, J., Le, Q., Litsch, K., . . . Dean, J. (2018). Scalable and accurate deep learning with electronic health records. npj Digital Medicine, 1, Article 18. https://doi.org/10.1038/s41746-018-0029-1

Reed, J., & Olenchock, S. (2003). Comparative analysis of risk-adjusted bypass surgery stratification models in a community hospital. Heart & Lung, 32(6), 383–390. https://doi.org/10.1016/j.hrtlng.2003.08.001

Reeve, K. A., Gelsomino, N. S., Venturini, M., Buddeberg, F., Zozman, M., Stocker, R., Kedda, M.-A., Meier, P., Moller, M., Wildhaber, S. P., & Dodsworth, B. T. (2025). Prospective external validation of the automated PIPRA multivariable prediction model for postoperative delirium on real-world data from a consecutive cohort of non-cardiac surgery inpatients. BMJ Health & Care Informatics, 32(1), Article e101291. https://doi.org/10.1136/bmjhci-2024-101291

Republic of Türkiye Ministry of Health. (n.d.). e-Nabız. Retrieved October 11, 2026, from https://enabiz.gov.tr

Republic of Türkiye Ministry of Health, General Directorate of Health Services. (2020). Sağlıkta kalite standartları: Hastane [Health quality standards: Hospital] (Version 6.1; Ministry of Health Publication No. 1156). https://dosyamerkez.saglik.gov.tr/Eklenti/41258/0/skshastane-seti-s-61–09082021pdf.pdf

Romero-Brufau, S., Wyatt, K. D., Boyum, P., Mickelson, M., Moore, M., & Cognetta-Rieke, C. (2020). Implementation of artificial intelligence-based clinical decision support to reduce hospital readmissions at a regional hospital. Applied Clinical Informatics, 11(4), 570–577. https://doi.org/10.1055/s-0040-1715827

Royal College of Physicians. (2017). National Early Warning Score (NEWS) 2: Standardising the assessment of acute-illness severity in the NHS [Updated report of a working party]. https://www.rcp.ac.uk/media/a4ibkkbf/news2-final-report_0_0.pdf

Satoh, M., Nakahori, T., & Shimada, T. (2026). Risk-adjusted inpatient falls as indicators of health system performance during the COVID-19 pandemic. Healthcare, 14(3), Article 358. https://doi.org/10.3390/healthcare14030358

Scardoni, A., Balzarini, F., Signorelli, C., Cabitza, F., & Odone, A. (2020). Artificial intelligence-based tools to control healthcare associated infections: A systematic review of the literature. Journal of Infection and Public Health, 13(8), 1061–1077. https://doi.org/10.1016/j.jiph.2020.06.006

Schnetler, R., van der Vegt, A., Kalke, V. R., Lane, P., & Scott, I. (2025). False hope of a single generalisable AI sepsis prediction model: Bias and proposed mitigation strategies for improving performance based on a retrospective multisite cohort study. BMJ Quality & Safety, 34(9), 580–589. https://doi.org/10.1136/bmjqs-2024-018328

Shewhart, W. A. (1931). Economic control of quality of manufactured product. Van Nostrand.

Shmueli, G. (2010). To explain or to predict? Statistical Science, 25(3), 289–310. https://doi.org/10.1214/10-STS330

Smith, G. B., Prytherch, D. R., Meredith, P., Schmidt, P. E., & Featherstone, P. I. (2013). The ability of the National Early Warning Score (NEWS) to discriminate patients at risk of early cardiac arrest, unanticipated intensive care unit admission, and death. Resuscitation, 84(4), 465–470. https://doi.org/10.1016/j.resuscitation.2012.12.016

Song, W., Kang, M.-J., Zhang, L., Jung, W., Song, J., Bates, D. W., & Dykes, P. C. (2021). Predicting pressure injury using nursing assessment phenotypes and machine learning methods. Journal of the American Medical Informatics Association, 28(4), 759–765. https://doi.org/10.1093/jamia/ocaa336

Steyerberg, E. W. (2019). Clinical prediction models: A practical approach to development, validation, and updating (2nd ed.). Springer. https://doi.org/10.1007/978-3-030-16399-0

Tang, S., Li, Q., Liu, L., He, L., Zhang, Y., & Chen, L. (2026). Longitudinal analysis and predictive modeling of nursing-sensitive quality indicators in hemodialysis. Journal of Nursing Management, 2026(1), Article 8833031. https://doi.org/10.1155/jonm/8833031

Tercan, H., & Meisen, T. (2022). Machine learning and deep learning based predictive quality in manufacturing: A systematic review. Journal of Intelligent Manufacturing, 33(7), 1879–1905. https://doi.org/10.1007/s10845-022-01963-8

Van Calster, B., McLernon, D. J., van Smeden, M., Wynants, L., & Steyerberg, E. W. (2019). Calibration: The Achilles heel of predictive analytics. BMC Medicine, 17(1), Article 230. https://doi.org/10.1186/s12916-019-1466-7

van Walraven, C., Dhalla, I. A., Bell, C., Etchells, E., Stiell, I. G., Zarnke, K., Austin, P. C., & Forster, A. J. (2010). Derivation and validation of an index to predict early death or unplanned readmission after discharge from hospital to the community. CMAJ, 182(6), 551–557. https://doi.org/10.1503/cmaj.091117

Vickers, A. J., & Elkin, E. B. (2006). Decision curve analysis: A novel method for evaluating prediction models. Medical Decision Making, 26(6), 565–574. https://doi.org/10.1177/0272989X06295361

Wang, H., Sambamoorthi, N., Sandlin, D., & Sambamoorthi, U. (2025). Interpretable machine learning models for prolonged emergency department wait time prediction. BMC Health Services Research, 25(1), Article 403. https://doi.org/10.1186/s12913-025-12535-w

Wolff, R. F., Moons, K. G. M., Riley, R. D., Whiting, P. F., Westwood, M., Collins, G. S., Reitsma, J. B., Kleijnen, J., & Mallett, S. (2019). PROBAST: A tool to assess the risk of bias and applicability of prediction model studies. Annals of Internal Medicine, 170(1), 51–58. https://doi.org/10.7326/M18-1376

Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough, J., DeTroyer-Cooley, O., Pestrue, J., Phillips, M., Konye, J., Penoza, C., Ghous, M., & Singh, K. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 181(8), 1065–1070. https://doi.org/10.1001/jamainternmed.2021.2626

World Health Organization. (2021). Ethics and governance of artificial intelligence for health: WHO guidance. https://www.who.int/publications/i/item/9789240029200

Xu, Y., Veillard, J., & Kong, J. D. (2026). AI for health care quality and patient safety: Scoping review of diagnostic, predictive, and decision support applications. Journal of Medical Internet Research, 28, Article e95157. https://doi.org/10.2196/95157

Ye, C., Li, J., Hao, S., Liu, M., Jin, H., Zheng, L., Xia, M., Jin, B., Zhu, C., Alfreds, S. T., Stearns, F., Kanov, L., Sylvester, K. G., Widen, E., McElhinney, D., & Ling, X. B. (2020). Identification of elders at higher risk for fall with statewide electronic health records and a machine learning algorithm. International Journal of Medical Informatics, 137, Article 104105. https://doi.org/10.1016/j.ijmedinf.2020.104105

Zhang, Y., Zhang, Y., Sholle, E., Abedian, S., Sharko, M., Turchioe, M. R., Wu, Y., & Ancker, J. S. (2020). Assessing the impact of social determinants of health on predictive models for potentially avoidable 30-day readmission or death. PLOS ONE, 15(6), Article e0235064. https://doi.org/10.1371/journal.pone.0235064