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

Development and internal validation of a machine learning–based prediction model and simplified screening score for in-hospital falls: a retrospective cohort study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Development and internal validation of a machine learning–based prediction model and simplified screening score for in-hospital falls: a retrospective cohort study Onishi Tatsuki, Tatsuyoshi Ikenoue, Norihide Itoh, Takumi Nishioka, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8540709/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 12 You are reading this latest preprint version Abstract 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. Fall risk Hospital safety Machine learning Random forest SHAP analysis Functional Independence Measure Resource allocation Screening score Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Background Hospital falls represent an important safety concern affecting patient outcomes and healthcare resource utilisation [ 1 – 5 ]. Current fall-prevention interventions, including assistance, supervision, and activity-level adjustments, are predominantly experience-based and individually tailored [ 6 , 7 ]. Nevertheless, the subjective and personalised nature of these interventions may lead to inconsistent implementation and suboptimal resource allocation, potentially allowing the occurrence of falls despite preventive efforts [ 8 ]. The complexity of fall-risk assessment stems from multiple factors, including patient demographics, functional status, medical conditions, and environmental factors [ 9 , 10 ]. Conventional approaches, which rely heavily on clinical judgement and standardised assessment tools, may not capture the nuanced interactions among various risk factors for falls [ 9 , 11 , 12 ]. Furthermore, the lack of objective and quantifiable risk assessment tools limits the ability of healthcare providers to systematically prioritise interventions and effectively allocate limited human resources [ 13 ]. Machine learning (ML) approaches offer promising solutions for developing sophisticated risk prediction models that can integrate multiple variables simultaneously [ 14 ]. In particular, random forest (RF) classifiers are effective in healthcare prediction tasks owing to their ability to handle complex interactions among variables and provide interpretable feature importance rankings [ 15 ]. Additionally, advanced explainability techniques, including SHapley Additive exPlanations (SHAP) analysis, provide insights into how individual patient characteristics contribute to fall-risk predictions [ 16 , 17 ]. Nonetheless, operationalising individual predictions remains a challenge. SHAP analysis explains why a specific patient is at risk but generates unique complex profiles that hinder the standardisation of care protocols. To bridge this gap, latent class analysis (LCA) is required to reduce the individual complexity into manageable patient-level phenotypes. LCA facilitates the development of practical standardised interventions beyond simple risk quantification by clustering heterogeneous high-risk patients into recognisable clinical syndromes (e.g. functional vs. metabolic types). The present study aimed to develop and internally validate a comprehensive ML-based fall-risk prediction model, characterise latent patient subgroups through SHAP analysis and LCA, and derive a simplified bedside screening score for standardised fall-prevention interventions in hospital settings [ 18 , 19 ]. Methods Reporting guideline This study was reported in accordance with the TRIPOD + AI statement and the original TRIPOD 2015 reporting guidelines, following the EQUATOR Network recommendations [ 20 , 21 ]. Study design and setting This single-centre retrospective cohort study was conducted from January 2023 to September 2024 using routinely collected electronic medical records at Saiseikai Moriyama Municipal Hospital (Shiga, Japan), a 199-bed community hospital with 51 beds in acute-care wards, 100 beds in recovery-phase rehabilitation wards, and 48 beds in chronic/long-term care wards. Eligible inpatients with examination records during the study period were included. Participants All hospital admissions during the study period were screened for patients who underwent at least one Functional Independence Measure (FIM) assessment and a panel of routine laboratory tests recorded in the hospital information system. Only adults aged ≥ 18 years were considered eligible for study inclusion. When patients were admitted multiple times during the study period, only the first (index) admission was considered to avoid within-patient clustering. Additionally, records with missing values for key demographic or functional variables (e.g., sex, age, motor FIM score, and cognitive FIM score) were excluded because these predictors were essential and we chose not to impute them. Out of 633 individuals initially identified, 22 patients with incomplete data were excluded, leaving 611 participants for the analysis. Outcome The primary outcome was the occurrence of in-hospital falls at index admission. Based on the World Health Organization criteria, a fall was defined as an event that caused a person to come to rest inadvertently on the ground, floor, or other lower levels. Falls were identified from incident reports, nursing documentations, or electronic safety reports recorded in the hospital information system. Outcome assessors were unaware of the model’s predictions. For patients who experienced multiple falls, only the first fall at admission was considered for the analysis, and the number of days from admission to the first fall was calculated. Predictors and data collection Routinely collected demographic (sex and age), anthropometric (height and weight), functional, and laboratory variables were regarded as candidate predictors. Body mass index (BMI) was calculated as weight divided by height squared (kg/m²). Functional status was assessed using the FIM, and both motor and cognitive FIM scores were recorded repeatedly during hospitalisation as part of routine care. Based on observational data on patients’ activities of daily living in the ward, the FIM scores were examined every 4 weeks from admission through a discussion between the primary physical therapist or occupational therapist and the nursing staff. The smoking index was included as a measure of cumulative tobacco exposure. Laboratory variables included serum albumin (ALB), B-type natriuretic peptide (BNP), C-reactive protein (CRP), prothrombin time–international normalised ratio (PT-INR), total bilirubin (T-Bil), total cholesterol (T-CHO), creatinine (CRE), estimated glomerular filtration rate (eGFR), γ-glutamyl transpeptidase (γ-GTP), fasting blood glucose (FBG), and haemoglobin (Hgb). Use of relevant medications such as antihypertensive, antidiabetic, and sedative–hypnotic agents was also extracted and coded as a binary variable. Measurements of time-varying predictors (motor and cognitive FIM scores and laboratory values) were aligned with a reference time point for each patient. The reference time point was defined as the day before the first fall for fallers (patients who experienced a fall) and as the day before discharge for non-fallers. For each predictor, the most recent measurements available at or before the reference time point were used. The last observation carried forward (LOCF) approach was applied when the FIM scores or laboratory values at the reference time point were missing but earlier measurements during the same admission were available. Patients with no recorded FIM assessment at any time during hospitalisation were excluded because the FIM was considered an essential predictor and was not imputed. Variables that remained entirely missing even after applying the LOCF approach were imputed using median values within the ML pipeline. Handling of missing values The following strategies were applied to variables with missing values even after using the LOCF approach for repeated measurements: records with missing values for sex, age, and motor or cognitive FIM score were excluded from the analysis; missing values for the smoking index were assumed to represent non-smokers and were imputed as 0; and continuous anthropometric and biochemical variables (height, weight, PT-INR, eGFR, and ALB, BNP, CRP, T-Bil, T-CHO, CRE, γ-GTP, FBG, and Hgb levels) were imputed using the cohort median. The normality of distributions was not formally assessed prior to median imputation, which represents a limitation. The proportion of missing values for each laboratory predictor before imputation is summarised descriptively in Supplementary Table S1 in the Additional File. Sample size considerations To avoid overfitting, effective model complexity was constrained relative to 120 observed fall events, consistent with the TRIPOD-AI recommendations and the formal sample size criteria for prediction models described by Riley et al., which emphasised adequate events per parameter, minimal optimism, and stable estimation of outcome proportions [ 22 ]. Multicollinearity assessment For numerical explanatory variables (excluding the fall outcome), variance inflation factors (VIF) within the regression-based pipeline were calculated to assess potential multicollinearity among candidate predictors. Height and weight showed substantial collinearity with BMI (VIF > 10) and were consequently excluded from the final set of predictors. The VIF values before and after height and weight removal are provided in Supplementary Table S2 in the Additional File. Model development and performance evaluation The performance of the fall-prediction model was evaluated by constructing receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC). This model was based on an RF classifier, and the target variable was the binary fall outcome (0 or 1). ROC curves were plotted using predicted fall probabilities (fall scores), with the AUC serving as a quantitative measure of discriminative ability. As the fall outcome was moderately imbalanced (19.6% of fall events), we evaluated whether class-weight adjustment was necessary. However, the imbalance was judged to be sufficiently mild, and the internal validation indicated no meaningful benefit from weighting. Therefore, we did not apply class weighting and trained the RF classifier with class_weight = none, ensuring that the model complexity aligned with the available number of events and followed contemporary recommendations for prediction model development. This decision was also consistent with the observed stability of ROC curve performance in cross-validation. In addition to discrimination, overall prediction accuracy was quantified using the Brier score, and calibration was assessed numerically by estimating the calibration intercept and slope from a logistic recalibration model with the logit of the predicted probabilities as the sole predictor. Two complementary methods were employed to determine the statistical significance of the AUC—namely, logistic regression and bootstrap resampling. A logistic regression model was fitted with the fall outcome as the dependent variable, and a bootstrap procedure with 1,000 iterations was performed to recalculate the AUC under resampling. The proportion of iterations with AUC ≤ 0.5 (i.e., equivalent to random classification) was used to estimate a one-sided p -value. ROC curves were visualised, and the AUC was reported as the primary indicator of model performance. The RF classifier was implemented using scikit-learn and default settings. In addition to discrimination, the calibration of the RF model was analysed using calibration plots. The predicted fall probabilities were grouped into deciles, and the observed fall proportion in each group was plotted against the mean predicted probability, with the 45-degree line representing perfect calibration. The model was based on an RF classifier implemented using Python 3.12 (Python Software Foundation, Delaware, the United States of America) and scikit-learn version 1.5.2, and the target variable was the binary fall outcome (0 or 1). Standard hyperparameter settings were applied, unless otherwise specified: n_estimators = 100, random state = 42, max_depth = None, max_features = "sqrt", class_weight = None, and bootstrap = True. These settings corresponded to the default configuration in scikit-learn. Fall-score distribution analysis The distribution of predicted fall scores was visualised using histograms stratified by outcome status. For patients who experienced a fall (fallers), the fall-score distribution was plotted to examine central tendency and skewness among high-risk individuals. Additionally, overlaid histograms for fallers and non-fallers were created using 20 bins to visually assess the separation between the groups. The RF classifier was trained without class weighting (class_weight = None), reflecting our judgement that the class imbalance was mild. Thus, the resultant fall-score distribution represented the unweighted separation between fallers and non-fallers. Feature importance analysis Feature importance scores after model training were extracted to evaluate the relative contribution of each predictor to the RF model. The predictors were ranked in descending order of importance and visualised using a horizontal bar plot, with the most influential features displayed at the top. Care-related variables analysis To explore the associations between care-related interventions and falls, 2×2 contingency tables comparing the fall outcome with four categorical variables (presence of a therapist during training, adherence to bed-rest restrictions, assistance provided, and monitoring status) were created. Pairwise statistical comparisons between intervention types were performed, with Bonferroni correction being applied to adjust for multiple testing. To determine whether the implementation of preventive interventions differed by the predicted risk, patients who experienced a fall (fall = 1) were analysed, and their fall scores were categorised into three groups: <0.33, 0.33 to < 0.66, and ≥ 0.66. The proportion of patients who received each preventive intervention was calculated for each group. These patterns were visualised using stacked bar charts, heat maps, and 100% stacked bar charts. LCA LCA was conducted using a Gaussian mixture model to identify latent patient subgroups based on fall-related features. All numeric predictors, except the fall outcome, were standardised (z-scores). The remaining missing values were imputed using median values, and rows with persistent missing data were excluded. The optimal number of clusters was determined using the elbow method, and the resultant clusters were characterised by summarising descriptive statistics, calculating the feature importance for cluster classification, comparing key predictors among clusters, and analysing the fall scores by cluster. This approach was consistent with the use of clustering methods in previous studies to identify clinically meaningful patient subgroups in risk prediction research [ 23 , 24 ]. SHAP analysis for model explainability To enhance model explainability, SHAP analysis was conducted to quantify the contribution of individual features to RF predictions. For the SHAP analysis, we used an RF regressor with the same predictor set and similar hyperparameters as the classification model, which was trained to approximate the predicted fall probability. This allowed us to decompose individual predictions into additive feature contributions on a probability scale. Eighteen numerical features, including demographics, anthropometric measures, FIM scores, and laboratory values, were included as predictors. The dataset was divided into training (80%) and test (20%) sets. An RF regressor with the same predictor set and similar hyperparameters was trained on the training data to predict fall probability. SHAP values were computed for the training data and visualised for the test data to interpret feature contributions at both global and individual levels. Sensitivity analysis Two main sensitivity analyses were conducted to assess robustness. For the primary analysis, examination findings up to the day before the fall event were used, and missing data were imputed. In the first sensitivity analysis, examination findings from the day of the fall event were instead used as the most recent results. In the second sensitivity analysis, records with missing examination data were excluded, and a complete-case analysis was performed without imputation. For each sensitivity analysis, ROC curves and histograms of fall scores were generated and compared with those of the primary analysis. Development of the screening score We sought a parsimonious, hand-calculated screening score. Candidate predictors were routinely available numeric variables, and BMI was not excluded a priori in this step. For operational reasons, age, sex, and smoking status were not considered for bedside scores, whereas laboratory markers and FIM scores were considered. A fixed pipeline (median imputation, standardisation, class-weighted logistic regression) was evaluated under stratified five-fold cross-validation (random state = 42). For each subset size 𝑘, all combinations were exhaustively scored, and the smallest subset that met our target (for sensitivity targeting: existence in every fold of an ROC operating point with sensitivity of ≥ 0.80 and specificity of ≥ 0.50) were selected; ties were resolved by median AUC. For the selected subset, each variable was binarised via univariate ROC with Youden’s 𝐽 on training folds [ 25 , 26 ]. The median threshold and direction (≥ or ≤) across folds were recorded. Subsequently, an equal-weight score (+ 1 per satisfied condition) and an integer-weight score were constructed using L1-penalised logistic regression on the binarised variables, scaling positive coefficients to small integers [ 27 ]. Performance was summarised by cross-validated AUC (95% confidence interval [CI]) and two-sided bootstrap p -value under H0 (AUC = 0.5). The final six-item subset comprised the motor FIM score, BNP level, PT-INR, eGFR, Hgb level, and BMI. For the final integer-weight score, calibration was assessed by constructing calibration plots of predicted versus observed fall risks across score-based risk groups. Results Study participants Out of 633 individuals identified during the study period, 22 patients with incomplete data on key demographic or functional variables were excluded, leaving 611 participants for the analysis (Fig. 1 ). Overall, 120 (19.6%) participants experienced at least one fall during hospitalisation. The median time from admission to the first fall was 32.0 (interquartile range: 19.0–54.0) days. Participant characteristics The baseline characteristics of the overall cohort stratified by fall status are summarised in Table 1 . Fallers were slightly older and had lower motor and cognitive FIM scores at the reference time point than non-fallers. Fallers also exhibited lower weight and BNP level; other laboratory parameters showed minimal differences between the groups. Antihypertensive and sedative–hypnotic agents were more frequently used among fallers (Table 1 ). Table 1 Characteristics of the participants Overall Fallers Non-fallers P -value n 611 120 491 Male, n (%) 229 (37.5) 48 (40.0) 181 (36.9) 0.595 Age, mean (SD) 78.6 (12.8) 80.6 (11.8) 78.1 (13.0) 0.046 Height, mean (SD) 155.6 (10.2) 154.6 (10.1) 155.9 (10.3) 0.224 Weight, mean (SD) 52.2 (12.5) 50.3 (11.4) 52.6 (12.7) 0.050 BMI, mean (SD) 21.4 (3.9) 21.0 (4.0) 21.5 (3.9) 0.176 Motor FIM score, median [Q1, Q3] 46.0 [25.0, 61.0]) 39.5 [24.0, 49.0] 48.0 [25.5, 63.0] < 0.001 Cognitive FIM score, median [Q1, Q3] 25.0 [16.0, 33.0] 22.0 [16.0, 28.0] 25.0 [16.0, 34.0] 0.008 Smoking, mean (SD) 218.2 (1167.0) 180.2 (977.7) 227.5 (1209.5) 0.652 ALB, mean (SD) 3.6 (0.5) 3.6 (0.5) 3.6 (0.5) 0.383 BNP, mean (SD) 92.9 (62.9) 82.9 (51.0) 95.3 (65.3) 0.025 CRP, mean (SD) 1.4 (2.3) 1.4 (2.1) 1.4 (2.3) 0.911 PT-INR, mean (SD) 1.1 (0.1) 1.0 (0.2) 1.1 (0.1) 0.351 T-Bil, mean (SD) 0.7 (0.3) 0.7 (0.2) 0.7 (0.3) 0.338 T-CHO, mean (SD) 165.1 (22.9) 164.9 (23.2) 165.1 (22.9) 0.926 CRE, mean (SD) 1.1 (1.0) 1.0 (0.9) 1.1 (1.0) 0.675 eGFR, mean (SD) 61.9 (20.0) 61.3 (17.1) 62.1 (20.7) 0.660 γ-GTP, mean (SD) 37.2 (45.7) 34.5 (30.3) 37.9 (48.7) 0.327 FBG, mean (SD) 132.6 (25.7) 132.8 (33.6) 132.5 (23.4) 0.922 Hgb, mean (SD) 11.8 (1.6) 11.8 (1.6) 11.8 (1.6) 0.843 Antihypertensive agents, n (%) 29 (4.7) 12 (10.0) 17 (3.5) 0.005 Antidiabetic agents, n (%) 138 (22.6) 32 (26.7) 106 (21.6) 0.284 Sedative–hypnotic agents, n (%) 233 (38.1) 59 (49.2) 174 (35.4) 0.008 FIM, Functional Independence Measure; BMI, body mass index; ALB, serum albumin; BNP, B-type natriuretic peptide; CRP, C-reactive protein, PT-INR, prothrombin time–international normalised ratio, T-Bil, total bilirubin; T-CHO, total cholesterol; CRE, creatinine; eGFR, estimated glomerular filtration rate; γ-GTP, γ-glutamyl transpeptidase; FBG, fasting blood glucose; Hgb, haemoglobin; SD, standard deviation; IQR, interquartile range Fall-score distribution The distribution of predicted fall scores among fallers skewed towards higher values, indicating that the model assigned higher estimated probabilities to several fall cases (Fig. 2 A). However, the fall-score distribution notably overlapped between fallers and non-fallers, highlighting the residual variability that was not fully captured by the model (Fig. 2 A). Model performance The RF model achieved an AUC of 0.96 (95% CI: 0.931–0.981, bootstrap p < 0.001), indicating excellent discriminative ability (Fig. 2 B). The average AUC from cross-validation was 0.96, suggesting stable performance across resamples. The calibration plot showed good agreement between the predicted and observed fall probabilities, although the risk was slightly underestimated in the group with the highest predicted risk. The Brier score for the RF model was 0.050, and the calibration slope and intercept from a logistic recalibration model were 2.02 and 0.64, respectively, indicating overall good prediction accuracy but some overestimation of risk and overfitting at higher predicted probabilities. Feature importance Among predictors, the motor FIM score at admission was the most important feature, followed by the cognitive FIM score, BMI, age, and several laboratory markers such as the PT-INR, eGFR, and CRE, Hgb, and ALB levels. These predictors were ranked (Fig. 2 C). Care-related interventions The analysis of care-related variables using contingency tables (Fig. 3 A) revealed varying patterns of preventive intervention implementation. Adherence to bed-rest restrictions had significantly higher implementation rates than assistance (43.3% vs. 3.33%, p < 0.001) and supervision (43.3% vs. 15.8%, p < 0.001). Assistance was implemented less frequently than supervision ( p = 0.006) (Fig. 2 D). The proportion of fallers under each intervention category is presented in Fig. 2 D, whereas intervention-by-outcome contingency tables are shown in Fig. 3 A, clarifying how often each intervention was implemented among fallers and non-fallers. Fall score–based intervention patterns Contingency tables of fall occurrence and supervision/intervention type are presented in Fig. 3 B. When the participants were stratified into three groups by fall scores (< 0.33, 0.33 to < 0.66, and ≥ 0.66), the heatmap revealed non-uniform patterns of intervention implementation across risk levels (Fig. 3 B). Some high-risk patients did not receive intensive interventions, whereas certain lower-risk patients did, suggesting a potential misalignment between the predicted risk and actual preventive care. LCA Inspection using the elbow method suggested three distinct latent classes. The resultant clusters represented patient phenotypes with different combinations of clinical and functional parameters. For instance, one cluster comprised patients with markedly impaired motor FIM score and high fall score, whereas another cluster included patients with relatively preserved functional status and low fall score. These patterns supported the presence of heterogeneous fall-risk profiles in hospitalised populations. SHAP analysis The SHAP analysis provided detailed insights into the feature contributions at the individual level. Patients with lower motor FIM score and somewhat unexpectedly younger age tended to have higher SHAP values, indicating an increased predicted fall risk. Higher FBG and γ-GTP levels, PT-INR, and eGFR were also associated with an increased fall risk in the model. These findings revealed the complex and sometimes counterintuitive relationships between predictors and predicted fall probabilities. Sensitivity analysis The results when using laboratory values from the day of the fall event are presented in Fig. 4 A–B, and the complete-case analysis is illustrated in Fig. 4 C–D. Sensitivity analyses yielded fall-score distributions and ROC curves that were broadly similar to those in the primary analysis. Using laboratory findings from the day of the fall event instead of the previous day or restricting the analysis to complete cases without imputation did not materially change the discrimination performance or overall distributional patterns of fall scores. Prediction of the six-item screening score (equal vs. integer weights) The selected subset comprised six predictors: motor FIM score, BNP level, PT-INR, eGFR, Hgb level, and BMI. Binarisation thresholds (median across folds; direction in parentheses) were as follows: motor FIM score ≤ 51 (≤), BNP level ≤ 66.1 (≤), PT-INR ≤ 1.01 (≤), eGFR ≥ 78.41 (≥), Hgb level ≥ 11.2 (≥), and BMI ≤ 20.8 (≤) (Table 2 ). The equal-weight score (+ 1 per condition) achieved an AUC of 0.668 (95% CI: 0.618–0.716, bootstrap p < 0.001). At high-sensitivity threshold of ≥ 2 points, the sensitivity and specificity were 0.908 and 0.279, respectively (positive predictive value: 0.235, negative predictive value: 0.926, accuracy: 0.403) (Fig. 5 A–B). The Youden optimal cut-off was ≥ 3 points (sensitivity: 0.533, specificity: 0.703). The calibration plot (Fig. 5 C) indicated that the predicted probabilities derived from the equal-weight score aligned well with the observed fall proportions across the risk strata. Table 2 Direction, threshold, and equal and integer weights for items Item Direction Threshold Equal-weight points Integer-weight points Motor FIM score ≤ 51.0 + 1 + 3 BNP ≤ 66.1 + 1 + 1 PT-INR ≤ 1.01 + 1 + 2 eGFR ≥ 78.41 + 1 + 1 Hgb ≥ 11.2 + 1 + 1 BMI ≤ 20.8 + 1 + 1 FIM, Functional Independence Measure; BNP, B-type natriuretic peptide; PT-INR, prothrombin time–international normalised ratio; eGFR, estimated glomerular filtration rate; Hgb, haemoglobin; BMI, body mass index The integer-weight score (L1 logistic-derived) assigned weights of 3 to the motor FIM score, 2 to PT-INR, and 1 to the remaining four items, yielding an AUC of 0.695 (95% CI: 0.644–0.742, bootstrap p < 0.001). At high-sensitivity threshold of ≥ 4 points, the sensitivity and specificity were 0.825 and 0.458, respectively (accuracy: 0.530) (Fig. 5 D–E). The calibration plot (Fig. 5 F) indicated that the predicted probabilities derived from the integer-weight score aligned well with the observed fall proportions across the risk strata. These results verified that retaining BMI improved the coherence with the RF importance profile and that the six-item bedside score provided a practical trade-off between parsimony and discrimination suitable for repeated inward screening. Discussion In the present study, we developed a comprehensive fall-risk prediction system using routinely collected hospital data and ML methods. Our approach integrated supervised learning for precise risk estimation with unsupervised learning for patient profiling. The RF model showed excellent discriminative performance (AUC: 0.96), providing an objective data-driven tool for identifying high-risk individuals. LCA revealed distinct patient subgroups. Our findings indicated that combining quantitative prediction with qualitative profiling could not only identify at-risk patients but also provide a basis for tailoring interventions according to clinical phenotypes. Key findings and clinical implications In our study, we identified the motor FIM score as the most important predictor; this result aligns with the established understanding that mobility impairment is a primary risk factor for falls. Our SHAP analysis further refined this insight by revealing that younger patients with poorer functional status may still be at a high risk, challenging the conventional focus on age alone. Additionally, the model highlighted the potential relevance of the cognitive FIM score and several laboratory parameters, particularly BMI, eGFR, FBG, and inflammatory markers such as CRP. The observed association between higher eGFR and increased fall risk may reflect sarcopenia-related phenomena, in which individuals with lower muscle mass exhibit higher eGFR calculated from serum CRE levels [ 28 , 29 ]. Nevertheless, these findings require further investigation. The overlap in fall-score distribution between fallers and non-fallers underscores the multifactorial nature of falls and suggests that even a high-performing model cannot fully resolve individual-level uncertainty. Therefore, prediction models should be used as decision-support tools rather than deterministic classifiers. In this context, the six-item bedside screening score derived from the model (motor FIM score ≤ 51, BNP level ≤ 66.1, PT-INR ≤ 1.01, eGFR ≥ 78.4, Hgb level ≥ 11.2, and BMI ≤ 20.8) achieved moderate discrimination (AUC: 0.668 with equal weights and 0.695 with integer weights) while remaining sufficiently simple for routine use. High-sensitivity cut-offs (≥ 2 points for the equal-weight score and ≥ 4 points for the integer-weight score) enabled repeated risk stratification during hospitalisation, potentially facilitating a timelier allocation of preventive resources and targeted supervision for patients at greatest risk [ 30 , 31 ]. The cognitive FIM score was one of the top predictors in the RF model and showed considerable SHAP contributions; however, it did not remain in the final six-item screening score, likely because of methodological differences between the full multivariable model and the constrained subset selection procedure. The logistic regression pipeline used for score derivation emphasises linearly separable thresholds that maximise cross-validated discrimination using a minimal number of binary rules. Although the cognitive FIM score contributed mainly through complex nonlinear interactions in the SHAP analysis, these patterns translated less efficiently into a single threshold-based rule. This implies that cognitive impairment does not merely add to the risk but acts as a ‘force multiplier’ that compounds other physical vulnerabilities—a dynamic nuance that is inherently difficult to capture with a static, single-point threshold. Consequently, the cognitive FIM score was not retained despite its importance in the ML model. This distinction illustrates the complementary roles of SHAP (model interpretability) and simplified scoring systems (operational feasibility). Rather than enforcing a single universal cut-off, we propose that the model-based fall scores be used as a continuous decision-support tool to contextualise FIM-based risk assessments, particularly in patients who experience falls despite not being classified to have a high risk by conventional criteria. Clinical rationale for the inclusion of BMI, BNP level, and PT-INR in the screening score Retaining the BMI at the bedside is consistent with the model’s importance profile and clinical plausibility. Low BMI may be a proxy for sarcopenia-related weakness, whereas high BMI may indicate relative strength limitations and mobility inefficiency; these mechanisms align with previously described relationships between BMI and falls [ 29 , 30 ]. In our combinatorial search under cross-validation, BMI contributed to robust monotonic effects that were well captured by a single-threshold-based rule (≤ 20.8), making it suitable for inclusion in a parsimonious screening score. Although the BNP level and PT-INR were selected in the six-item subset, these variables should not be overinterpreted as primary fall-risk determinants. Their inclusion reflects the combinatorial optimisation process; under cross-validation, they provided small but consistent incremental discrimination when paired with high-importance predictors such as the motor FIM score and BMI. The SHAP analysis revealed that the BNP level and PT-INR contributed to modest global effects but occasionally showed meaningful local contributions in individuals with metabolic or cardiovascular instability. These findings implied that the BNP level and PT-INR functioned as auxiliary markers that captured systemic vulnerability rather than as direct fall-risk drivers. Their modest integer weights (1–2 points) were chosen to reflect this secondary role, ensuring that their operational impact remained limited while preserving their incremental predictive value. Intervention patterns and opportunities for improvement Our analysis of care-related interventions revealed substantial variation in implementation rates across intervention types, with adherence to bed-rest restrictions being more common than direct assistance or supervision. This pattern likely reflects the resource constraints and practical considerations in busy clinical settings [ 32 ]. The fall-score–based intervention analysis indicated that the current allocation of preventive care does not always align with the predicted risk, suggesting opportunities to design score-based intervention protocols that prioritise high-risk patients while avoiding unnecessary restrictions on lower-risk individuals. However, in this study, the fall scores were primarily used as a continuous measure to interrogate patterns of care rather than to define prescriptive risk categories. The three fall-score strata used in our descriptive analyses (< 0.33, 0.33 to < 0.66, and ≥ 0.66) were chosen for interpretability and to visualise the mismatch between model-predicted risk and delivered preventive care; they are not proposed as ready-to-use clinical cut-offs. In practice, we envisage that the model-derived fall scores will complement existing FIM-based assessments by highlighting patients whose predicted risk is discordant with their subjectively assessed risk, whereas any operational threshold for score-based interventions should be tailored and prospectively evaluated within each institution. The six-item screening score enables efficient identification of patients at high risk; however, it does not fully explain why specific patients fall. The identification of three latent patient clusters supports the existence of heterogeneous fall-risk phenotypes, which a single cumulative score may obscure. Our findings suggest that an effective fall-prevention strategy requires a two-step approach combining quantitative risk stratification and qualitative profiling. First, the developed six-item screening score acts as a sensitive filter for identifying high-risk patients requiring immediate attention. However, a high score alone does not indicate the specific type of intervention required. This is where the value of the LCA lies. Once patients are flagged to have a high risk, classifying them into a specific clinical phenotype, as identified by our LCA, enables tailored interventions. For example, patients fitting the “functional impairment” cluster should be prioritised for intensive physical therapy and environmental adjustments (e.g., sensor mats), whereas those in the “metabolically vulnerable” cluster require a review of medications, hydration, and nutritional management rather than physical restrictions. This workflow transforms the prediction model from a simple warning system into a comprehensive decision support tool, potentially resolving the mismatch between patient needs and resource allocation. Contributions of SHAP analysis The SHAP analysis enhanced our understanding of how the RF model combines predictors to generate fall-risk estimates. By orthogonalising the predictor contributions, the SHAP analysis enabled the interpretation of complex interactions without relying solely on marginal associations. In this study, younger patients with poorer functional status may still have a high fall risk; this finding challenges the conventional focus on age alone and underscores the need to prioritise functional decline irrespective of chronological age. The positive association between certain laboratory parameters and fall risk highlights the importance of comprehensive medical management in fall-prevention strategies [ 33 ]. Importantly, the SHAP analysis informed the score-development process by clarifying which predictors contributed primarily through strong monotonic effects—amenable to thresholding—and which predictors contributed through complex interaction patterns that were less suitable for a simplified additive score. The motor FIM score, BMI, and renal markers showed clear and directionally consistent SHAP effects, supporting their inclusion in the score. In contrast, the cognitive FIM score exhibited substantial but highly interaction-dependent SHAP contributions, explaining its exclusion from the threshold-based score despite its importance in the full RF model. Thus, the SHAP analysis served not only as an interpretability tool but also as a guide for selecting variables compatible with a clinically implementable risk-stratification tool. Limitations This study has some limitations. First, given that this was a single-centre retrospective study, selection bias and unmeasured confounding factors cannot be excluded, and the generalisability of our findings to other hospitals or healthcare systems is uncertain. External validation in independent cohorts and diverse settings is essential before this model can be adopted in routine practice. Second, although calibration was evaluated descriptively using calibration plots (Fig. 5 C and F), summary calibration indices, such as the calibration slope and Brier score were not calculated, limiting our ability to fully quantify the accuracy of predicted probabilities. Third, missing values for laboratory predictors were handled using simple median imputation, and the normality of distribution was not formally assessed. This might have introduced bias, especially if the missingness was informative. Fourth, the categorical nature of care-related variables prevented us from quantifying the intensity, duration, and quality of interventions, and we did not account for fall severity or circumstances. Fifth, our analyses primarily considered baseline or most recent measurements, and we did not explicitly model temporal changes in risk during hospitalisation. Even if immediate examination findings predict falls at some point during admission, the lead time until a fall is not constant, complicating real-time resource allocation. Finally, the model is internally validated. Implementation in practice would require not only external validation and potential recalibration but also careful consideration of the risk of automation bias and unintended consequences of over-reliance on algorithmic outputs [ 34 , 35 ]. Although fall events were moderately imbalanced, we did not apply class weighting because the imbalance was considered mild and internal validation showed stable performance without rebalancing. Consequently, model predictions reflected the natural distribution of the fall outcome rather than weighted estimates. Future directions Future research should focus on prospective validation of the prediction model, integration of additional risk factors (e.g., detailed medication profiles and environmental assessments), and development of dynamic models that update risk predictions as patient status evolves. Incorporating real-time monitoring data and automated alert systems can further enhance the clinical utility of fall-prediction models. Investigations into cluster-specific intervention strategies and cost-effectiveness analyses comparing score-based and usual care approaches would provide valuable evidence for healthcare decision-makers. Conclusions We developed an ML-based fall-risk prediction model and a simplified six-item bedside screening score using routinely collected hospital data. The model provides objective and quantifiable risk estimates that can support standardised fall-prevention interventions and efficient human resource allocation. Although the model shows excellent discriminative performance, challenges in achieving perfect prediction accuracy, ensuring good calibration, and demonstrating external validity remain. Combining ML approaches with clinical expertise and careful evaluation is essential for improving patient safety and optimising resource utilisation in in-hospital fall-prevention programmes. Abbreviations AUC Area under the curve BMI Body mass index BNP B-type natriuretic peptide CI Confidence interval CRE Creatinine CRP C-reactive protein eGFR Estimated glomerular filtration rate FBG Fasting blood glucose FIM Functional Independence Measure Hgb Haemoglobin LCA Latent class analysis LOCF Last observation carried forward ML Machine learning NPV Negative predictive value PT-INR Prothrombin time–international normalised ratio RF Random forest ROC Receiver operating characteristic SD Standard deviation SHAP SHapley Additive exPlanations γ-GTP Gamma-glutamyl transpeptidase Declarations Ethics approval and consent to participate This retrospective study was approved by the Institutional Review Board of Saiseikai Moriyama Municipal Hospital (approval no. 30). The requirement for individual informed consent was waived because de-identified routinely collected data were used and the study posed minimal risk to the participants. Consent for publication Not applicable. Availability of data and materials The datasets generated and/or analysed during the current study, as well as the analysis code used for data preprocessing and model development, are not publicly available because of institutional restrictions but are available from the corresponding author upon reasonable request. Competing interests The authors declare that they have no competing interests. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Authors’ contributions TO and TI conceived the study. TO curated the data and performed the statistical analyses. TI contributed to the study design, data interpretation, and clinical contextualisation. TO drafted the manuscript, and TI critically revised it for important intellectual content. Both authors have read and approved the final version of the manuscript. Acknowledgements The authors thank the clinical and administrative staff of Moriyama Municipal Hospital for their assistance with data extraction and continuous efforts in fall-prevention activities. Authors’ information Not applicable. References Kim DW, Seo J, Kwon S, Park CM, Han C, Kim Y, et al. Predicting In-hospital Fall risk Using Machine Learning with Real-Time Location System and Electronic Medical Records. J Cachexia Sarcopenia Muscle. 2025;16:e13713. doi:10.1002/jcsm.13713. Hill KD, Vu M, Walsh W. Falls in the acute hospital setting — Impact on resource utilisation. Aust Health Rev. 2007;31:471-7. doi.org/10.1071/AH070471. Oliver D, Healey F, Haines TP. Preventing falls and fall-related injuries in hospitals. Clin Geriatr Med. 2010;26:645-92. doi:10.1016/j.cger.2010.06.005. Montero-Odasso MM, Kamkar N, Pieruccini-Faria F, Osman A, Sarquis-Adamson Y, Close J, et al. Evaluation of clinical practice guidelines on fall prevention and management for older adults: A systematic review. JAMA Netw Open. 2021;4:e2138911. doi:10.1001/jamanetworkopen.2021.38911. Tricco AC, Thomas SM, Veroniki AA, Hamid JS, Cogo E, Strifler L, et al. Comparisons of interventions for preventing falls in older adults: A systematic review and meta-analysis. JAMA. 2017;318:1687-99. doi:10.1001/jama.2017.15006. Guirguis-Blake JM, Michael YL, Perdue LA, Coppola EL, Beil TL. Interventions to prevent falls in older adults: Updated evidence report and systematic review for the US Preventive Services Task Force. JAMA . 2018;319:1705–16. doi:10.1001/jama.2017.21962. Spoelstra SL, Given BA, Given CW. Fall prevention in hospitals: An integrative review. Clin Nurs Res. 2011;21:92-112. doi:10.1177/1054773811418106. Randell R, McVey L, Wright J, Zaman H, Cheong VL, Woodcock DM, et al. Practices of falls risk assessment and prevention in acute hospital settings: A realist investigation. Health Soc Care Deliv Res. 2024;12:1-194. Oliver D, Daly F, Martin FC, McMurdo MET. Risk factors and risk assessment tools for falls in hospital in-patients: A systematic review. Age Ageing. March 2004;33:122–30, https://doi.org/10.1093/ageing/afh017. Strini V, Schiavolin R, Prendin A. Fall risk Assessment Scales: A Systematic Literature Review. Nurs Rep. 2021;11:430-43. https://doi.org/10.3390/nursrep11020041. Rubenstein LZ. Falls in older people: Epidemiology, risk factors and strategies for prevention. Age Ageing. September 2006;35 Suppl 2, Issue suppl_2:ii37–41, https://doi.org/10.1093/ageing/afl084. Callis N. Falls prevention: Identification of predictive fall risk factors. Appl Nurs Res. 2016;29:53-8. doi:10.1016/j.apnr.2015.05.007. Greene BR, Redmond SJ, Caulfield B. Fall risk Assessment through Automatic Combination of Clinical Fall risk Factors and Body-Worn Sensor Data. IEEE J Biomed Health Inform. 2017;21:725-31. doi:10.1109/JBHI.2016.2539098. Schultebraucks K, Galatzer-Levy IR. Machine learning for prediction of posttraumatic stress and resilience following trauma: An overview of basic concepts and recent advances. J Trauma Stress. 2019;32:215-25. doi:10.1002/jts.22384. H Mohammed S, Ahamad S, Abdul Hameed DM. Enhancements in random forest algorithms for improving healthcare applications. In: 4th Indian International Conference on Industrial Engineering and Operations Management ; 2024, November, https://doi.org/10.46254/IN04.20240192. Mishra AK, Skubic M, Despins LA, Popescu M, Keller J, Rantz M, et al. Explainable Fall risk Prediction in Older Adults Using Gait and Geriatric Assessments. Front Digit Health. 2022;4:869812. doi:10.3389/fdgth.2022.869812. Tang YT, Romero-Ortuno R. Using explainable AI (XAI) for the prediction of falls in the older population. Algorithms. 2022;15:353. https://doi.org/10.3390/a15100353. Yip MYT, Lim G, Lim ZW, Nguyen QD, Chong CCY, Yu M, et al. Technical and imaging factors influencing performance of deep learning systems for diabetic retinopathy. NPJ Digit Med. 2020;3:40. He J, Baxter SL, Xu J, Xu J, Zhou X, Zhang K. The practical implementation of artificial intelligence technologies in medicine. Nat Med. 2019;25:30–6. Collins GS, Dhiman P, Andaur Navarro CL, et al. TRIPOD-AI: An updated reporting guideline for prediction model studies involving artificial intelligence. BMJ. 2024;385:e079044. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): The TRIPOD statement. Ann Intern Med. 2015;162:55–63. Riley RD, Ensor J, Snell KIE, Harrell FE Jr, Martin GP, Reitsma JB, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441. BMJ. 2019;368:m441. Wong ES, Yoon J, Piegari RI, Rosland AMM, Fihn SD, Chang ET. Identifying latent subgroups of high-risk patients using risk score trajectories. J Gen Intern Med. 2018 ; 33:2120–6. https://doi.org/10.1007/s11606-018-4653-x. Wang Y, Zhao Y, Therneau TM, Atkinson EJ, Tafti AP, Zhang N, et al. Unsupervised machine learning for the discovery of latent disease clusters and patient subgroups using electronic health records. J Biomed Inform. 2020;102:103364. doi:10.1016/j.jbi.2019.103364. Marschollek M, Gövercin M, Rust S, Gietzelt M, Schulze M, Wolf KH, et al. Mining geriatric assessment data for in-patient fall prediction models and high-risk subgroups. BMC Med Inform Decis Mak. 2012;12:19. https://doi.org/10.1186/1472-6947-12-19. Youden WJ. Index for rating diagnostic tests. Cancer. 1950 Jan;3:32-5. doi: 10.1002/1097-0142(1950)3:13.0.co;2-3. PMID: 15405679. Daneshvar A, Mousa G. Regression shrinkage and selection via least quantile shrinkage and selection operator. PLOS One. 2023;18:e0266267. https://doi.org/10.1371/journal.pone.0266267. Tap L, Boyé NDA, Hartholt KA, van der Cammen TJM, Mattace-Raso FUS. Association of estimated glomerular filtration rate with muscle function in older persons who have fallen. Age Ageing. March 2018;47:269–74, https://doi.org/10.1093/ageing/afx180. Kao PY, Yeh HC, Hsia YF, Hsiao YL, Wang JS, Chang DR, et al. Paradoxical mortality of high estimated glomerular filtration rate reversed by 24-h urine creatinine excretion rate adjustment: Sarcopenia matters. J Cachexia Sarcopenia Muscle. 2022;13:1704-16. doi:10.1002/jcsm.12951. Juraschek SP, Daya N, Appel LJ, Miller ER, Matsushita K, Michos ED, et al. Subclinical cardiovascular Disease and Fall risk in Older Adults: Results from the Atherosclerosis Risk in Communities Study. J Am Geriatr Soc. 2019;67:1795-802. doi:10.1111/jgs.16041. Inoue S, Otaka Y, Mori N, Matsuura D, Tsujikawa M, Kawakami M, et al. Blind spots in hospital fall prevention: Falls in stroke patients occurred not only in those at a high risk of falling. J Am Med Dir Assoc. 2024;25:160–166.e1. doi:10.1016/j.jamda.2023.10.034. Cordani C, Battel I. Do implementation interventions improve evidence-based care in acute stroke settings? A Cochrane Review summary with commentary. NeuroRehabilitation. 2024;54:343-6. doi:10.3233/NRE-246002. Hoyer EH, Young DL, Zhang C, Colantuoni E, Ghobadi K. Falls in hospitals: Challenging traditional risk assessments with new insights into patient mobility. J Adv Nurs. Published online February 28, 2025. 2025;81:5824–30. doi:10.1111/jan.16866. Goddard K, Roudsari A, Wyatt JC. Automation bias: A systematic review of frequency, effect mediators, and mitigators. J Am Med Inform Assoc. 2012;19:121–7. Cabitza F, Rasoini R, Gensini GF. Unintended consequences of machine learning in medicine. JAMA. 2017;318:517–8. Additional Declarations No competing interests reported. Supplementary Files SupplementaryFile.docx File format: .docx Title of data: Supplementary Table S1 Proportion of missing values for each predictor before imputation; Supplementary Table S2 Variance inflation factors (VIFs) with and without height and weight Description of data AdditionalFile.docx Additional File File name: Additional File Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 15 Apr, 2026 Reviews received at journal 16 Mar, 2026 Reviewers agreed at journal 16 Mar, 2026 Reviewers agreed at journal 15 Mar, 2026 Reviewers agreed at journal 13 Mar, 2026 Reviews received at journal 02 Mar, 2026 Reviewers agreed at journal 08 Feb, 2026 Reviewers invited by journal 20 Jan, 2026 Editor invited by journal 12 Jan, 2026 Editor assigned by journal 09 Jan, 2026 Submission checks completed at journal 09 Jan, 2026 First submitted to journal 07 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8540709","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":578141293,"identity":"87c091c1-651a-47e2-9eff-d480dedae125","order_by":0,"name":"Onishi Tatsuki","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzElEQVRIiWNgGAWjYBACAwglwcAPohIKiNVyAKhFsgGkxYB4LUDGASRL8QJz9h7j1x8qLPKNz69O/PDAgEGeX+wAfi2WPWfMLA6ckbDcduPtZgmgwwxnzk4g4LAbOWYGB9skDMxunN0A0pJgcJtYLcYzzm7+QawW4wcgLQb8vduIs8Wy51gZw5kzEgYSN3i3WSQYSBD2izl78+YPFRV1Bvz9Zzff/FFhI88vTUALELBJgCkJsEoJgspBgPkDmOI/QJTqUTAKRsEoGIEAAJeoRVkFK/y3AAAAAElFTkSuQmCC","orcid":"","institution":"Shiga University","correspondingAuthor":true,"prefix":"","firstName":"Onishi","middleName":"","lastName":"Tatsuki","suffix":""},{"id":578141294,"identity":"7d9ac6fc-e7d7-4b89-9c06-1b3c236a82b6","order_by":1,"name":"Tatsuyoshi Ikenoue","email":"","orcid":"","institution":"Shiga University","correspondingAuthor":false,"prefix":"","firstName":"Tatsuyoshi","middleName":"","lastName":"Ikenoue","suffix":""},{"id":578141295,"identity":"b915f209-fd45-4e2f-abf3-761580ea6951","order_by":2,"name":"Norihide Itoh","email":"","orcid":"","institution":"Saiseikai Moriyama Municipal Hospital","correspondingAuthor":false,"prefix":"","firstName":"Norihide","middleName":"","lastName":"Itoh","suffix":""},{"id":578141296,"identity":"16f15933-a0dd-4d26-a698-feb1c8b60e5a","order_by":3,"name":"Takumi Nishioka","email":"","orcid":"","institution":"Saiseikai Moriyama Municipal Hospital","correspondingAuthor":false,"prefix":"","firstName":"Takumi","middleName":"","lastName":"Nishioka","suffix":""},{"id":578141297,"identity":"c5f69dfb-04cf-4c98-80cf-1cb5a2415bf8","order_by":4,"name":"Keima Nagasaka","email":"","orcid":"","institution":"Saiseikai Moriyama Municipal Hospital","correspondingAuthor":false,"prefix":"","firstName":"Keima","middleName":"","lastName":"Nagasaka","suffix":""},{"id":578141298,"identity":"439a055f-9d13-4115-ad51-f9f9e6e27b9d","order_by":5,"name":"Ryo Okochi","email":"","orcid":"","institution":"Saiseikai Moriyama Municipal Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ryo","middleName":"","lastName":"Okochi","suffix":""},{"id":578141299,"identity":"16989c8f-9c9b-481f-9ac3-79094744066a","order_by":6,"name":"Haru Adachi","email":"","orcid":"","institution":"Saiseikai Moriyama Municipal Hospital","correspondingAuthor":false,"prefix":"","firstName":"Haru","middleName":"","lastName":"Adachi","suffix":""},{"id":578141300,"identity":"60ec554e-f56f-44cc-be52-71dc60338b0c","order_by":7,"name":"Naoko Matsuo","email":"","orcid":"","institution":"Saiseikai Moriyama Municipal Hospital","correspondingAuthor":false,"prefix":"","firstName":"Naoko","middleName":"","lastName":"Matsuo","suffix":""},{"id":578141301,"identity":"8652a17c-795a-4b21-a35a-48a67133d60d","order_by":8,"name":"Yoshiya Ueno","email":"","orcid":"","institution":"Saiseikai Moriyama Municipal Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yoshiya","middleName":"","lastName":"Ueno","suffix":""}],"badges":[],"createdAt":"2026-01-07 11:23:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8540709/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8540709/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100929837,"identity":"890f7602-1615-4427-9bcd-8ebead2cf99f","added_by":"auto","created_at":"2026-01-23 00:39:08","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":25808,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/6bf51334f047558465b8ee0d.docx"},{"id":100929864,"identity":"2e6fbd3d-c777-4e71-9392-edc4dbf20bd0","added_by":"auto","created_at":"2026-01-23 00:39:10","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":389888,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/59093ff4df7b257d20c8aa3a.docx"},{"id":100929840,"identity":"fe5fbb18-9789-41e8-8248-a7a1c0505825","added_by":"auto","created_at":"2026-01-23 00:39:08","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71145,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/7d37594cc103c9de60182d87.docx"},{"id":100929834,"identity":"5f0f791a-e78b-499a-ab16-a5c7c6cb1b6b","added_by":"auto","created_at":"2026-01-23 00:39:07","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":357059,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/50b3593474ce01b9da181b3e.docx"},{"id":100951347,"identity":"0c8eba9d-1934-419f-8198-4a016b4bfc0f","added_by":"auto","created_at":"2026-01-23 07:10:31","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":409270,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/3be229b2b601cceac2acd27d.docx"},{"id":100930015,"identity":"2efc8ff9-672d-4ab1-91e7-f30ba2187b63","added_by":"auto","created_at":"2026-01-23 00:39:26","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":623735,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/b6dd9660b6a540a673feedfb.docx"},{"id":100951762,"identity":"f05f15f0-0eb5-42c7-9b8a-3d268bb5bf74","added_by":"auto","created_at":"2026-01-23 07:11:12","extension":"json","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11140,"visible":true,"origin":"","legend":"","description":"","filename":"6fa28115512d40ff97666410ac5315c4.json","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/1618a367c630a01e9d6dea38.json"},{"id":100929878,"identity":"49ead991-ed8f-4f88-9ec1-0193e7e16d8d","added_by":"auto","created_at":"2026-01-23 00:39:12","extension":"docx","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":21536,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFile.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/bf95d13780a38a80ec91c862.docx"},{"id":100930051,"identity":"d78cbf42-b63a-48f4-bae0-19189afdc1e4","added_by":"auto","created_at":"2026-01-23 00:39:27","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128591,"visible":true,"origin":"","legend":"","description":"","filename":"6fa28115512d40ff97666410ac5315c41enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/1f7187822d963c31402d94a5.xml"},{"id":100929947,"identity":"c99a3ddd-5c07-4998-bc90-f22f8af03b46","added_by":"auto","created_at":"2026-01-23 00:39:13","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10019,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/3c9bd3479592c112e091bee5.png"},{"id":100929735,"identity":"f823290d-4741-46f4-b5f5-9543fbf8c595","added_by":"auto","created_at":"2026-01-23 00:38:55","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":375432,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/4a50ffaa2453c92e3c4e41d5.png"},{"id":100929950,"identity":"581582dd-0b69-4960-9aff-d664488e2f85","added_by":"auto","created_at":"2026-01-23 00:39:13","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342585,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/8e69e63c149ed9539bf02f7a.png"},{"id":100929994,"identity":"b511130d-507b-4667-a29b-192df0f1f5de","added_by":"auto","created_at":"2026-01-23 00:39:24","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":394809,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/1394e09d3c97238542e44430.png"},{"id":100929773,"identity":"a631272f-d7b0-47f3-8819-34f3dfdbc223","added_by":"auto","created_at":"2026-01-23 00:39:00","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":609396,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/987d40db2445b9739de05b09.png"},{"id":100929788,"identity":"2c3d6e10-3e95-40a4-ab42-e977fe087848","added_by":"auto","created_at":"2026-01-23 00:39:01","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5844,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/2df5900e44d5cdee6a49935c.png"},{"id":100929933,"identity":"8d8fd47b-caba-49a1-aae8-e94bc8a669c6","added_by":"auto","created_at":"2026-01-23 00:39:13","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69737,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/f57067d614d0395b7de49b14.png"},{"id":100929806,"identity":"cab6b154-80d7-49f4-b44d-04aa63e15200","added_by":"auto","created_at":"2026-01-23 00:39:03","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69390,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/a4ceb4975c08bb881985d69d.png"},{"id":100929956,"identity":"49d48ff8-78f3-4d6c-b3b8-42578827b17b","added_by":"auto","created_at":"2026-01-23 00:39:15","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":77689,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/c45cebea364c3c601d1b5b5c.png"},{"id":100929856,"identity":"0b854b20-0164-40ee-9f05-dc5af8461c45","added_by":"auto","created_at":"2026-01-23 00:39:09","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":105172,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/5144d9c56f208da630a08c3e.png"},{"id":100929857,"identity":"5b8da9c7-6470-488b-9cf4-24dcb936aab9","added_by":"auto","created_at":"2026-01-23 00:39:09","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128785,"visible":true,"origin":"","legend":"","description":"","filename":"6fa28115512d40ff97666410ac5315c41structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/49a2666d74bcd0a717307a06.xml"},{"id":100929858,"identity":"337d3ab0-2ef3-4cb7-b451-952c43b8b998","added_by":"auto","created_at":"2026-01-23 00:39:09","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":140829,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/19796864a57ba4b47b54c44d.html"},{"id":100929862,"identity":"ab3ea9db-275f-4b64-b68c-406e6670b7fe","added_by":"auto","created_at":"2026-01-23 00:39:10","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":21466,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram of participants\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/b650e12f6af16e3ca80aa03d.jpg"},{"id":100929786,"identity":"0a883641-3480-4e4e-bd6c-539825df9a2a","added_by":"auto","created_at":"2026-01-23 00:39:01","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":79657,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram of participants\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/bae963ae5d0b13aa4717da1a.jpg"},{"id":100929740,"identity":"c1d2d068-7194-45eb-ad7b-dd33e2a25df0","added_by":"auto","created_at":"2026-01-23 00:38:57","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":65578,"visible":true,"origin":"","legend":"\u003cp\u003e(\u003cstrong\u003eA\u003c/strong\u003e) Fall-score distribution in all cases and by outcome (fall = 0/1). (\u003cstrong\u003eB\u003c/strong\u003e) ROC curve of the RF-derived fall score. (\u003cstrong\u003eC\u003c/strong\u003e) Feature importance ranking (top contributors highlighted). (\u003cstrong\u003eD\u003c/strong\u003e) Conducted ratio of fallers. \u003cem\u003eNote:\u003c/em\u003ePanels \u003cstrong\u003eA\u003c/strong\u003e and \u003cstrong\u003eB\u003c/strong\u003e illustrate distributional overlap and discrimination, panel \u003cstrong\u003eC\u003c/strong\u003e lists the top features, and panel \u003cstrong\u003eD\u003c/strong\u003esummarises the incidence by intervention\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/f60b0214e6e583ea0b0ed6a8.jpg"},{"id":100929737,"identity":"fb5495a4-f326-478d-b1f6-c64b5518acfb","added_by":"auto","created_at":"2026-01-23 00:38:56","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":80478,"visible":true,"origin":"","legend":"\u003cp\u003e\u0026nbsp;(\u003cstrong\u003eA\u003c/strong\u003e) Contingency tables of fall occurrence and supervision/intervention type. (\u003cstrong\u003eB\u003c/strong\u003e) Heat map of fall scores and supervision/intervention. The analysis indicated varied intervention rates across fall-score categories, suggesting potential opportunities for more systematic score-based intervention allocation. Training: occurrence with the presence of a therapist during training. Adherence: occurrence with adherence to bed-rest restrictions. Assistance: occurrence with assistance provided\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/0773d7f9eaa9e32476d43d2c.jpg"},{"id":100929883,"identity":"b16d09f2-2566-419c-99bc-8bc05e083b37","added_by":"auto","created_at":"2026-01-23 00:39:12","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":84930,"visible":true,"origin":"","legend":"\u003cp\u003eManual six-item fall-risk screening score. (\u003cstrong\u003eA\u003c/strong\u003e) Distribution of the equal-weight score stratified by outcome; bins are integer ticks. (\u003cstrong\u003eB\u003c/strong\u003e) ROC curve of the equal-weight score (AUC: 0.668). (\u003cstrong\u003eC\u003c/strong\u003e) Calibration plot of the equal-weight score (predicted probability vs. observed fall proportion across score-based risk groups). (\u003cstrong\u003eD\u003c/strong\u003e) Distribution of the integer-weight score (weights: 3 for the motor FIM score, 2 for PT-INR, 1 for others). (\u003cstrong\u003eE\u003c/strong\u003e) ROC curve of the integer-weight score (AUC: 0.695). (\u003cstrong\u003eF\u003c/strong\u003e) Calibration plot of the integer-weight score (predicted probability vs. observed fall proportion across score-based risk groups). Vertical dashed lines indicate example operating points (high-sensitivity cut-offs of ≥2 points for the equal-weight score and ≥4 points for the integer-weight score)\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/c076e7068260badb627fe802.jpg"},{"id":100953977,"identity":"21fba6a5-9488-4cfb-affe-a6a39b2dbee5","added_by":"auto","created_at":"2026-01-23 07:23:05","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1511835,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/fb7654af-4527-4908-bc19-d14a059de1d4.pdf"},{"id":100929739,"identity":"bdd24a6a-1340-4ad8-8078-111d8088c91b","added_by":"auto","created_at":"2026-01-23 00:38:57","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":21536,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFile format: \u003c/strong\u003e.docx\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTitle of data: \u003c/strong\u003eSupplementary Table S1 Proportion of missing values for each predictor before imputation; Supplementary Table S2 Variance inflation factors (VIFs) with and without height and weight\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDescription of data\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"SupplementaryFile.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/db53b79012057e8998da0b09.docx"},{"id":100930073,"identity":"ce7af292-8da7-4b69-bb68-282583edcafa","added_by":"auto","created_at":"2026-01-23 00:39:28","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":21511,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAdditional File\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFile name: \u003c/strong\u003eAdditional File\u003c/p\u003e","description":"","filename":"AdditionalFile.docx","url":"https://assets-eu.researchsquare.com/files/rs-8540709/v1/d8ef96177d4f954bd0673b15.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and internal validation of a machine learning–based prediction model and simplified screening score for in-hospital falls: a retrospective cohort study","fulltext":[{"header":"Background","content":"\u003cp\u003eHospital falls represent an important safety concern affecting patient outcomes and healthcare resource utilisation [\u003cspan additionalcitationids=\"CR2 CR3 CR4\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Current fall-prevention interventions, including assistance, supervision, and activity-level adjustments, are predominantly experience-based and individually tailored [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Nevertheless, the subjective and personalised nature of these interventions may lead to inconsistent implementation and suboptimal resource allocation, potentially allowing the occurrence of falls despite preventive efforts [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe complexity of fall-risk assessment stems from multiple factors, including patient demographics, functional status, medical conditions, and environmental factors [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Conventional approaches, which rely heavily on clinical judgement and standardised assessment tools, may not capture the nuanced interactions among various risk factors for falls [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Furthermore, the lack of objective and quantifiable risk assessment tools limits the ability of healthcare providers to systematically prioritise interventions and effectively allocate limited human resources [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMachine learning (ML) approaches offer promising solutions for developing sophisticated risk prediction models that can integrate multiple variables simultaneously [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. In particular, random forest (RF) classifiers are effective in healthcare prediction tasks owing to their ability to handle complex interactions among variables and provide interpretable feature importance rankings [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Additionally, advanced explainability techniques, including SHapley Additive exPlanations (SHAP) analysis, provide insights into how individual patient characteristics contribute to fall-risk predictions [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eNonetheless, operationalising individual predictions remains a challenge. SHAP analysis explains \u003cem\u003ewhy\u003c/em\u003e a specific patient is at risk but generates unique complex profiles that hinder the standardisation of care protocols. To bridge this gap, latent class analysis (LCA) is required to reduce the individual complexity into manageable patient-level phenotypes. LCA facilitates the development of practical standardised interventions beyond simple risk quantification by clustering heterogeneous high-risk patients into recognisable clinical syndromes (e.g. functional vs. metabolic types).\u003c/p\u003e \u003cp\u003eThe present study aimed to develop and internally validate a comprehensive ML-based fall-risk prediction model, characterise latent patient subgroups through SHAP analysis and LCA, and derive a simplified bedside screening score for standardised fall-prevention interventions in hospital settings [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eReporting guideline\u003c/h2\u003e \u003cp\u003eThis study was reported in accordance with the TRIPOD\u0026thinsp;+\u0026thinsp;AI statement and the original TRIPOD 2015 reporting guidelines, following the EQUATOR Network recommendations [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy design and setting\u003c/h3\u003e\n\u003cp\u003e This single-centre retrospective cohort study was conducted from January 2023 to September 2024 using routinely collected electronic medical records at Saiseikai Moriyama Municipal Hospital (Shiga, Japan), a 199-bed community hospital with 51 beds in acute-care wards, 100 beds in recovery-phase rehabilitation wards, and 48 beds in chronic/long-term care wards. Eligible inpatients with examination records during the study period were included.\u003c/p\u003e\n\u003ch3\u003eParticipants\u003c/h3\u003e\n\u003cp\u003eAll hospital admissions during the study period were screened for patients who underwent at least one Functional Independence Measure (FIM) assessment and a panel of routine laboratory tests recorded in the hospital information system. Only adults aged\u0026thinsp;\u0026ge;\u0026thinsp;18 years were considered eligible for study inclusion. When patients were admitted multiple times during the study period, only the first (index) admission was considered to avoid within-patient clustering. Additionally, records with missing values for key demographic or functional variables (e.g., sex, age, motor FIM score, and cognitive FIM score) were excluded because these predictors were essential and we chose not to impute them. Out of 633 individuals initially identified, 22 patients with incomplete data were excluded, leaving 611 participants for the analysis.\u003c/p\u003e\n\u003ch3\u003eOutcome\u003c/h3\u003e\n\u003cp\u003eThe primary outcome was the occurrence of in-hospital falls at index admission. Based on the World Health Organization criteria, a fall was defined as an event that caused a person to come to rest inadvertently on the ground, floor, or other lower levels. Falls were identified from incident reports, nursing documentations, or electronic safety reports recorded in the hospital information system. Outcome assessors were unaware of the model\u0026rsquo;s predictions. For patients who experienced multiple falls, only the first fall at admission was considered for the analysis, and the number of days from admission to the first fall was calculated.\u003c/p\u003e\n\u003ch3\u003ePredictors and data collection\u003c/h3\u003e\n\u003cp\u003eRoutinely collected demographic (sex and age), anthropometric (height and weight), functional, and laboratory variables were regarded as candidate predictors. Body mass index (BMI) was calculated as weight divided by height squared (kg/m\u0026sup2;). Functional status was assessed using the FIM, and both motor and cognitive FIM scores were recorded repeatedly during hospitalisation as part of routine care. Based on observational data on patients\u0026rsquo; activities of daily living in the ward, the FIM scores were examined every 4 weeks from admission through a discussion between the primary physical therapist or occupational therapist and the nursing staff. The smoking index was included as a measure of cumulative tobacco exposure. Laboratory variables included serum albumin (ALB), B-type natriuretic peptide (BNP), C-reactive protein (CRP), prothrombin time\u0026ndash;international normalised ratio (PT-INR), total bilirubin (T-Bil), total cholesterol (T-CHO), creatinine (CRE), estimated glomerular filtration rate (eGFR), γ-glutamyl transpeptidase (γ-GTP), fasting blood glucose (FBG), and haemoglobin (Hgb). Use of relevant medications such as antihypertensive, antidiabetic, and sedative\u0026ndash;hypnotic agents was also extracted and coded as a binary variable.\u003c/p\u003e \u003cp\u003eMeasurements of time-varying predictors (motor and cognitive FIM scores and laboratory values) were aligned with a reference time point for each patient. The reference time point was defined as the day before the first fall for fallers (patients who experienced a fall) and as the day before discharge for non-fallers. For each predictor, the most recent measurements available at or before the reference time point were used. The last observation carried forward (LOCF) approach was applied when the FIM scores or laboratory values at the reference time point were missing but earlier measurements during the same admission were available. Patients with no recorded FIM assessment at any time during hospitalisation were excluded because the FIM was considered an essential predictor and was not imputed. Variables that remained entirely missing even after applying the LOCF approach were imputed using median values within the ML pipeline.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eHandling of missing values\u003c/h2\u003e \u003cp\u003eThe following strategies were applied to variables with missing values even after using the LOCF approach for repeated measurements: records with missing values for sex, age, and motor or cognitive FIM score were excluded from the analysis; missing values for the smoking index were assumed to represent non-smokers and were imputed as 0; and continuous anthropometric and biochemical variables (height, weight, PT-INR, eGFR, and ALB, BNP, CRP, T-Bil, T-CHO, CRE, γ-GTP, FBG, and Hgb levels) were imputed using the cohort median. The normality of distributions was not formally assessed prior to median imputation, which represents a limitation. The proportion of missing values for each laboratory predictor before imputation is summarised descriptively in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e in the Additional File.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSample size considerations\u003c/h3\u003e\n\u003cp\u003eTo avoid overfitting, effective model complexity was constrained relative to 120 observed fall events, consistent with the TRIPOD-AI recommendations and the formal sample size criteria for prediction models described by Riley et al., which emphasised adequate events per parameter, minimal optimism, and stable estimation of outcome proportions [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eMulticollinearity assessment\u003c/h3\u003e\n\u003cp\u003eFor numerical explanatory variables (excluding the fall outcome), variance inflation factors (VIF) within the regression-based pipeline were calculated to assess potential multicollinearity among candidate predictors. Height and weight showed substantial collinearity with BMI (VIF\u0026thinsp;\u0026gt;\u0026thinsp;10) and were consequently excluded from the final set of predictors. The VIF values before and after height and weight removal are provided in Supplementary Table S2 in the Additional File.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eModel development and performance evaluation\u003c/h2\u003e \u003cp\u003eThe performance of the fall-prediction model was evaluated by constructing receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC). This model was based on an RF classifier, and the target variable was the binary fall outcome (0 or 1). ROC curves were plotted using predicted fall probabilities (fall scores), with the AUC serving as a quantitative measure of discriminative ability. As the fall outcome was moderately imbalanced (19.6% of fall events), we evaluated whether class-weight adjustment was necessary. However, the imbalance was judged to be sufficiently mild, and the internal validation indicated no meaningful benefit from weighting. Therefore, we did not apply class weighting and trained the RF classifier with class_weight\u0026thinsp;=\u0026thinsp;none, ensuring that the model complexity aligned with the available number of events and followed contemporary recommendations for prediction model development. This decision was also consistent with the observed stability of ROC curve performance in cross-validation. In addition to discrimination, overall prediction accuracy was quantified using the Brier score, and calibration was assessed numerically by estimating the calibration intercept and slope from a logistic recalibration model with the logit of the predicted probabilities as the sole predictor.\u003c/p\u003e \u003cp\u003eTwo complementary methods were employed to determine the statistical significance of the AUC\u0026mdash;namely, logistic regression and bootstrap resampling. A logistic regression model was fitted with the fall outcome as the dependent variable, and a bootstrap procedure with 1,000 iterations was performed to recalculate the AUC under resampling. The proportion of iterations with AUC\u0026thinsp;\u0026le;\u0026thinsp;0.5 (i.e., equivalent to random classification) was used to estimate a one-sided \u003cem\u003ep\u003c/em\u003e-value. ROC curves were visualised, and the AUC was reported as the primary indicator of model performance. The RF classifier was implemented using scikit-learn and default settings.\u003c/p\u003e \u003cp\u003eIn addition to discrimination, the calibration of the RF model was analysed using calibration plots. The predicted fall probabilities were grouped into deciles, and the observed fall proportion in each group was plotted against the mean predicted probability, with the 45-degree line representing perfect calibration.\u003c/p\u003e \u003cp\u003eThe model was based on an RF classifier implemented using Python 3.12 (Python Software Foundation, Delaware, the United States of America) and scikit-learn version 1.5.2, and the target variable was the binary fall outcome (0 or 1). Standard hyperparameter settings were applied, unless otherwise specified: n_estimators\u0026thinsp;=\u0026thinsp;100, random state\u0026thinsp;=\u0026thinsp;42, max_depth\u0026thinsp;=\u0026thinsp;None, max_features = \"sqrt\", class_weight\u0026thinsp;=\u0026thinsp;None, and bootstrap\u0026thinsp;=\u0026thinsp;True. These settings corresponded to the default configuration in scikit-learn.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eFall-score distribution analysis\u003c/h2\u003e \u003cp\u003eThe distribution of predicted fall scores was visualised using histograms stratified by outcome status. For patients who experienced a fall (fallers), the fall-score distribution was plotted to examine central tendency and skewness among high-risk individuals. Additionally, overlaid histograms for fallers and non-fallers were created using 20 bins to visually assess the separation between the groups. The RF classifier was trained without class weighting (class_weight\u0026thinsp;=\u0026thinsp;None), reflecting our judgement that the class imbalance was mild. Thus, the resultant fall-score distribution represented the unweighted separation between fallers and non-fallers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eFeature importance analysis\u003c/h2\u003e \u003cp\u003eFeature importance scores after model training were extracted to evaluate the relative contribution of each predictor to the RF model. The predictors were ranked in descending order of importance and visualised using a horizontal bar plot, with the most influential features displayed at the top.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eCare-related variables analysis\u003c/h2\u003e \u003cp\u003eTo explore the associations between care-related interventions and falls, 2\u0026times;2 contingency tables comparing the fall outcome with four categorical variables (presence of a therapist during training, adherence to bed-rest restrictions, assistance provided, and monitoring status) were created. Pairwise statistical comparisons between intervention types were performed, with Bonferroni correction being applied to adjust for multiple testing.\u003c/p\u003e \u003cp\u003eTo determine whether the implementation of preventive interventions differed by the predicted risk, patients who experienced a fall (fall\u0026thinsp;=\u0026thinsp;1) were analysed, and their fall scores were categorised into three groups: \u0026lt;0.33, 0.33 to \u0026lt;\u0026thinsp;0.66, and \u0026ge;\u0026thinsp;0.66. The proportion of patients who received each preventive intervention was calculated for each group. These patterns were visualised using stacked bar charts, heat maps, and 100% stacked bar charts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eLCA\u003c/h2\u003e \u003cp\u003eLCA was conducted using a Gaussian mixture model to identify latent patient subgroups based on fall-related features. All numeric predictors, except the fall outcome, were standardised (z-scores). The remaining missing values were imputed using median values, and rows with persistent missing data were excluded.\u003c/p\u003e \u003cp\u003eThe optimal number of clusters was determined using the elbow method, and the resultant clusters were characterised by summarising descriptive statistics, calculating the feature importance for cluster classification, comparing key predictors among clusters, and analysing the fall scores by cluster. This approach was consistent with the use of clustering methods in previous studies to identify clinically meaningful patient subgroups in risk prediction research [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eSHAP analysis for model explainability\u003c/h2\u003e \u003cp\u003eTo enhance model explainability, SHAP analysis was conducted to quantify the contribution of individual features to RF predictions. For the SHAP analysis, we used an RF regressor with the same predictor set and similar hyperparameters as the classification model, which was trained to approximate the predicted fall probability. This allowed us to decompose individual predictions into additive feature contributions on a probability scale. Eighteen numerical features, including demographics, anthropometric measures, FIM scores, and laboratory values, were included as predictors. The dataset was divided into training (80%) and test (20%) sets. An RF regressor with the same predictor set and similar hyperparameters was trained on the training data to predict fall probability. SHAP values were computed for the training data and visualised for the test data to interpret feature contributions at both global and individual levels.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eSensitivity analysis\u003c/h2\u003e \u003cp\u003eTwo main sensitivity analyses were conducted to assess robustness. For the primary analysis, examination findings up to the day before the fall event were used, and missing data were imputed. In the first sensitivity analysis, examination findings from the day of the fall event were instead used as the most recent results. In the second sensitivity analysis, records with missing examination data were excluded, and a complete-case analysis was performed without imputation. For each sensitivity analysis, ROC curves and histograms of fall scores were generated and compared with those of the primary analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eDevelopment of the screening score\u003c/h2\u003e \u003cp\u003eWe sought a parsimonious, hand-calculated screening score. Candidate predictors were routinely available numeric variables, and BMI was not excluded \u003cem\u003ea priori\u003c/em\u003e in this step. For operational reasons, age, sex, and smoking status were not considered for bedside scores, whereas laboratory markers and FIM scores were considered. A fixed pipeline (median imputation, standardisation, class-weighted logistic regression) was evaluated under stratified five-fold cross-validation (random state\u0026thinsp;=\u0026thinsp;42). For each subset size \u0026#119896;, all combinations were exhaustively scored, and the smallest subset that met our target (for sensitivity targeting: existence in every fold of an ROC operating point with sensitivity of \u0026ge;\u0026thinsp;0.80 and specificity of \u0026ge;\u0026thinsp;0.50) were selected; ties were resolved by median AUC.\u003c/p\u003e \u003cp\u003eFor the selected subset, each variable was binarised via univariate ROC with Youden\u0026rsquo;s \u0026#119869; on training folds [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. The median threshold and direction (\u0026ge;\u0026thinsp;or \u0026le;) across folds were recorded. Subsequently, an equal-weight score (+\u0026thinsp;1 per satisfied condition) and an integer-weight score were constructed using L1-penalised logistic regression on the binarised variables, scaling positive coefficients to small integers [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Performance was summarised by cross-validated AUC (95% confidence interval [CI]) and two-sided bootstrap \u003cem\u003ep\u003c/em\u003e-value under H0 (AUC\u0026thinsp;=\u0026thinsp;0.5).\u003c/p\u003e \u003cp\u003eThe final six-item subset comprised the motor FIM score, BNP level, PT-INR, eGFR, Hgb level, and BMI. For the final integer-weight score, calibration was assessed by constructing calibration plots of predicted versus observed fall risks across score-based risk groups.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\n \u003ch2\u003eStudy participants\u003c/h2\u003e\n \u003cp\u003eOut of 633 individuals identified during the study period, 22 patients with incomplete data on key demographic or functional variables were excluded, leaving 611 participants for the analysis (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e). Overall, 120 (19.6%) participants experienced at least one fall during hospitalisation. The median time from admission to the first fall was 32.0 (interquartile range: 19.0\u0026ndash;54.0) days.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\n \u003ch2\u003eParticipant characteristics\u003c/h2\u003e\n \u003cp\u003eThe baseline characteristics of the overall cohort stratified by fall status are summarised in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. Fallers were slightly older and had lower motor and cognitive FIM scores at the reference time point than non-fallers. Fallers also exhibited lower weight and BNP level; other laboratory parameters showed minimal differences between the groups. Antihypertensive and sedative\u0026ndash;hypnotic agents were more frequently used among fallers (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eCharacteristics of the participants\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOverall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFallers\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNon-fallers\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e-value\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003en\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e611\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e120\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e491\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMale, \u003cem\u003en\u003c/em\u003e (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e229 (37.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e48 (40.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e181 (36.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.595\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e78.6 (12.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e80.6 (11.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e78.1 (13.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.046\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHeight, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e155.6 (10.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e154.6 (10.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e155.9 (10.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.224\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWeight, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e52.2 (12.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50.3 (11.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e52.6 (12.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.050\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBMI, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21.4 (3.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21.0 (4.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21.5 (3.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.176\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMotor FIM score, median [Q1, Q3]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e46.0 [25.0, 61.0])\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e39.5 [24.0, 49.0]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e48.0 [25.5, 63.0]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCognitive FIM score, median [Q1, Q3]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25.0 [16.0, 33.0]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22.0 [16.0, 28.0]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e25.0 [16.0, 34.0]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSmoking, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e218.2 (1167.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e180.2 (977.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e227.5 (1209.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.652\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eALB, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.6 (0.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.6 (0.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.6 (0.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.383\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBNP, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e92.9 (62.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e82.9 (51.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e95.3 (65.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.025\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCRP, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.4 (2.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.4 (2.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.4 (2.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.911\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePT-INR, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.1 (0.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.0 (0.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.1 (0.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.351\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT-Bil, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7 (0.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7 (0.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7 (0.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.338\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT-CHO, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e165.1 (22.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e164.9 (23.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e165.1 (22.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.926\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCRE, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.1 (1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.0 (0.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1.1 (1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.675\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eeGFR, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e61.9 (20.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e61.3 (17.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e62.1 (20.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.660\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026gamma;-GTP, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e37.2 (45.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e34.5 (30.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e37.9 (48.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.327\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFBG, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e132.6 (25.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e132.8 (33.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e132.5 (23.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.922\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHgb, mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11.8 (1.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11.8 (1.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11.8 (1.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.843\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAntihypertensive agents, \u003cem\u003en\u003c/em\u003e (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e29 (4.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12 (10.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e17 (3.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAntidiabetic agents, \u003cem\u003en\u003c/em\u003e (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e138 (22.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e32 (26.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e106 (21.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.284\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSedative\u0026ndash;hypnotic agents, \u003cem\u003en\u003c/em\u003e (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e233 (38.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e59 (49.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e174 (35.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003eFIM, Functional Independence Measure; BMI, body mass index; ALB, serum albumin; BNP, B-type natriuretic peptide; CRP, C-reactive protein, PT-INR, prothrombin time\u0026ndash;international normalised ratio, T-Bil, total bilirubin; T-CHO, total cholesterol; CRE, creatinine; eGFR, estimated glomerular filtration rate; \u0026gamma;-GTP, \u0026gamma;-glutamyl transpeptidase; FBG, fasting blood glucose; Hgb, haemoglobin; SD, standard deviation; IQR, interquartile range\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\n \u003ch2\u003eFall-score distribution\u003c/h2\u003e\n \u003cp\u003eThe distribution of predicted fall scores among fallers skewed towards higher values, indicating that the model assigned higher estimated probabilities to several fall cases (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eA). However, the fall-score distribution notably overlapped between fallers and non-fallers, highlighting the residual variability that was not fully captured by the model (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eA).\u003c/p\u003e\n \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\n \u003ch2\u003eModel performance\u003c/h2\u003e\n \u003cp\u003eThe RF model achieved an AUC of 0.96 (95% CI: 0.931\u0026ndash;0.981, bootstrap \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), indicating excellent discriminative ability (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eB). The average AUC from cross-validation was 0.96, suggesting stable performance across resamples. The calibration plot showed good agreement between the predicted and observed fall probabilities, although the risk was slightly underestimated in the group with the highest predicted risk. The Brier score for the RF model was 0.050, and the calibration slope and intercept from a logistic recalibration model were 2.02 and 0.64, respectively, indicating overall good prediction accuracy but some overestimation of risk and overfitting at higher predicted probabilities.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\n \u003ch2\u003eFeature importance\u003c/h2\u003e\n \u003cp\u003eAmong predictors, the motor FIM score at admission was the most important feature, followed by the cognitive FIM score, BMI, age, and several laboratory markers such as the PT-INR, eGFR, and CRE, Hgb, and ALB levels. These predictors were ranked (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eC).\u003c/p\u003e\n \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e\n \u003ch2\u003eCare-related interventions\u003c/h2\u003e\n \u003cp\u003eThe analysis of care-related variables using contingency tables (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003eA) revealed varying patterns of preventive intervention implementation. Adherence to bed-rest restrictions had significantly higher implementation rates than assistance (43.3% vs. 3.33%, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and supervision (43.3% vs. 15.8%, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Assistance was implemented less frequently than supervision (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.006) (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eD). The proportion of fallers under each intervention category is presented in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eD, whereas intervention-by-outcome contingency tables are shown in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003eA, clarifying how often each intervention was implemented among fallers and non-fallers.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e\n \u003ch2\u003eFall score\u0026ndash;based intervention patterns\u003c/h2\u003e\n \u003cp\u003eContingency tables of fall occurrence and supervision/intervention type are presented in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003eB. When the participants were stratified into three groups by fall scores (\u0026lt;\u0026thinsp;0.33, 0.33 to \u0026lt;\u0026thinsp;0.66, and \u0026ge;\u0026thinsp;0.66), the heatmap revealed non-uniform patterns of intervention implementation across risk levels (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003eB). Some high-risk patients did not receive intensive interventions, whereas certain lower-risk patients did, suggesting a potential misalignment between the predicted risk and actual preventive care.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e\n \u003ch2\u003eLCA\u003c/h2\u003e\n \u003cp\u003eInspection using the elbow method suggested three distinct latent classes. The resultant clusters represented patient phenotypes with different combinations of clinical and functional parameters. For instance, one cluster comprised patients with markedly impaired motor FIM score and high fall score, whereas another cluster included patients with relatively preserved functional status and low fall score. These patterns supported the presence of heterogeneous fall-risk profiles in hospitalised populations.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e\n \u003ch2\u003eSHAP analysis\u003c/h2\u003e\n \u003cp\u003eThe SHAP analysis provided detailed insights into the feature contributions at the individual level. Patients with lower motor FIM score and somewhat unexpectedly younger age tended to have higher SHAP values, indicating an increased predicted fall risk. Higher FBG and \u0026gamma;-GTP levels, PT-INR, and eGFR were also associated with an increased fall risk in the model. These findings revealed the complex and sometimes counterintuitive relationships between predictors and predicted fall probabilities.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec29\" class=\"Section2\"\u003e\n \u003ch2\u003eSensitivity analysis\u003c/h2\u003e\n \u003cp\u003eThe results when using laboratory values from the day of the fall event are presented in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003eA\u0026ndash;B, and the complete-case analysis is illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003eC\u0026ndash;D. Sensitivity analyses yielded fall-score distributions and ROC curves that were broadly similar to those in the primary analysis. Using laboratory findings from the day of the fall event instead of the previous day or restricting the analysis to complete cases without imputation did not materially change the discrimination performance or overall distributional patterns of fall scores.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003ePrediction of the six-item screening score (equal vs. integer weights)\u003c/h3\u003e\n\u003cp\u003eThe selected subset comprised six predictors: motor FIM score, BNP level, PT-INR, eGFR, Hgb level, and BMI. Binarisation thresholds (median across folds; direction in parentheses) were as follows: motor FIM score\u0026thinsp;\u0026le;\u0026thinsp;51 (\u0026le;), BNP level\u0026thinsp;\u0026le;\u0026thinsp;66.1 (\u0026le;), PT-INR\u0026thinsp;\u0026le;\u0026thinsp;1.01 (\u0026le;), eGFR\u0026thinsp;\u0026ge;\u0026thinsp;78.41 (\u0026ge;), Hgb level\u0026thinsp;\u0026ge;\u0026thinsp;11.2 (\u0026ge;), and BMI\u0026thinsp;\u0026le;\u0026thinsp;20.8 (\u0026le;) (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). The equal-weight score (+\u0026thinsp;1 per condition) achieved an AUC of 0.668 (95% CI: 0.618\u0026ndash;0.716, bootstrap \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). At high-sensitivity threshold of \u0026ge;\u0026thinsp;2 points, the sensitivity and specificity were 0.908 and 0.279, respectively (positive predictive value: 0.235, negative predictive value: 0.926, accuracy: 0.403) (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003eA\u0026ndash;B). The Youden optimal cut-off was \u0026ge;\u0026thinsp;3 points (sensitivity: 0.533, specificity: 0.703). The calibration plot (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003eC) indicated that the predicted probabilities derived from the equal-weight score aligned well with the observed fall proportions across the risk strata.\u003c/p\u003e\n\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDirection, threshold, and equal and integer weights for items\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eItem\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDirection\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eThreshold\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eEqual-weight points\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eInteger-weight points\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMotor FIM score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026le;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e51.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBNP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026le;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e66.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePT-INR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026le;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eeGFR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026ge;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e78.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHgb\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026ge;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e11.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026le;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e20.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003eFIM, Functional Independence Measure; BNP, B-type natriuretic peptide; PT-INR, prothrombin time\u0026ndash;international normalised ratio; eGFR, estimated glomerular filtration rate; Hgb, haemoglobin; BMI, body mass index\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe integer-weight score (L1 logistic-derived) assigned weights of 3 to the motor FIM score, 2 to PT-INR, and 1 to the remaining four items, yielding an AUC of 0.695 (95% CI: 0.644\u0026ndash;0.742, bootstrap \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). At high-sensitivity threshold of \u0026ge;\u0026thinsp;4 points, the sensitivity and specificity were 0.825 and 0.458, respectively (accuracy: 0.530) (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003eD\u0026ndash;E). The calibration plot (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003eF) indicated that the predicted probabilities derived from the integer-weight score aligned well with the observed fall proportions across the risk strata.\u003c/p\u003e\n\u003cp\u003eThese results verified that retaining BMI improved the coherence with the RF importance profile and that the six-item bedside score provided a practical trade-off between parsimony and discrimination suitable for repeated inward screening.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn the present study, we developed a comprehensive fall-risk prediction system using routinely collected hospital data and ML methods. Our approach integrated supervised learning for precise risk estimation with unsupervised learning for patient profiling. The RF model showed excellent discriminative performance (AUC: 0.96), providing an objective data-driven tool for identifying high-risk individuals. LCA revealed distinct patient subgroups. Our findings indicated that combining quantitative prediction with qualitative profiling could not only identify at-risk patients but also provide a basis for tailoring interventions according to clinical phenotypes.\u003c/p\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003eKey findings and clinical implications\u003c/h2\u003e \u003cp\u003eIn our study, we identified the motor FIM score as the most important predictor; this result aligns with the established understanding that mobility impairment is a primary risk factor for falls. Our SHAP analysis further refined this insight by revealing that younger patients with poorer functional status may still be at a high risk, challenging the conventional focus on age alone. Additionally, the model highlighted the potential relevance of the cognitive FIM score and several laboratory parameters, particularly BMI, eGFR, FBG, and inflammatory markers such as CRP. The observed association between higher eGFR and increased fall risk may reflect sarcopenia-related phenomena, in which individuals with lower muscle mass exhibit higher eGFR calculated from serum CRE levels [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Nevertheless, these findings require further investigation.\u003c/p\u003e \u003cp\u003eThe overlap in fall-score distribution between fallers and non-fallers underscores the multifactorial nature of falls and suggests that even a high-performing model cannot fully resolve individual-level uncertainty. Therefore, prediction models should be used as decision-support tools rather than deterministic classifiers. In this context, the six-item bedside screening score derived from the model (motor FIM score\u0026thinsp;\u0026le;\u0026thinsp;51, BNP level\u0026thinsp;\u0026le;\u0026thinsp;66.1, PT-INR\u0026thinsp;\u0026le;\u0026thinsp;1.01, eGFR\u0026thinsp;\u0026ge;\u0026thinsp;78.4, Hgb level\u0026thinsp;\u0026ge;\u0026thinsp;11.2, and BMI\u0026thinsp;\u0026le;\u0026thinsp;20.8) achieved moderate discrimination (AUC: 0.668 with equal weights and 0.695 with integer weights) while remaining sufficiently simple for routine use. High-sensitivity cut-offs (\u0026ge;\u0026thinsp;2 points for the equal-weight score and \u0026ge;\u0026thinsp;4 points for the integer-weight score) enabled repeated risk stratification during hospitalisation, potentially facilitating a timelier allocation of preventive resources and targeted supervision for patients at greatest risk [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe cognitive FIM score was one of the top predictors in the RF model and showed considerable SHAP contributions; however, it did not remain in the final six-item screening score, likely because of methodological differences between the full multivariable model and the constrained subset selection procedure. The logistic regression pipeline used for score derivation emphasises linearly separable thresholds that maximise cross-validated discrimination using a minimal number of binary rules. Although the cognitive FIM score contributed mainly through complex nonlinear interactions in the SHAP analysis, these patterns translated less efficiently into a single threshold-based rule. This implies that cognitive impairment does not merely add to the risk but acts as a \u0026lsquo;force multiplier\u0026rsquo; that compounds other physical vulnerabilities\u0026mdash;a dynamic nuance that is inherently difficult to capture with a static, single-point threshold. Consequently, the cognitive FIM score was not retained despite its importance in the ML model. This distinction illustrates the complementary roles of SHAP (model interpretability) and simplified scoring systems (operational feasibility).\u003c/p\u003e \u003cp\u003eRather than enforcing a single universal cut-off, we propose that the model-based fall scores be used as a continuous decision-support tool to contextualise FIM-based risk assessments, particularly in patients who experience falls despite not being classified to have a high risk by conventional criteria.\u003c/p\u003e \u003cdiv id=\"Sec33\" class=\"Section3\"\u003e \u003ch2\u003eClinical rationale for the inclusion of BMI, BNP level, and PT-INR in the screening score\u003c/h2\u003e \u003cp\u003eRetaining the BMI at the bedside is consistent with the model\u0026rsquo;s importance profile and clinical plausibility. Low BMI may be a proxy for sarcopenia-related weakness, whereas high BMI may indicate relative strength limitations and mobility inefficiency; these mechanisms align with previously described relationships between BMI and falls [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. In our combinatorial search under cross-validation, BMI contributed to robust monotonic effects that were well captured by a single-threshold-based rule (\u0026le;\u0026thinsp;20.8), making it suitable for inclusion in a parsimonious screening score.\u003c/p\u003e \u003cp\u003eAlthough the BNP level and PT-INR were selected in the six-item subset, these variables should not be overinterpreted as primary fall-risk determinants. Their inclusion reflects the combinatorial optimisation process; under cross-validation, they provided small but consistent incremental discrimination when paired with high-importance predictors such as the motor FIM score and BMI. The SHAP analysis revealed that the BNP level and PT-INR contributed to modest global effects but occasionally showed meaningful local contributions in individuals with metabolic or cardiovascular instability. These findings implied that the BNP level and PT-INR functioned as auxiliary markers that captured systemic vulnerability rather than as direct fall-risk drivers. Their modest integer weights (1\u0026ndash;2 points) were chosen to reflect this secondary role, ensuring that their operational impact remained limited while preserving their incremental predictive value.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section3\"\u003e \u003ch2\u003eIntervention patterns and opportunities for improvement\u003c/h2\u003e \u003cp\u003eOur analysis of care-related interventions revealed substantial variation in implementation rates across intervention types, with adherence to bed-rest restrictions being more common than direct assistance or supervision. This pattern likely reflects the resource constraints and practical considerations in busy clinical settings [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. The fall-score\u0026ndash;based intervention analysis indicated that the current allocation of preventive care does not always align with the predicted risk, suggesting opportunities to design score-based intervention protocols that prioritise high-risk patients while avoiding unnecessary restrictions on lower-risk individuals.\u003c/p\u003e \u003cp\u003eHowever, in this study, the fall scores were primarily used as a continuous measure to interrogate patterns of care rather than to define prescriptive risk categories. The three fall-score strata used in our descriptive analyses (\u0026lt;\u0026thinsp;0.33, 0.33 to \u0026lt;\u0026thinsp;0.66, and \u0026ge;\u0026thinsp;0.66) were chosen for interpretability and to visualise the mismatch between model-predicted risk and delivered preventive care; they are not proposed as ready-to-use clinical cut-offs. In practice, we envisage that the model-derived fall scores will complement existing FIM-based assessments by highlighting patients whose predicted risk is discordant with their subjectively assessed risk, whereas any operational threshold for score-based interventions should be tailored and prospectively evaluated within each institution.\u003c/p\u003e \u003cp\u003eThe six-item screening score enables efficient identification of patients at high risk; however, it does not fully explain why specific patients fall. The identification of three latent patient clusters supports the existence of heterogeneous fall-risk phenotypes, which a single cumulative score may obscure. Our findings suggest that an effective fall-prevention strategy requires a two-step approach combining quantitative risk stratification and qualitative profiling. First, the developed six-item screening score acts as a sensitive filter for identifying high-risk patients requiring immediate attention. However, a high score alone does not indicate the specific type of intervention required. This is where the value of the LCA lies. Once patients are flagged to have a high risk, classifying them into a specific clinical phenotype, as identified by our LCA, enables tailored interventions. For example, patients fitting the \u0026ldquo;functional impairment\u0026rdquo; cluster should be prioritised for intensive physical therapy and environmental adjustments (e.g., sensor mats), whereas those in the \u0026ldquo;metabolically vulnerable\u0026rdquo; cluster require a review of medications, hydration, and nutritional management rather than physical restrictions. This workflow transforms the prediction model from a simple warning system into a comprehensive decision support tool, potentially resolving the mismatch between patient needs and resource allocation.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e\n\u003ch3\u003eContributions of SHAP analysis\u003c/h3\u003e\n\u003cp\u003eThe SHAP analysis enhanced our understanding of how the RF model combines predictors to generate fall-risk estimates. By orthogonalising the predictor contributions, the SHAP analysis enabled the interpretation of complex interactions without relying solely on marginal associations. In this study, younger patients with poorer functional status may still have a high fall risk; this finding challenges the conventional focus on age alone and underscores the need to prioritise functional decline irrespective of chronological age. The positive association between certain laboratory parameters and fall risk highlights the importance of comprehensive medical management in fall-prevention strategies [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Importantly, the SHAP analysis informed the score-development process by clarifying which predictors contributed primarily through strong monotonic effects\u0026mdash;amenable to thresholding\u0026mdash;and which predictors contributed through complex interaction patterns that were less suitable for a simplified additive score. The motor FIM score, BMI, and renal markers showed clear and directionally consistent SHAP effects, supporting their inclusion in the score. In contrast, the cognitive FIM score exhibited substantial but highly interaction-dependent SHAP contributions, explaining its exclusion from the threshold-based score despite its importance in the full RF model. Thus, the SHAP analysis served not only as an interpretability tool but also as a guide for selecting variables compatible with a clinically implementable risk-stratification tool.\u003c/p\u003e\n\u003ch3\u003eLimitations\u003c/h3\u003e\n\u003cp\u003eThis study has some limitations. First, given that this was a single-centre retrospective study, selection bias and unmeasured confounding factors cannot be excluded, and the generalisability of our findings to other hospitals or healthcare systems is uncertain. External validation in independent cohorts and diverse settings is essential before this model can be adopted in routine practice.\u003c/p\u003e \u003cp\u003eSecond, although calibration was evaluated descriptively using calibration plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC and F), summary calibration indices, such as the calibration slope and Brier score were not calculated, limiting our ability to fully quantify the accuracy of predicted probabilities.\u003c/p\u003e \u003cp\u003eThird, missing values for laboratory predictors were handled using simple median imputation, and the normality of distribution was not formally assessed. This might have introduced bias, especially if the missingness was informative.\u003c/p\u003e \u003cp\u003eFourth, the categorical nature of care-related variables prevented us from quantifying the intensity, duration, and quality of interventions, and we did not account for fall severity or circumstances.\u003c/p\u003e \u003cp\u003eFifth, our analyses primarily considered baseline or most recent measurements, and we did not explicitly model temporal changes in risk during hospitalisation. Even if immediate examination findings predict falls at some point during admission, the lead time until a fall is not constant, complicating real-time resource allocation.\u003c/p\u003e \u003cp\u003eFinally, the model is internally validated. Implementation in practice would require not only external validation and potential recalibration but also careful consideration of the risk of automation bias and unintended consequences of over-reliance on algorithmic outputs [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. Although fall events were moderately imbalanced, we did not apply class weighting because the imbalance was considered mild and internal validation showed stable performance without rebalancing. Consequently, model predictions reflected the natural distribution of the fall outcome rather than weighted estimates.\u003c/p\u003e \u003cdiv id=\"Sec37\" class=\"Section2\"\u003e \u003ch2\u003eFuture directions\u003c/h2\u003e \u003cp\u003eFuture research should focus on prospective validation of the prediction model, integration of additional risk factors (e.g., detailed medication profiles and environmental assessments), and development of dynamic models that update risk predictions as patient status evolves. Incorporating real-time monitoring data and automated alert systems can further enhance the clinical utility of fall-prediction models. Investigations into cluster-specific intervention strategies and cost-effectiveness analyses comparing score-based and usual care approaches would provide valuable evidence for healthcare decision-makers.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusions","content":"\u003cp\u003eWe developed an ML-based fall-risk prediction model and a simplified six-item bedside screening score using routinely collected hospital data. The model provides objective and quantifiable risk estimates that can support standardised fall-prevention interventions and efficient human resource allocation. Although the model shows excellent discriminative performance, challenges in achieving perfect prediction accuracy, ensuring good calibration, and demonstrating external validity remain. Combining ML approaches with clinical expertise and careful evaluation is essential for improving patient safety and optimising resource utilisation in in-hospital fall-prevention programmes.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAUC\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Area under the curve\u003c/p\u003e\n\u003cp\u003eBMI\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Body mass index\u003c/p\u003e\n\u003cp\u003eBNP\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;B-type natriuretic peptide\u003c/p\u003e\n\u003cp\u003eCI\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Confidence interval\u003c/p\u003e\n\u003cp\u003eCRE\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Creatinine\u003c/p\u003e\n\u003cp\u003eCRP\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;C-reactive protein\u003c/p\u003e\n\u003cp\u003eeGFR\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Estimated glomerular filtration rate\u003c/p\u003e\n\u003cp\u003eFBG\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Fasting blood glucose\u003c/p\u003e\n\u003cp\u003eFIM\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Functional Independence Measure\u003c/p\u003e\n\u003cp\u003eHgb\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Haemoglobin\u003c/p\u003e\n\u003cp\u003eLCA\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Latent class analysis\u003c/p\u003e\n\u003cp\u003eLOCF\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Last observation carried forward\u003c/p\u003e\n\u003cp\u003eML\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Machine learning\u003c/p\u003e\n\u003cp\u003eNPV\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Negative predictive value\u003c/p\u003e\n\u003cp\u003ePT-INR\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Prothrombin time–international normalised ratio\u003c/p\u003e\n\u003cp\u003eRF\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Random forest\u003c/p\u003e\n\u003cp\u003eROC\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Receiver operating characteristic\u003c/p\u003e\n\u003cp\u003eSD\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Standard deviation\u003c/p\u003e\n\u003cp\u003eSHAP\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;SHapley Additive exPlanations\u003c/p\u003e\n\u003cp\u003eγ-GTP \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Gamma-glutamyl transpeptidase\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective study was approved by the Institutional Review Board of Saiseikai Moriyama Municipal Hospital (approval no. 30). The requirement for individual informed consent was waived because de-identified routinely collected data were used and the study posed minimal risk to\u0026nbsp;the participants.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and/or analysed during the current study, as well as the analysis code used for data preprocessing and model development, are not publicly available because of institutional restrictions but are available from the corresponding author\u0026nbsp;upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors’ contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTO and TI conceived\u0026nbsp;the study. TO curated the data and performed the statistical analyses. TI contributed to\u0026nbsp;the study design, data interpretation, and clinical contextualisation. TO drafted the manuscript, and TI critically revised it for important intellectual content. Both authors have read and approved the final\u0026nbsp;version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors thank the clinical and administrative staff of Moriyama Municipal Hospital for their assistance with data extraction and continuous efforts in\u0026nbsp;fall-prevention\u0026nbsp;activities.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors’ information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eKim DW, Seo J, Kwon S, Park CM, Han C, Kim Y, et al. Predicting In-hospital Fall risk Using Machine Learning with Real-Time Location System and Electronic Medical Records. J Cachexia Sarcopenia Muscle. 2025;16:e13713. doi:10.1002/jcsm.13713.\u003c/li\u003e\n \u003cli\u003eHill KD, Vu M, Walsh W. Falls in the acute hospital setting \u0026mdash; Impact on resource utilisation. Aust Health Rev. 2007;31:471-7. doi.org/10.1071/AH070471.\u003c/li\u003e\n \u003cli\u003eOliver D, Healey F, Haines TP. Preventing falls and fall-related injuries in hospitals. Clin Geriatr Med. 2010;26:645-92. doi:10.1016/j.cger.2010.06.005.\u003c/li\u003e\n \u003cli\u003eMontero-Odasso MM, Kamkar N, Pieruccini-Faria F, Osman A, Sarquis-Adamson Y, Close J, et al. Evaluation of clinical practice guidelines on fall prevention and management for older adults: A systematic review. JAMA Netw Open. 2021;4:e2138911. doi:10.1001/jamanetworkopen.2021.38911.\u003c/li\u003e\n \u003cli\u003eTricco AC, Thomas SM, Veroniki AA, Hamid JS, Cogo E, Strifler L, et al. Comparisons of interventions for preventing falls in older adults: A systematic review and meta-analysis. JAMA. 2017;318:1687-99. doi:10.1001/jama.2017.15006.\u003c/li\u003e\n \u003cli\u003eGuirguis-Blake JM, Michael YL, Perdue LA, Coppola EL, Beil TL. Interventions to prevent falls in older adults: Updated evidence report and systematic review for the US Preventive Services Task Force. JAMA\u003cem\u003e.\u003c/em\u003e 2018;319:1705\u0026ndash;16. doi:10.1001/jama.2017.21962.\u003c/li\u003e\n \u003cli\u003eSpoelstra SL, Given BA, Given CW. Fall prevention in hospitals: An integrative review. Clin Nurs Res. 2011;21:92-112. doi:10.1177/1054773811418106.\u003c/li\u003e\n \u003cli\u003eRandell R, McVey L, Wright J, Zaman H, Cheong VL, Woodcock DM, et al. Practices of falls risk assessment and prevention in acute hospital settings: A realist investigation. Health Soc Care Deliv Res. 2024;12:1-194.\u003c/li\u003e\n \u003cli\u003eOliver D, Daly F, Martin FC, McMurdo MET. Risk factors and risk assessment tools for falls in hospital in-patients: A systematic review. Age Ageing. March 2004;33:122\u0026ndash;30, https://doi.org/10.1093/ageing/afh017.\u003c/li\u003e\n \u003cli\u003eStrini V, Schiavolin R, Prendin A. Fall risk Assessment Scales: A Systematic Literature Review. Nurs Rep. 2021;11:430-43. https://doi.org/10.3390/nursrep11020041.\u003c/li\u003e\n \u003cli\u003eRubenstein LZ. Falls in older people: Epidemiology, risk factors and strategies for prevention. Age Ageing. September 2006;35 Suppl 2, Issue suppl_2:ii37\u0026ndash;41, https://doi.org/10.1093/ageing/afl084.\u003c/li\u003e\n \u003cli\u003eCallis N. Falls prevention: Identification of predictive fall risk factors. Appl Nurs Res. 2016;29:53-8. doi:10.1016/j.apnr.2015.05.007.\u003c/li\u003e\n \u003cli\u003eGreene BR, Redmond SJ, Caulfield B. Fall risk Assessment through Automatic Combination of Clinical Fall risk Factors and Body-Worn Sensor Data. IEEE J Biomed Health Inform. 2017;21:725-31. doi:10.1109/JBHI.2016.2539098.\u003c/li\u003e\n \u003cli\u003eSchultebraucks K, Galatzer-Levy IR. Machine learning for prediction of posttraumatic stress and resilience following trauma: An overview of basic concepts and recent advances. J Trauma Stress. 2019;32:215-25. doi:10.1002/jts.22384.\u003c/li\u003e\n \u003cli\u003eH Mohammed S, Ahamad S, Abdul Hameed DM. Enhancements in random forest algorithms for improving healthcare applications. In: \u003cem\u003e4th Indian International Conference on Industrial Engineering and Operations Management\u003c/em\u003e; 2024, November, https://doi.org/10.46254/IN04.20240192.\u003c/li\u003e\n \u003cli\u003eMishra AK, Skubic M, Despins LA, Popescu M, Keller J, Rantz M, et al. Explainable Fall risk Prediction in Older Adults Using Gait and Geriatric Assessments. Front Digit Health. 2022;4:869812. doi:10.3389/fdgth.2022.869812.\u003c/li\u003e\n \u003cli\u003eTang YT, Romero-Ortuno R. Using explainable AI (XAI) for the prediction of falls in the older population. Algorithms. 2022;15:353. https://doi.org/10.3390/a15100353.\u003c/li\u003e\n \u003cli\u003eYip MYT, Lim G, Lim ZW, Nguyen QD, Chong CCY, Yu M, et al. Technical and imaging factors influencing performance of deep learning systems for diabetic retinopathy. NPJ Digit Med. 2020;3:40.\u003c/li\u003e\n \u003cli\u003eHe J, Baxter SL, Xu J, Xu J, Zhou X, Zhang K. The practical implementation of artificial intelligence technologies in medicine. Nat Med. 2019;25:30\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eCollins GS, Dhiman P, Andaur Navarro CL, et al. TRIPOD-AI: An updated reporting guideline for prediction model studies involving artificial intelligence. BMJ. 2024;385:e079044.\u003c/li\u003e\n \u003cli\u003eCollins GS, Reitsma JB, Altman DG, Moons KGM. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): The TRIPOD statement. Ann Intern Med. 2015;162:55\u0026ndash;63.\u003c/li\u003e\n \u003cli\u003eRiley RD, Ensor J, Snell KIE, Harrell FE Jr, Martin GP, Reitsma JB, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441. BMJ. 2019;368:m441.\u003c/li\u003e\n \u003cli\u003eWong ES, Yoon J, Piegari RI, Rosland AMM, Fihn SD, Chang ET. Identifying latent subgroups of high-risk patients using risk score trajectories. J Gen Intern Med. 2018\u003cem\u003e;\u003c/em\u003e33:2120\u0026ndash;6. https://doi.org/10.1007/s11606-018-4653-x.\u003c/li\u003e\n \u003cli\u003eWang Y, Zhao Y, Therneau TM, Atkinson EJ, Tafti AP, Zhang N, et al. Unsupervised machine learning for the discovery of latent disease clusters and patient subgroups using electronic health records. J Biomed Inform. 2020;102:103364. doi:10.1016/j.jbi.2019.103364.\u003c/li\u003e\n \u003cli\u003eMarschollek M, G\u0026ouml;vercin M, Rust S, Gietzelt M, Schulze M, Wolf KH, \u003cem\u003eet al.\u003c/em\u003e Mining geriatric assessment data for in-patient fall prediction models and high-risk subgroups. BMC Med Inform Decis Mak. 2012;12:19. https://doi.org/10.1186/1472-6947-12-19.\u003c/li\u003e\n \u003cli\u003eYouden WJ. Index for rating diagnostic tests. Cancer. 1950 Jan;3:32-5. doi: 10.1002/1097-0142(1950)3:1\u0026lt;32::aid-cncr2820030106\u0026gt;3.0.co;2-3. PMID: 15405679.\u003c/li\u003e\n \u003cli\u003eDaneshvar A, Mousa G. Regression shrinkage and selection via least quantile shrinkage and selection operator. PLOS One. 2023;18:e0266267. https://doi.org/10.1371/journal.pone.0266267.\u003c/li\u003e\n \u003cli\u003eTap L, Boy\u0026eacute; NDA, Hartholt KA, van der Cammen TJM, Mattace-Raso FUS. Association of estimated glomerular filtration rate with muscle function in older persons who have fallen. Age Ageing. March 2018;47:269\u0026ndash;74, https://doi.org/10.1093/ageing/afx180.\u003c/li\u003e\n \u003cli\u003eKao PY, Yeh HC, Hsia YF, Hsiao YL, Wang JS, Chang DR, et al. Paradoxical mortality of high estimated glomerular filtration rate reversed by 24-h urine creatinine excretion rate adjustment: Sarcopenia matters. J Cachexia Sarcopenia Muscle. 2022;13:1704-16. doi:10.1002/jcsm.12951.\u003c/li\u003e\n \u003cli\u003eJuraschek SP, Daya N, Appel LJ, Miller ER, Matsushita K, Michos ED, et al. Subclinical cardiovascular Disease and Fall risk in Older Adults: Results from the Atherosclerosis Risk in Communities Study. J Am Geriatr Soc. 2019;67:1795-802. doi:10.1111/jgs.16041.\u003c/li\u003e\n \u003cli\u003eInoue S, Otaka Y, Mori N, Matsuura D, Tsujikawa M, Kawakami M, et al. Blind spots in hospital fall prevention: Falls in stroke patients occurred not only in those at a high risk of falling. J Am Med Dir Assoc. 2024;25:160\u0026ndash;166.e1. doi:10.1016/j.jamda.2023.10.034.\u003c/li\u003e\n \u003cli\u003eCordani C, Battel I. Do implementation interventions improve evidence-based care in acute stroke settings? A Cochrane Review summary with commentary. NeuroRehabilitation. 2024;54:343-6. doi:10.3233/NRE-246002.\u003c/li\u003e\n \u003cli\u003eHoyer EH, Young DL, Zhang C, Colantuoni E, Ghobadi K. Falls in hospitals: Challenging traditional risk assessments with new insights into patient mobility. J Adv Nurs. Published online February 28, 2025. 2025;81:5824\u0026ndash;30. doi:10.1111/jan.16866.\u003c/li\u003e\n \u003cli\u003eGoddard K, Roudsari A, Wyatt JC. Automation bias: A systematic review of frequency, effect mediators, and mitigators. J Am Med Inform Assoc. 2012;19:121\u0026ndash;7.\u003c/li\u003e\n \u003cli\u003eCabitza F, Rasoini R, Gensini GF. Unintended consequences of machine learning in medicine. JAMA. 2017;318:517\u0026ndash;8.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Fall risk, Hospital safety, Machine learning, Random forest, SHAP analysis, Functional Independence Measure, Resource allocation, Screening score","lastPublishedDoi":"10.21203/rs.3.rs-8540709/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8540709/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eConventional 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.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis 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 \u0026ldquo;last observation carried forward\u0026rdquo; 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.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eOut 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\u0026ndash;54.0) days. The RF model showed excellent discriminative performance (area under the curve [AUC]: 0.96, 95% confidence interval [CI]: 0.931\u0026ndash;0.981, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;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., \u0026ldquo;functional impairment\u0026rdquo; vs. \u0026ldquo;metabolically vulnerable\u0026rdquo;) with distinct risk profiles, highlighting the heterogeneity of high-risk patients. The derived six-item screening score (motor FIM score\u0026thinsp;\u0026le;\u0026thinsp;51, B-type natriuretic peptide level\u0026thinsp;\u0026le;\u0026thinsp;66.1, prothrombin time\u0026ndash;international normalised ratio\u0026thinsp;\u0026le;\u0026thinsp;1.01, estimated glomerular filtration rate\u0026thinsp;\u0026ge;\u0026thinsp;78.4, haemoglobin level\u0026thinsp;\u0026ge;\u0026thinsp;11.2, BMI\u0026thinsp;\u0026le;\u0026thinsp;20.8) yielded an AUC of 0.668 (95% CI: 0.618\u0026ndash;0.716, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) with equal weights and 0.695 (95% CI: 0.644\u0026ndash;0.742, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) with integer weights. High-sensitivity thresholds of \u0026ge;\u0026thinsp;2 points for the equal-weight score and \u0026ge;\u0026thinsp;4 points for the integer-weight score achieved sensitivity/specificity of 0.908/0.279 and 0.825/0.458, respectively.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThe 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.\u003c/p\u003e\u003ch2\u003eTrial registration:\u003c/h2\u003e \u003cp\u003eNot applicable to this retrospective observational study, which was not prospectively registered.\u003c/p\u003e","manuscriptTitle":"Development and internal validation of a machine learning–based prediction model and simplified screening score for in-hospital falls: a retrospective cohort study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-23 00:38:18","doi":"10.21203/rs.3.rs-8540709/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-04-15T09:17:57+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-17T03:36:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"200787290969803865805014884405070457750","date":"2026-03-16T07:42:12+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"49731747499688833702333625575255829650","date":"2026-03-16T01:39:17+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"295524991084838073681162350824407347171","date":"2026-03-13T17:58:11+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-02T14:26:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"236883612103807918241882445444071486799","date":"2026-02-08T11:41:30+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-20T17:37:35+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-01-12T15:13:33+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-09T14:19:52+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-09T14:17:32+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2026-01-07T11:01:30+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"45a9fa60-ae13-4403-937f-11705be3cd66","owner":[],"postedDate":"January 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-01-23T00:38:23+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-23 00:38:18","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8540709","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8540709","identity":"rs-8540709","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}