Development and internal validation of a machine learning–based prediction model and simplified screening score for in-hospital falls: a retrospective cohort study

Background Conventional fall-prevention programmes rely heavily on clinical judgement and experience-based assessments, limiting their consistency and efficiency. This study aimed to develop and internally validate a machine learning (ML)-based fall-risk prediction model, characterise key risk factors and latent patient subgroups, and derive a simplified bedside screening score to support standardised fall-prevention strategies. Methods This retrospective cohort study was conducted from January 2023 to September 2024 using electronic medical records at Saiseikai Moriyama Municipal Hospital. Candidate predictors included demographics, Functional Independence Measure (FIM) scores, and routine laboratory tests. Missing values were handled by using median imputation and the “last observation carried forward” approach, as appropriate. A random forest (RF) classifier was trained and internally validated. A six-item bedside screening score was derived via cross-validated subset selection, univariate binarisation, and equal or integer weighting. Results Out of 611 participants, 120 (19.6%) experienced at least one fall during hospitalisation; the median time from admission to the first fall among fallers was 32.0 (interquartile range: 19.0–54.0) days. The RF model showed excellent discriminative performance (area under the curve [AUC]: 0.96, 95% confidence interval [CI]: 0.931–0.981, p  < 0.001). Motor and cognitive FIM scores, body mass index (BMI), age, and renal/inflammatory markers were identified as key predictors. Latent class analysis identified three phenotypic clusters (e.g., “functional impairment” vs. “metabolically vulnerable”) with distinct risk profiles, highlighting the heterogeneity of high-risk patients. The derived six-item screening score (motor FIM score ≤ 51, B-type natriuretic peptide level ≤ 66.1, prothrombin time–international normalised ratio ≤ 1.01, estimated glomerular filtration rate ≥ 78.4, haemoglobin level ≥ 11.2, BMI ≤ 20.8) yielded an AUC of 0.668 (95% CI: 0.618–0.716, p  < 0.001) with equal weights and 0.695 (95% CI: 0.644–0.742, p  < 0.001) with integer weights. High-sensitivity thresholds of ≥ 2 points for the equal-weight score and ≥ 4 points for the integer-weight score achieved sensitivity/specificity of 0.908/0.279 and 0.825/0.458, respectively. Conclusions The ML-based model and derived six-item screening score enable objective risk quantification to support efficient resource allocation. The identification of distinct risk phenotypes suggests that combining quantitative screening with qualitative profiling is essential for optimising effective fall-prevention interventions. Trial registration: Not applicable to this retrospective observational study, which was not prospectively registered.