Explainable AI to Guide the Design of Interpretable Prediction Models From High-Dimensional Data, Applied to Falls and Related Injuries in Adults Aged 50 Years or Older: Algorithm Development and Validation
Journal of Medical Internet Research · 2026-11-12 · 28 · e105364
DOI: 10.2196/105364 ↗Abstract
Background Transparent prediction models, such as Cox regression, remain the standard for clinical risk prediction. However, they are challenging to design optimally by hand when tens or even hundreds of features require selection, transformation, or interaction modeling. While complex machine learning models offer high performance, their “black-box” nature limits the clinical trust, transparency, and interpretability required for decision-making. Objective This study aimed to develop an explainable AI-based recommender that provides data-driven feature-related recommendations that, when incorporated, improve the predictive performance of existing interpretable statistical models on high-dimensional data, and to demonstrate it by developing and internally validating a Cox model for the 1-year risk of falls or related injuries (FRIs). Methods We used linked primary and secondary care records from DataLoch (Lothian, Scotland) for adults aged 50 years or older registered with a contributing general practice between January 2015 and February 2020. The outcome was the first FRI within 1 year of the index date. A random survival forest was explained with Shapley Additive Explanations (SHAP), and its feature attributions were translated into recommendations for a Cox model: feature exclusions, nonlinear terms, and interactions. Data were split 80/20 into training and test sets. We compared the recommendations-augmented Cox model with the baseline Cox model, a Cox model with restricted cubic splines, and a penalized Cox model with all pairwise interactions. The metrics were Harrell concordance index (C-index), Uno C, calibration slope, and integrated Brier score. We repeated the analysis with all 409 available features and in 3 public datasets (German Breast Cancer Study Group 2 [GBSG2], AIDS Clinical Trial [ACT], and PeakVO2). Results Of 220,341 patients (mean age 64, SD 12 years), 8489 (3.85%) had an FRI within 1 year. From 59 candidate features, the recommender excluded 15 (8 medication classes and 7 comorbidities). The final model retained 44 predictors (demographic and lifestyle factors, frailty, 21 comorbidities, and 15 medication classes), with nonlinear terms for age, frailty, BMI, and area deprivation, and interactions of age with stroke, osteoporosis, loop diuretics, and frailty. The augmented model had a C-index of 0.719 (95% CI 0.708-0.731) compared with 0.715 (95% CI 0.702-0.727) for the baseline model (difference 0.004, 95% CI 0.001-0.007), with 70 vs 73 parameters, and similar calibration (slope 0.962 vs 0.961). With all 409 features, the recommender recommended 304 exclusions, 4 nonlinear terms, and 2 interactions. The augmented model had a C-index of 0.722 compared with 0.713, with 129 vs 423 parameters, and better calibration (slope 0.952 vs 0.846). The method also proved effective on 3 additional public datasets, demonstrating wider applicability. Conclusions The proposed exploratory AI recommender demonstrates the potential of explainable AI to guide the design of transparent predictive models.
EUVIMED