Process Mining 101 for Oncology Hospitals: Seeing and Managing Real Patient Flow

An oncology hospital can have a tidy organisation chart, clear procedures and well-defined clinical pathways, and still have very little idea of the route its patients actually take. A patient arrives with a suspicious finding, waits for pathology, is sent for imaging, is biopsied again, goes to the tumour board, has extra tests requested after the board, has a treatment plan drawn up and only then starts treatment. Classic hospital information system reports show pieces of this: monthly attendances, mean length of stay, machine utilisation, average time to radiotherapy. They do not show in what order things happened, or why delay concentrates in some patient groups and not others. Averages also hide the extreme cases and the skew of the distribution.

Process mining addresses exactly this blind spot. Van der Aalst (2012) describes it as taking an X-ray of organisational processes from event data: when the time-stamped records each patient leaves in hospital systems are correctly assembled, the activities patients actually passed through can be reconstructed. This chapter is a manager-oriented introduction to the method for oncology hospitals, from event logs to process discovery, from pathway conformance to bottlenecks and decision support.

What process mining is, and what it is not

Process mining is the discipline of discovering, checking and improving processes using records of the events that occurred in them. For every event three things must be known: which case it belongs to, which activity it represents, and when it happened.

• It is not just a mapping technique. Traditional process maps are drawn from meetings, observation and expert opinion and represent the designed or perceived process. A mined model is derived from what actually happened. The strongest use combines the two: the data model read together with clinicians and managers.
• It is not classic data mining. Machine learning typically looks at relationships between variables in rows and columns; process mining preserves the order of events and their links over time. Whether a radiotherapy plan was revised before or after contouring, and how many times, matters.
• It does not replace business intelligence. A dashboard can warn that time to radiotherapy has lengthened; process mining shows whether the delay lies in organ-at-risk contouring, physics checks, replanning or cancelled appointments.
• It is not a staff surveillance system. Delays usually come from workload distribution, missing information, prioritisation rules, inter-unit dependencies and capacity mismatch, not from individual slowness.

Core concepts

• Event: a recorded occurrence at a point in time, such as “pathology result approved” or “first chemotherapy cycle given”. Recording time is not always the clinical time.
• Activity: the standardised name of the work. “RT sim”, “CT simulation”, “planning CT” and “CT-SIM” must be merged into one activity in a shared dictionary.
• Case: the process instance to which events are linked. In oncology, using the patient ID can merge years of different cancers into one trace; a cancer episode is often more meaningful.
• Timestamp: request, appointment, actual start, completion and result approval times are different and must be kept separate.
• Trace and variant: a trace is a case’s events in time order; cases following the same order share a variant. Many variants are not automatically poor quality, since oncology patients are clinically heterogeneous.
• Deviation: a departure from a reference model (guideline, protocol or target flow). Sometimes it is error or waste, sometimes an appropriate clinical adaptation.
• Bottleneck: the point where demand exceeds capacity and creates waiting. The longest activity is not necessarily the bottleneck.

The event log and getting it right

The event log is a table where each row is an event, with at minimum a case ID, activity name and timestamp, and ideally patient or episode attributes, unit, role, lifecycle status, cancellation reason, priority, diagnosis group and outcome. Keeping planned, cancelled and occurred appointments as separate rows preserves rescheduling loops that a single “last appointment date” would erase. Oncology data is fragmented across the hospital information system, laboratory, pathology, PACS, radiotherapy information system, chemotherapy software and pharmacy, so a linking table for patient, episode and procedure IDs is usually needed.

The chapter sets out five steps from raw data to a reliable log: write a measurable analytical question; draw the process boundaries (and state what lies outside them); build an activity dictionary approved by clinicians and process owners; run quality checks for chronology, impossible negative durations, duplicates and system-change periods (a software upgrade alone can change the mined model; Kurniati et al., 2019); and clinically validate the log against randomly selected patient files. Without that last step, a colourful process map is digital decoration, not decision support.

Three core analyses

Discovery produces the actual flow from the log without a predefined model, revealing common paths, parallel activities, loops, repetitions and rare variants. In one oncology clinic it was used to map both outpatient flow and the longitudinal pathway from first assessment to discharge, to support a patient navigation programme (Santos et al., 2026). Complex models should be shown in layers (all cases, high-volume variants, delayed cases, clinical subgroups) rather than as one spaghetti diagram.

Conformance checking compares real paths with the target process. Savino et al. (2023) turned the ESMO rectal cancer guideline into a computer-interpretable model and compared it with the real pathways of 453 patients; a subgroup deviated because of radiotherapy dose and the addition of oxaliplatin. The lesson is to report the clinical reason for a deviation, not just its rate. The chapter proposes three tiers: deviations unacceptable for patient safety (for example starting treatment before a mandatory check), clinical deviations requiring a documented reason, and managerial deviations that miss a time target without disturbing clinical order.

Enhancement adds time, resource, cost, quality and outcome information to the discovered model. In real-world melanoma data, process mining compared treatment sequences against survival, progression-free survival and response (Wicky et al., 2023). A path is not automatically good because it is fast: some longer paths create more clinical value.

Performance analysis in oncology

Most of the total time in an oncology hospital is not active clinical work but waiting: for results, approvals, appointments, boards and resources. Radiation oncology is a priority area. Bakhtiari (2025) found organ-at-risk contouring and physics checks to be the main bottlenecks, and targeted interventions reduced planning completion time. Mastroleo et al. (2026) analysed 43,183 events from 8,608 radiotherapy courses and showed that treatment interruptions significantly affected the time from prescription to treatment start. A manager’s bottleneck panel should show at least four things together: cases waiting before each activity, median and upper-percentile waiting time, rework rate, and the activity’s contribution to total delay. Variant comparisons must be risk-adjusted for stage, intent, comorbidity, performance status and urgency, or the units treating the most complex patients will look unfairly slow.

Algorithms: how much does a manager need to know?

Not the mathematics, but enough to understand the assumptions. Directly-follows graphs are intuitive but can misrepresent parallelism and loops; Alpha Miner is historic but fragile with noisy data; Heuristics Miner copes better with noise; Inductive Miner produces sound, replayable models suited to conformance checking; Fuzzy Miner simplifies highly variable processes. Model quality is judged on fitness, precision, generalisation and simplicity, and the chapter adds a fifth criterion for management: actionability.

Advanced uses

Predictive process mining estimates remaining time, next event or risk of an adverse outcome for an ongoing case; prescriptive mining suggests the next best action, such as assigning a navigator, always under human oversight. Object-centric process mining handles linked objects (patient, episode, appointment, specimen, imaging request, drug preparation, treatment plan) that a single case ID distorts. Combining mining with simulation lets managers test “what if” questions, such as adding a linear accelerator, changing physicist shifts or holding the tumour board more often, before changing the real system (Salas et al., 2024).

Quality, safety, data quality and ethics

Mandatory sequence rules, required checks and time thresholds can be monitored as event-based rules, with process mining used for learning and early warning rather than punishment. Outcomes such as complications, readmissions and cancellations can be linked back to process variants, while remembering that observational event data does not establish causation. Every analysis should come with a data quality card covering completeness, accuracy, temporal reliability, identity matching, semantic consistency and cross-system agreement. Oncology data is highly sensitive: rare diagnoses, date sequences and treatment combinations can re-identify patients even after direct identifiers are removed (Pika et al., 2020). Data minimisation is the basic principle, research logs and operational early-warning systems should be kept separate, and staff-level data should never be turned into performance penalties.

A seven-stage project model and priority uses

Select a high-value problem with clear start and end events (radiotherapy planning, pathology-to-treatment, post-board treatment start or day-case chemotherapy, not “the whole hospital”); design a precise question and success criteria; assess data readiness; discover and validate with domain experts (Pereira et al., 2020); run focused analyses with root-cause interviews; intervene with a named owner and measurable target, then re-measure the same way; and institutionalise. Priority uses for managers: diagnosis-to-treatment time, radiotherapy workflow, the tumour board, chemotherapy and systemic treatment pathways, patient navigation, and quality and safety rules.

Conclusion

Process mining moves the unit of management analysis from single indicators to real processes. Its success depends less on the algorithm than on the right question, the right case definition and a trustworthy event log. Because health care has natural clinical variability, not every deviation is an error and not every standardisation is quality. The method is a strong management tool when its results are read together with clinical expertise, patient experience, quality management and ethical governance.

Source chapter: Kurutkan, M. N. (2026). Process mining 101: Seeing and managing real patient flow in oncology hospitals. Oncology Hospital Management Training Series, supplementary chapter (original in Turkish).

References