The use of interpretability methods in machine learning models for disease prediction: a systematic review — The Association Specialists
  1. Schedule
  2. Oral Presentations - Advancing Epidemiological Methods 2
  3. The use of interpretability methods in machine learning models for disease prediction: a systematic review

The use of interpretability methods in machine learning models for disease prediction: a systematic review (24224)

Iqbal Madakkatel 1 2 , Kitty Pham 1 2 , Elina Hyppönen 1 2
  1. South Australian Health and Medical Research Institute, Adelaide, SOUTH AUSTRALIA, Australia
  2. Adelaide University, Adealide, SOUTH AUSTRALIA, Australia

Background

Machine learning (ML) methods are increasingly used to predict disease risks using population-level data, but with uptake continuing to be limited by their “black box” nature. Interpretable models are critical for assessing validity, identifying bias and informing public-health decision-making. We conducted a systematic review to examine how interpretability methods are currently applied in ML-based disease-prediction studies.

Methods

We searched PubMed, MEDLINE, bioRxiv and medRxiv for primary studies that developed or validated ML models to predict disease outcomes in general‑population. We excluded studies restricted by care setting or pre-existing condition, those using unstructured data and those focussing on intermediary outcomes, COVID‑19, mortality, prognosis, or hospital/surgical endpoints. Risk‑of‑bias was assessed using a tool tailored to interpretability. The review followed PRISMA guidelines and was registered in PROSPERO (CRD42024552595).

Results

Eighty-five studies met the inclusion criteria. Across the studies 24 interpretability methods were used, with SHAP being the most common (52%), followed by built‑in model feature importance (26%), permutation feature importance (13%), partial dependency plots (6%), and LIME (6%). Most studies reported only global explanations (76%), typically presented graphically.  Only 11% met all seven risk-of-bias criteria. Explanation stability was assessed in 38% of the included studies, subgroup/bias‑related variation in explanations 20%, and calibration of predicted probabilities was assessed in 32% studies. Real world or prototype implementation of interpretability was uncommon (8%).

Conclusions

Interpretability methods are widely used in ML-based disease prediction studies, but their evaluation and reporting are inconsistent. We recommend standardised interpretability reporting, calibration prior to model explanation, and use of multiple interpretability methods. Further development needs relate to quantification of explanation variability, routine assessment of subgroup differences and greater involvement of end users in evaluating model explanations.