Reviewed article: Alanezi T., Li B., Al-Omran L. et al. (2026). Machine learning in the development and application of patient-reported outcome measures (PROMs) for surgical patients: a systematic review. Journal of Patient-Reported Outcomes. https://doi.org/10.1186/s41687-026-00992-8
1. Why This Review Matters
Patient-Reported Outcome Measures (PROMs) have long stood at the awkward intersection of two competing imperatives in surgical care: the clinical need to capture the patient’s subjective recovery trajectory with fidelity, and the operational reality that lengthy, rigid questionnaires erode response rates, completion quality, and ultimately the signal we are trying to measure. Alanezi and colleagues position their review precisely at this tension point, asking whether artificial intelligence (AI) and machine learning (ML) can simultaneously reduce patient burden and increase predictive value — two goals that, until recently, have been treated as a zero-sum trade-off.
The review synthesizes 22 studies stratified into three functional buckets: (i) six studies developing Computer Adaptive Test (CAT) PROMs, (ii) seven evaluating the psychometric properties of CAT PROMs, and (iii) five deploying ML algorithms for post-surgical outcome prediction, with additional studies addressing PROM score reconstruction and short-form derivation. Methodological appraisal is carried out using two distinct tools — COSMIN for measurement-property studies and PROBAST for prediction-modelling studies — a decision the authors defend explicitly and, in my view, correctly.
From a health management perspective, this review is timely for three reasons. First, value-based purchasing, quality accreditation frameworks, and patient-centered care mandates increasingly require routine PROM collection, but implementation remains hampered by completion burden. Second, ML-based outcome prediction aligns directly with preoperative risk stratification and perioperative resource allocation — two of the costliest operational decisions in hospital management. Third, the distinction between CAT PROMs (an AI-enabled measurement technology) and ML prediction models (an AI-enabled inference technology) is frequently collapsed in the applied literature, and this review is among the first to disentangle them systematically.
2. Principal Findings and Their Significance
2.1. CAT PROMs deliver efficiency without fidelity loss
The core psychometric finding is coherent and quantitatively defensible. Across the included studies, CAT PROMs demonstrated root mean square error of approximation (RMSEA) values of 0.05–0.10 — all within the conventional ≤0.10 threshold for acceptable unidimensionality. Construct validity correlations with established PROMs measuring the same health domain consistently exceeded r = 0.6, with many surpassing r = 0.8. Crucially, simulated CAT scores against original full-length PROM scores yielded near-perfect agreement (r ≥ 0.98; ICC ≥ 0.98), meaning that the CAT versions are not measuring something subtly different — they are measuring the same construct with fewer items.
The item-reduction figures are the headline operational finding: a 22.4% reduction for CLEFT-Q, 30% for HOOS, 37.5% for BREAST-Q (rising to 75% for group-level measurement), and, in the Lötsch et al. machine-learned short-form approach, a dramatic 90.4% reduction (from 73 items to 7) for predicting persistent post-surgical pain. For a hospital administrator calculating the aggregate patient-hours lost to PROM completion across a surgical service line, these reductions translate directly into measurable workflow savings.
One caveat the authors handle well: when CAT administration was truncated to one or two items, measurement error exceeded the 0.55 threshold for group-based research. This is a useful operational boundary — CATs are efficient, but there is a floor below which efficiency becomes inaccuracy.
2.2. No ML algorithm dominates, and this is the most important finding
Across the prediction-modelling studies, the absence of consistent superiority of any single algorithm — logistic regression, XGBoost, random forests, neural networks, Wide & Deep architectures, LASSO — is arguably the review’s most consequential contribution. Zhou et al. explicitly found traditional logistic regression (AUC = 0.712) outperforming classification tree (0.657), XGBoost (0.676), and random forest (0.671) for predicting improvement after total knee arthroplasty. The authors’ interpretation — that with well-defined outcomes and tabular PROM data, algorithmic sophistication yields diminishing returns because diverse algorithms converge on similar solutions when optimizing comparable objective functions — is theoretically grounded in the optimization literature (Bottou, Curtis & Nocedal, 2018) and empirically supported by the included studies.
This finding has direct policy implications for health institutions. It suggests that investment in data quality, feature engineering, and clinical interpretability should be prioritized over investment in algorithmic complexity. For health management departments designing decision-support tools, a transparent logistic regression model with well-curated PROM inputs is likely to match a black-box neural network in predictive accuracy while offering substantially better auditability, regulatory acceptance, and clinician trust.
2.3. Cross-PROM score reconstruction as a pragmatic solution
The Tenan et al. studies showing that PROMIS-PF and PROMIS-PI data can reconstruct ASES and IKDC scores within the minimal clinically important difference (MCID) threshold represent a quietly transformative finding. Registry interoperability has been a persistent problem in orthopedic outcomes research: different institutions adopt different disease-specific PROMs, making cross-institutional benchmarking difficult. If generic PROMIS domains can be mapped onto specialty-specific instruments through general additive mixed models, the path toward unified, comparable outcome registries becomes substantially shorter.
3. Methodological Strengths
The review has several commendable methodological features. The use of two separate risk-of-bias tools (COSMIN and PROBAST) is appropriate given the divergent designs of the included studies, and the authors articulate the rationale transparently. PROSPERO pre-registration (CRD42024591696) adds credibility. The four-domain search syntax — PROMs, AI/ML, measurement characteristics, surgical populations — is well-constructed and explicitly adapts the COSMIN PubMed filter. The construct validity summary in Table 4 offers an unusually clean cross-study comparison by reporting correlation coefficients alongside expected direction, which allows readers to evaluate convergent and divergent validity evidence side by side.
4. Critical Limitations the Authors Acknowledge Insufficiently
Despite its strengths, the review leaves several methodological and conceptual issues either under-examined or unaddressed. These are not minor editorial concerns — they materially constrain the external validity of the conclusions.
4.1. Single-database search
The review relies exclusively on PubMed. This is a substantial limitation for an AI/ML topic. Much of the computational methods literature — particularly work from engineering, computer science, and health informatics venues — is indexed in IEEE Xplore, ACM Digital Library, Scopus, Web of Science, and Embase but not consistently in PubMed. The authors’ claim of comprehensiveness is therefore overstated. A realistic estimate is that 20–40% of relevant ML prediction studies in surgical contexts are missed by PubMed-only searches, particularly those published in computer science conference proceedings.
4.2. Orthopedic and plastic surgery dominance
The authors note, almost in passing, that most included studies come from orthopedic and plastic surgery populations. This is not a neutral observation. It means the review’s conclusions about the performance parity of ML algorithms, the efficiency of CAT PROMs, and the feasibility of PROM-based prediction are derived from a surgical ecosystem where recovery is heavily self-reported, outcomes are measurable on continuous scales, and digital literacy is relatively high. Whether these findings generalize to cardiothoracic, neurosurgical, oncological, transplant, or emergency general surgery populations remains entirely unknown — yet the review’s framing invites that inference.
4.3. Equity, digital exclusion, and algorithmic fairness are treated cursorily
The review mentions digital exclusion and digital literacy in a single paragraph in the limitations section. This is insufficient for a 2026 publication. There is now a well-developed literature on differential CAT performance by age, educational attainment, primary language, and cognitive status; none of this is engaged with. Similarly, no included study appears to have disaggregated ML prediction model performance by race, ethnicity, socioeconomic status, or gender — a now-standard reporting expectation under frameworks such as TRIPOD-AI and the CONSORT-AI extension. From a health management standpoint, deploying a prediction model whose performance characteristics are known only in aggregate is a governance risk, not merely an academic limitation.
4.4. No cost-effectiveness or implementation analysis
The review documents the technical feasibility of AI-enabled PROMs but does not engage with their economic or implementation dimensions. What is the marginal cost per avoided readmission of a PROM-driven early-warning system? What are the IT infrastructure requirements for real-time CAT administration at scale? What clinician workflow disruptions result from integrating ML predictions into preoperative consent? These are the questions a hospital chief medical officer needs answered before committing capital, and the review does not position the included evidence to answer any of them.
4.5. Heterogeneity of MCID thresholds and comparison baselines
Although the authors correctly note that data pooling was not feasible, the qualitative synthesis itself sometimes glosses over the heterogeneity. MCID thresholds used as benchmarks for prediction accuracy vary substantially across studies (e.g., ASES MCID of 21.7 vs. IKDC MCID estimates), yet the review treats “within MCID” as a uniformly interpreted criterion. A sensitivity analysis or standardized effect-size comparison would have strengthened the quantitative discussion considerably.
4.6. Temporal validation and model drift are absent
None of the included prediction studies appears to have evaluated model performance over time — that is, the stability of predictive accuracy as surgical techniques, patient populations, and perioperative protocols evolve. The review does not flag this as a concern. For health management, this is critical: a model calibrated on 2018–2020 data may be dangerously miscalibrated in 2026 post-pandemic practice patterns, and periodic recalibration should be a governance expectation, not an afterthought.
4.7. Publication and language bias
Although the authors placed no language restrictions on their search, PubMed’s linguistic coverage is heavily English-dominant. Relevant work from Turkish, German, Japanese, Chinese, and Spanish surgical registries — some of which have made substantial investments in PROM-based outcome prediction — is likely underrepresented.
5. A Forward Research Agenda
Given the gaps identified above, I propose a structured research agenda organized into four tiers. Items marked (★) are, in my judgment, the most urgent for the field to address in the 2026–2028 window.
Tier 1 — Methodological Foundations
★ Multi-database systematic reviews with explicit ML methodology reporting. Future reviews in this space should combine PubMed, Embase, Scopus, Web of Science, IEEE Xplore, and ACM to avoid the PubMed-only bias. Reporting should follow TRIPOD-AI, CONSORT-AI, and PRISMA-AI extensions uniformly.
★ Benchmark datasets for surgical PROM prediction. The field lacks the equivalent of MIMIC or eICU for surgical PROMs — an openly available, harmonized, multi-institutional dataset against which competing algorithms can be evaluated on identical inputs. Without this, the “no algorithm dominates” finding will remain empirically underpowered.
Standardized MCID reporting conventions. Research groups should report not only whether predictions fall within MCID but also the distribution of absolute prediction errors stratified by pre-operative severity, because prediction accuracy at the extremes of the scale is clinically more consequential than mean accuracy.
Tier 2 — Equity, Fairness, and External Validity
★ Algorithmic fairness audits as a publication requirement. Every ML prediction model in surgical PROMs should report performance disaggregated by age, sex, race/ethnicity (where ethically collected), socioeconomic proxies, and primary language. Models failing fairness thresholds should be flagged, not silently deployed.
CAT PROM performance in low digital literacy populations. The assumption that CATs reduce patient burden is valid only in populations capable of navigating digital adaptive instruments. Controlled comparisons in elderly, low-literacy, and non-native-language populations are essential before CAT adoption is recommended as a universal solution.
Extension to under-represented surgical specialties. Deliberate research programs are needed in cardiothoracic, neurosurgical, transplant, oncological, and emergency general surgery — specialties where the CAT/ML evidence base is currently thin.
Tier 3 — Implementation Science and Health Management
★ Cost-effectiveness analyses of CAT PROM deployment. Studies should quantify the incremental cost per completed high-quality PROM, per avoided readmission, and per quality-adjusted life year attributable to AI-enabled PROM workflows versus status-quo paper/fixed-length digital instruments.
Implementation frameworks (CFIR, NASSS, RE-AIM) applied to AI-PROM integration. The included studies evaluate technical performance but not implementation outcomes. Hybrid effectiveness-implementation designs should become the norm.
Governance and monitoring protocols for deployed models. Health systems need standardized processes for model performance monitoring, drift detection, recalibration triggers, and de-escalation criteria when model performance degrades. This is a gap that affects every clinical AI deployment, not just PROMs.
Integration with hospital accreditation and quality frameworks. In jurisdictions with accreditation standards such as JCI, SKS Hospital Set, and NHS PROMs programs, research is needed on how CAT and ML PROM workflows align with or conflict with existing documentation, consent, and data-governance requirements.
Tier 4 — Methodological Frontiers
★ Longitudinal CAT PROMs with time-varying item banks. Current CATs select items adaptively within a single administration. Future work should explore longitudinal adaptivity — selecting items not only based on within-session responses but on the patient’s trajectory across multiple time points, potentially leveraging recurrent neural network architectures.
Multimodal PROM–biomarker–imaging–wearable integration. The review’s discussion correctly identifies this as a future direction but does not engage with the methodological complexity. Multimodal fusion architectures, missing-modality robustness, and interpretable attention mechanisms are active research problems that deserve dedicated systematic review attention.
Large Language Model-based PROM generation and interpretation. Post-2023 developments in foundation models open possibilities for free-text patient narratives to be automatically structured into PROM-equivalent scores. This is entirely absent from the included literature (the search cutoff was August 2024, and deployable clinical LLMs emerged largely after this window) and represents a critical near-term research frontier.
Causal inference layered on prediction models. Current ML models are predictive, not causal. For surgical decision-making — “will this patient benefit from this operation more than alternative management?” — causal identification strategies (doubly-robust estimation, targeted maximum likelihood, instrumental variables where applicable) should be integrated with the predictive pipeline.
Federated learning for multi-institutional PROM models. Given privacy constraints on cross-institutional data sharing, federated learning architectures should be systematically evaluated against centralized alternatives in the surgical PROM context.
6. Concluding Assessment
Alanezi and colleagues have produced a useful, methodologically careful, and timely synthesis of a rapidly moving literature. The review’s principal contributions — documenting the measurement equivalence of CAT PROMs to their full-length counterparts, quantifying the item-reduction efficiencies, and establishing that algorithm choice matters less than data quality in surgical PROM prediction — are well-supported by the included evidence and will serve as a useful reference point for clinicians, researchers, and health administrators.
At the same time, the review is best understood as a snapshot of a field in an early, largely orthopedic-dominated phase of development. Its conclusions should be extrapolated to other surgical specialties, to non-English-language contexts, to low digital literacy populations, and to economically constrained health systems only with considerable caution. The research agenda outlined above is intended not to diminish the contribution of this review but to mark the path from its conclusions toward the generalizable, equitable, and implementable AI-PROM infrastructure that patient-centered surgical care will require in the remainder of this decade.
For health management scholarship specifically, the most productive posture is neither uncritical enthusiasm nor reflexive skepticism, but rather disciplined attention to the governance, equity, implementation, and economic dimensions that the purely technical literature has so far treated as someone else’s problem.
Review perspective: Health management, patient safety, and quality in health services research.
Declared interests: None pertaining to the reviewed article.