Key points are not available for this paper at this time.
Inpatient falls remain a persistent patient safety problem, increasing morbidity, mortality, and costs. This study integrates machine learning (ML), Bayesian Belief Networks (BBNs), and Chain Event Graphs (CEGs) with the Systems Engineering Initiative for Patient Safety (SEIPS) to enhance fall-risk prediction and support prevention. A retrospective cohort of 6661 inpatient admissions from an academic hospital was analyzed, with 1.8% experiencing at least one fall. In line with recent guideline recommendations that discourage the use of stand-alone fall risk scores for screening, this work uses routinely collected electronic health record (EHR) data to support locally defined multifactorial fall prevention rather than proposing a new point-of-care score. The selected SEIPS-based Autoencoder, evaluated under nested stratified cross-validation, showed more balanced performance than purely data-driven models and achieved a recall of 0.74 for fall events. Feature analysis identified clinically plausible contributors spanning patients, tasks, tools, and technology, organizational conditions, and physical environment. BBNs captured probabilistic dependencies among these predictors, and CEGs summarized high-risk pathways involving combinations of impaired mobility, lower cognitive status, intermediate neurological responsiveness, and higher potassium values. Uplift modeling suggested that patients who received fall education had an estimated 7.5% lower probability of falling, under standard causal assumptions for observational data. Cost analyses using published per-fall estimates indicated that, if similar effects were achieved in practice, implementation could meaningfully reduce fall-related expenditures for hospitals adopting systems-oriented, data-driven fall prevention programs.
Toffaha et al. (Fri,) studied this question.