Admission clinical and functional assessments for predicting inpatient fall risk using machine learning models

Abstract

Objectives

Falls are a major patient safety concern in inpatient care. This study aimed to develop machine learning models using multidimensional clinical factors to improve early identification of high-risk patients.

Methods

A retrospective dataset of 483,174 adult inpatients was analyzed. Logistic Regression, Support Vector Machine, Random Forest, Extreme Gradient Boosting, and Multi-Layer Perceptron were applied. Models were trained with and without SMOTENC and evaluated using five-fold cross-validation, with emphasis on AUC-ROC, recall, and F1-score.

Results

Key predictors included fall history, medication use, dizziness, sensory impairment, and reduced functional status. Without SMOTENC, models achieved moderate AUC-ROC values ranging from 0.62 to 0.72, but exhibited extremely low recall and F1-scores for fall events. After applying SMOTENC, AUC-ROC remained stable in the range of 0.67 to 0.70, while recall improved substantially to approximately 50 to 60 percent. Random Forest and XGBoost demonstrated relatively stable AUC-ROC performance, whereas MLP showed limited generalization ability.

Conclusions

Machine learning models can capture clinically relevant fall risk factors, with AUC-ROC indicating moderate discrimination. Severe class imbalance substantially limits minority event detection. Although SMOTENC improves recall, it introduces a precision–recall trade-off, underscoring the need to interpret AUC-ROC alongside class-specific metrics in imbalanced clinical datasets.

Keywords

fall risk prediction machine learning models multi-layer perceptron polypharmacy activities of daily living

Introduction

Falls represent a potential risk for all hospitalized patients, regardless of age, sex, or diagnosis. They not only lead to physical injuries and mobility restrictions but may also trigger psychological consequences such as fear and loss of confidence, thereby negatively affecting overall care and recovery. The occurrence of falls is often sudden and unpredictable, and even younger or middle-aged patients may be affected. These findings highlight that falls are not confined to specific patient groups but are broadly present across the continuum of hospital care. Current research suggests that multifactorial interventions may effectively reduce fall incidence,¹ though such measures may restrict patient mobility. Thus, identifying optimal assessment tools or intervention strategies requires interdisciplinary collaboration.² Some studies have reported that surgical wards have the lowest fall incidence,³ which warrants further investigation. Moreover, fall events substantially increase healthcare costs compared to patients without falls.⁴ In clinical practice, nurses are primarily responsible for fall risk identification and prevention. Previous studies have shown that the accuracy and consistency of nursing assessments may be influenced by factors such as staffing patterns, patient load, years of experience, and workplace support.⁵ If assessment outcomes are not promptly translated into effective care actions, opportunities for fall prevention may be missed. Beyond the limitations of human judgment, fall occurrence is also closely associated with specific risk factors, including the use of central nervous system medications, sedatives, and hypnotics.⁶

In recent years, numerous studies have focused on developing predictive models for fall prevention. Some scholars have collected data specifically from medical inpatients and demonstrated that incorporating fall-prevention variables into the models significantly improved predictive accuracy.⁷ Others analyzed data from 15 hospital wards and evaluated variables at four distinct time points during hospitalization, finding that multi-temporal assessments enhanced predictive performance.⁸ Another study trained a model using 46 features and reported an accuracy as high as 95.11%.⁹ Research on patients undergoing spinal and lower extremity surgery further revealed that with each additional year of age, the risk of falls increased, underscoring age as a critical risk factor.¹⁰ In addition, face-to-face interviews combined with anthropometric and functional assessments were used to predict falls within a 6-month follow-up period.¹¹ Other investigators have utilized inpatient fall records with feature selection techniques to identify optimal subsets of predictors and subsequently developed three displacement-based models, demonstrating that machine learning approaches achieved favorable performance when applied to the Fall Risk Assessment Tool (FRAT).¹² More recently, the use of real-time location systems to capture patients’ dynamic movement patterns has been proposed to enhance predictive accuracy, with findings indicating that higher fall probability was significantly associated with sedative use and increased red blood cell distribution width.¹³

Despite these advancements, several important gaps remain. Many existing studies emphasize overall accuracy or AUC without adequately reporting class-specific performance, such as recall or F1-score for fall events, which are critical in highly imbalanced clinical datasets. The issue of class imbalance is often insufficiently addressed or not explicitly discussed, potentially leading to overestimation of model performance and limited real-world applicability. Few studies systematically evaluate the impact of different imbalance handling strategies, particularly in large-scale inpatient datasets with mixed-type clinical variables. To address these limitations, the present study develops and evaluates multiple machine learning models using a large-scale inpatient dataset and explicitly investigates the effect of class imbalance on model performance. Unlike prior work, this study not only compares models using both discrimination metrics and class-specific measures, but also examines the effectiveness of resampling techniques, including SMOTE and SMOTENC, with particular attention to the latter’s ability to handle mixed categorical and numerical features. By doing so, this study aims to provide a more realistic and clinically meaningful evaluation of fall prediction models and to clarify the methodological challenges associated with imbalanced clinical data. This study aims to develop and systematically evaluate machine learning models for predicting fall risk among hospitalized patients by integrating multidimensional clinical factors, including demographic characteristics, functional status, medication use, sensory function, and activities of daily living (ADL). Given the highly imbalanced nature of fall-related data and the complexity of clinical tabular features, appropriate data preprocessing strategies are incorporated to enhance model learning and improve the detection of rare fall events. Multiple machine learning algorithms with distinct modeling mechanisms are compared to assess their predictive performance, generalization ability, and clinical applicability. Falls represent clinically significant adverse events and key indicators of patient safety in inpatient care, often leading to physical injury, functional decline, psychological distress, and increased healthcare utilization. Therefore, this study seeks to establish a robust and quantifiable prediction framework to support early risk stratification and facilitate targeted fall prevention strategies, ultimately contributing to improved patient safety and healthcare quality.

Materials and methods

Ethical approval

The study was approved by the Institutional Review Board of Changhua Christian Hospital (approval no.: 241105). The Institutional Review Board waived the need for informed consent considering the retrospective nature of data collected. All methods were performed in accordance with the relevant guidelines and regulations or declaration of Helsinki.

Study setting, design, and ethical considerations

This study was a retrospective analysis conducted at Changhua Christian Hospital in central Taiwan. The study population consisted of adult inpatients aged 18 to 111 years. Nursing and clinical information documented at the time of admission were collected and analyzed. To address the substantial imbalance between fall and non-fall cases, fall-related data were collected over a 10-year period from 2014 to 2024. A total of 483,174 records were included in the analysis, comprising 2,557 fall cases and 480,617 non-fall cases. The dataset was randomly divided into a training dataset (N = 386,539; falls = 2,046; non-falls = 384,493) and an independent testing dataset (N = 96,635; falls = 511; non-falls = 96,124). To ensure robust model development and evaluation, five-fold cross-validation was applied within the training dataset. In each fold, approximately 80% of the data were used for model training (N ≈ 309,231–309,232), and the remaining 20% were used for validation (N ≈ 77,307–77,308), while maintaining class distribution consistency across folds, as shown in Figure 1. The study protocol was reviewed and approved by the Institutional Review Board of Changhua Christian Hospital. All data were de-identified prior to analysis to ensure patient confidentiality. Due to the retrospective design and the use of secondary data, the requirement for informed consent was waived.

Figure 1.

Dataset distribution and five-fold cross-validation scheme for fall prediction modeling.

Model training

Based on effect size measures derived from univariate analysis (Cramer’s V and Cohen’s d) and clinical relevance, a set of variables was selected as input features. Based on effect size analysis and clinical relevance, a set of key variables was selected as input features for model development. These included age, Barthel Index total score, unstable gait, need for walking aids, history of falls within one year, dizziness or vertigo, abnormal foot sensation, use of fall-risk medications, number of medications, activity level, mobility, stair climbing, walking on level ground, transfer ability, and toileting.

This study compared multiple machine learning models, including Logistic Regression, Random Forest, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). These models were selected due to their distinct learning mechanisms and representativeness across different categories of machine learning approaches. Logistic Regression is a linear model that provides high interpretability and serves as a baseline for comparison.¹⁴ Random Forest is an ensemble learning method based on bagging, which improves robustness and reduces variance through multiple decision trees.¹⁵ XGBoost is a gradient boosting framework, focuses on sequential learning and error correction, often achieving superior performance in structured data tasks.¹⁶ SVM is a margin-based classifier that is effective in high-dimensional spaces and can model non-linear relationships through kernel functions.¹⁷ MLP is a neural network-based model capable of capturing complex non-linear patterns through multiple hidden layers.¹⁸

The Synthetic Minority Over-sampling Technique for Nominal and Continuous features (SMOTENC) was applied prior to model training to address class imbalance. Unlike standard SMOTE, SMOTENC is specifically designed for datasets containing both numerical and categorical variables, allowing synthetic samples to be generated while preserving the integrity of categorical feature distributions. SMOTENC generates synthetic minority class samples within the feature space to reduce bias caused by imbalanced class distributions, while minimizing distortion of categorical attributes. This approach improves the model’s ability to learn decision boundaries for minority classes without introducing unrealistic data patterns. In addition, it enhances model robustness, reduces overfitting to majority class patterns, and improves sensitivity in detecting rare events such as falls. To systematically evaluate the impact of SMOTENC, two sets of models were constructed: one without SMOTENC and one with SMOTENC, and their performances were compared. All models were trained and validated using 5-fold cross-validation to ensure consistency in class distribution across folds. This design enabled a comprehensive assessment of performance differences under different sampling strategies and evaluated whether SMOTENC improved the models’ ability to identify minority fall cases.

Experimental environment

The experiments in this study were conducted on a workstation equipped with the Windows 10 64-bit operating system, an Intel Core i9 processor, 128 GB of RAM, and an NVIDIA RTX 4090 GPU. All software implementations were developed in Python 3.10. Machine learning models were implemented using the scikit-learn library version 1.6.0 and the XGBoost package version 2.0.3. Data processing and analysis were performed using NumPy version 1.26.0 and pandas version 2.3.3.

The hyperparameters of all machine learning models were predefined to ensure stable training and fair comparison. Class imbalance was addressed using class weighting where applicable. Model complexity and overfitting were controlled through regularization and early stopping strategies, and all models were implemented with fixed random seeds to ensure reproducibility as shown in Table 1.

Table 1.

Detailed hyperparameter settings for all machine learning models.

Parameter name	Value
Logistic Regression	max_iter=2000; solver=liblinear; class_weight=balanced
Random Forest	n_estimators=300; max_depth=10; min_samples_split=10; min_samples_leaf=5; max_features=sqrt; class_weight=balanced
XGBoost	n_estimators=200; max_depth=3; learning_rate=0.03; subsample=0.8; colsample_bytree=0.8; min_child_weight=10; gamma=2; reg_alpha=0.5; reg_lambda=2
SVM	C=0.5; max_iter=2000; class_weight=balanced
MLP	hidden_layers=(128,64,32); activation=ReLU; batch_size=256; learning_rate=0.0005; early_stopping=True

Model evolution

Confusion matrix

The performance of the models was evaluated using multiple classification metrics. Accuracy was employed to measure the overall proportion of correct predictions, as shown in Eqs. (1). Recall reflects the model’s ability to correctly identify positive cases among all actual positive samples, as presented in Eqs. (2). Precision emphasizes the proportion of true positives among all samples predicted as positive, as illustrated in Eqs. (3). The F1-score, defined as the harmonic mean of Precision and Recall, provides a balanced measure that accounts for both completeness and correctness of positive predictions, making it particularly suitable for imbalanced datasets, as shown in Eqs. (4).

A c c u r a c y = \frac{T P + T N}{T P + T N + F P + F N}

(1)

R e c a l l = \frac{T P}{T P + F N}

(2)

P r e c i s i o n = \frac{T P}{T P + F P}

(3)

F 1 - s c o r e = \frac{2 * P r e c i s i o n * R e c a l l}{P r e c i s i o n + R e c a l l}

(4)

Statistical analysis

This study first performed descriptive statistical analyses to examine the distributional characteristics of each variable and to explore their potential associations with fall events, serving as the basis for subsequent variable selection. Categorical variables were presented as frequencies and percentages. The Chi-square test of independence was applied to assess whether the distributional differences between the fall and non-fall groups were statistically significant. In addition to p-values, effect sizes were quantified using Cramér’s V to evaluate the strength of associations. For continuous variables, given the potential deviation from normality, the data were summarized using the median, first quartile (Q1), and third quartile (Q3). The Mann–Whitney U test, a nonparametric method, was then employed to compare differences between the two groups. Effect sizes for continuous variables were further assessed using Cohen’s d to quantify the magnitude of differences between groups. To quantify the direction and strength of associations between predictors and fall events, odds ratios (ORs) with 95% confidence intervals were estimated, providing an interpretable measure of risk for both categorical and continuous variables.

Results

Table 2 summarizes the distribution of key categorical variables between the non-fall and fall groups, while detailed results are provided in Supplementary Table S1. Given the large sample size, interpretation focused on odds ratios and Cramer’s V rather than p-values alone. The fall group showed a higher prevalence of male sex and smoking in the demographic domain. Clinically, a history of falls within the past year demonstrated the strongest association with fall events, followed by lack of caregiver support and the use of fall-risk medications. Medication-related factors, particularly the use of sedatives, psychiatric medications, and antiepileptics, were more frequently observed in the fall group, indicating a substantial pharmacological contribution to fall risk. In terms of neurological and sensory factors, dizziness or vertigo and abnormal foot sensation were notably more common among patients who experienced falls. Functional status showed the most pronounced differences between groups. Patients in the fall group exhibited significantly higher levels of impairment across multiple domains, including mobility, activity level, transfer ability, and walking ability. These findings suggest that functional dependence and mobility-related impairments are more strongly associated with inpatient falls than demographic characteristics alone.

Table 2.

Baseline characteristics associated with fall risk.

	Feature		Non-fall (n=480,617)	Fall (n=2,557)	Odds ratio	Cramer’s V	p-value
Demographics	Gender	Famale	47.5%	38%	0.678	0.014	<0.001
	Gender	Man	52.5%	62%	1.475	0.014	<0.001
	Smoke		14.6%	20.5%	1.510	0.012	<0.001
	Defecation		89.7%	84.2%	0.614	0.013	<0.001
	History of falls within one year		9.2%	24.1%	3.155	0.037	<0.001
	Inpatient without caregiver		6.1%	9.4%	1.594	0.010	<0.001
	Use of fall-risk medications		7%	12.2%	1.848	0.015	<0.001
	Sedatives		0.4%	1.3%	3.332	0.010	<0.001
	Psychiatric medications		0.8%	2.7%	3.528	0.016	<0.001
	Antiepileptics		0.1%	0.4%	3.639	0.005	<0.001
Neurological/Sensory	Dizziness or vertigo		8.7%	19.4%	2.524	0.027	<0.001
	Abnormal foot sensation		12.7%	27.2%	2.567	0.031	<0.001
	Activity	1	11.4%	11.5%	1.005	0.033	<0.001
		2	11.3%	16.9%	1.597
		3	18.9%	33.5%	2.162
		4	58.4%	38.1%	0.439
	Mobility	1	4%	2.9%	0.711	0.032	<0.001
		2	9.8%	14.4%	1.554
		3	23.6%	39.5%	2.119
		4	62.6%	43.2%	0.453
	Level-ground walking	0	14.4%	16.9%	1.204	0.038	<0.001
		5	9.2%	18.3%	2.208
		10	19.7%	32.2%	1.934
		15	56.6%	32.6%	0.37
	Transfer ability	0	11.2%	11.4%	1.021	0.038	<0.001
		5	11.2%	21.7%	2.192
		10	21.4%	34%	1.891
		15	56.2%	33%	0.383

Table 3 presents the comparison of continuous variables between the non-fall and fall groups. Due to the large sample size, emphasis was placed on effect sizes (Cohen’s d) rather than statistical significance alone. The fall group was older on average, although the magnitude of this difference was small. Similarly, lower body weight and BMI were observed in the fall group, but these differences also showed small effect sizes, indicating limited practical significance. No meaningful difference was observed in height. The Barthel Index showed a moderate effect size, with lower scores in the fall group, indicating greater dependence in activities of daily living. This suggests that functional status is a more clinically relevant factor in fall risk than age or body composition alone.

Table 3.

Comparison of continuous variables between fall and non-fall groups.

	Non fall (n=480617)	Fall (n=2557)	Odds ratio	Cohen’s d	p-value
Age	59.57(18-111)	64.08(18-99)	1.015	-0.254	<0.001
Height	161.34(121-206)	161.17(125-201)	0.998	0.019	0.832
Weight	63.3(30.05-198)	61.34(30.1-159)	0.990	0.139	<0.001
BMI	24.24(10.09-80.86)	23.55(12.12-70.72)	0.966	0.150	<0.001
Barthel Index total score	77.77(0-100)	68.74(0-100)	0.992	0.286	<0.001

The performance of all models without applying SMOTENC is presented in Tables 4–6. Classification results, including accuracy, area under the receiver operating characteristic curve, specificity, and sensitivity, are reported as mean ± standard deviation across five-fold cross-validation for the training, validation, and test sets, reflecting model behavior under the original imbalanced data distribution. Further examination of class-specific performance revealed a pronounced degradation in the ability to detect fall events, with most models exhibiting markedly low recall and F1-scores for the minority class, while maintaining consistently high performance for non-fall cases. This disparity indicates a strong bias toward the majority class and is further supported by the confusion matrix results, where predictions were predominantly assigned to the non-fall category. SVM, XGBoost, and MLP demonstrated near-zero true positive predictions across all data splits, indicating a severe prediction collapse toward the majority class. These results provide clear evidence that, in the absence of SMOTENC, most models fail to learn meaningful decision boundaries for the minority class, resulting in poor detection of fall events despite relatively high overall accuracy. This finding underscores the critical impact of class imbalance and highlights the necessity of appropriate strategies to address it in clinical prediction tasks.

Table 4.

Mean ± standard deviation of overall model performance metrics without SMOTENC across five-fold cross-validation for training, validation, and test sets.

Model	Mode	Accuracy	AUC_ROC	Specificity	Sensitivity
Linear Regression	Train	67.57% ± 0.35%	63.09% ± 0.26%	67.62% ± 0.35%	80.57% ± 0.25%
	Validation	67.53% ± 0.36%	62.82% ± 0.88%	67.58% ± 0.37%	80.54% ± 0.26%
	Test	67.58% ± 0.33%	64.63% ± 0.17%	67.61% ± 0.33%	80.58% ± 0.24%
Random Forest	Train	75.10% ± 1.05%	72.91% ± 0.36%	75.12% ± 1.06%	85.71% ± 0.69%
	Validation	74.94% ± 1.15%	62.44% ± 1.78%	75.07% ± 1.18%	85.63% ± 0.76%
	Test	74.93% ± 1.07%	62.13% ± 0.35%	75.07% ± 1.08%	85.63% ± 0.70%
XGBoost	Train	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
	Validation	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
	Test	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
SVM	Train	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
	Validation	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
	Test	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
MLP	Train	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
	Validation	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%
	Test	99.47% ± 0.00%	50.00% ± 0.00%	100.00% ± 0.00%	99.73% ± 0.00%

Table 5.

Mean ± standard deviation of class-specific performance metrics without SMOTENC for fall and non-fall predictions across five-fold cross-validation.

Model	Mode	Precision_Fall	Recall_Fall	F1_Fall	Precision_NonFall	Recall_NonFall	F1_NonFall
Linear Regression	Train	69.21% ± 0.32%	1.24% ± 0.03%	0.95% ± 0.01%	58.57% ± 0.42%	1.88% ± 0.02%	99.67% ± 0.00%
	Validation	68.89% ± 1.24%	1.27% ± 0.11%	0.94% ± 0.03%	58.06% ± 2.00%	1.86% ± 0.05%	99.67% ± 0.01%
	Test	69.75% ± 0.07%	1.20% ± 0.01%	1.00% ± 0.01%	61.64% ± 0.48%	1.97% ± 0.01%	99.70% ± 0.00%
Random Forest	Train	80.12% ± 0.29%	3.28% ± 0.15%	1.49% ± 0.06%	70.70% ± 0.42%	2.92% ± 0.11%	99.79% ± 0.00%
	Validation	69.18% ± 1.26%	1.46% ± 0.35%	1.05% ± 0.05%	49.80% ± 4.66%	2.06% ± 0.11%	99.65% ± 0.03%
	Test	69.68% ± 0.22%	1.24% ± 0.02%	1.04% ± 0.02%	49.20% ± 1.48%	2.03% ± 0.04%	99.64% ± 0.01%
XGBoost	Train	72.57% ± 0.30%	1.83% ± 0.07%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
	Validation	71.29% ± 1.02%	1.61% ± 0.26%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
	Test	72.32% ± 0.29%	1.45% ± 0.02%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
SVM	Train	69.20% ± 0.32%	1.23% ± 0.02%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
	Validation	68.87% ± 1.24%	1.26% ± 0.11%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
	Test	69.72% ± 0.07%	1.18% ± 0.01%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
MLP	Train	71.34% ± 0.77%	1.64% ± 0.12%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
	Validation	70.42% ± 1.42%	1.47% ± 0.25%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%
	Test	71.15% ± 0.46%	1.36% ± 0.07%	0.00% ± 0.00%	0.00% ± 0.00%	0.00% ± 0.00%	99.47% ± 0.00%

Table 6.

Mean ± standard deviation of confusion matrix components without SMOTENC across five-fold cross-validation for training, validation, and test sets.

Mode	Mode	TN	FP	FN	TP
Linear Regression	Train	207986.20 ± 1074.67	99608.20 ± 1075.15	678.20 ± 6.83	958.60 ± 7.09
	Validation	51965.20 ± 286.73	24933.40 ± 286.29	171.60 ± 8.08	237.60 ± 8.38
	Test	64989.40 ± 319.97	31134.60 ± 319.97	196.00 ± 2.45	315.00 ± 2.45
Random Forest	Train	231065.00 ± 3253.09	76529.40 ± 3253.47	479.60 ± 7.02	1157.20 ± 6.76
	Validation	57729.20 ± 908.52	19169.40 ± 908.18	205.40 ± 18.98	203.80 ± 19.19
	Test	72162.00 ± 1039.64	23962.00 ± 1039.64	259.60 ± 7.57	251.40 ± 7.57
XGBoost	Train	307594.40 ± 0.55	0.00 ± 0.00	1636.80 ± 0.45	0.00 ± 0.00
	Validation	76898.60 ± 0.55	0.00 ± 0.00	409.20 ± 0.45	0.00 ± 0.00
	Test	96124.00 ± 0.00	0.00 ± 0.00	511.00 ± 0.00	0.00 ± 0.00
SVM	Train	307594.40 ± 0.55	0.00 ± 0.00	1636.80 ± 0.45	0.00 ± 0.00
	Validation	76898.60 ± 0.55	0.00 ± 0.00	409.20 ± 0.45	0.00 ± 0.00
	Test	96124.00 ± 0.00	0.00 ± 0.00	511.00 ± 0.00	0.00 ± 0.00
MLP	Train	307594.40 ± 0.55	0.00 ± 0.00	1636.80 ± 0.45	0.00 ± 0.00
	Validation	76898.60 ± 0.55	0.00 ± 0.00	409.20 ± 0.45	0.00 ± 0.00
	Test	96124.00 ± 0.00	0.00 ± 0.00	511.00 ± 0.00	0.00 ± 0.00

The performance of all models with the application of SMOTENC is presented in Tables 7–9. Overall classification results, including accuracy, area under the receiver operating characteristic curve, specificity, and sensitivity, are reported as mean ± standard deviation across five-fold cross-validation for the training, validation, and test sets. Compared with the results without SMOTENC, the application of SMOTENC led to notable changes in model behavior, particularly in improving sensitivity for fall detection. Random Forest and XGBoost demonstrated relatively stable performance across all data splits, with moderate accuracy and balanced sensitivity and specificity. Logistic Regression and SVM showed comparable performance, with moderate improvements in sensitivity compared to the non-SMOTE setting. In contrast, although MLP achieved high training performance, its test and validation AUC values were substantially reduced, indicating limited generalization ability. Class-specific performance further revealed that SMOTENC substantially improved recall for fall cases across most models. Logistic Regression, Random Forest, XGBoost, and SVM achieved recall values ranging from approximately 50% to 60% on the test and validation sets, representing a marked improvement compared to near-zero recall observed without SMOTENC. This improvement in recall was accompanied by extremely low precision for fall predictions, generally below 1%, indicating a high number of false positive predictions. F1-scores for fall events remained low despite the increase in recall. The confusion matrix results provide additional insight into this trade-off. While the number of true positive predictions increased considerably across all models after applying SMOTENC, there was also a substantial increase in false positives, reflecting a shift in model behavior toward detecting more minority cases at the expense of specificity. This pattern was consistently observed across Logistic Regression, Random Forest, XGBoost, and SVM. For MLP, although training performance appeared strong, the model produced relatively low true positive counts in the test and validation sets, suggesting instability and limited robustness under resampled conditions. These results indicate that the application of SMOTENC effectively mitigates the prediction collapse observed in the non-SMOTE setting by improving the ability of models to detect fall events. This improvement comes at the cost of a substantial increase in false positive predictions, resulting in low precision and limited overall effectiveness in identifying true fall cases. These findings highlight the inherent trade-off between sensitivity and precision when applying resampling techniques in highly imbalanced clinical datasets.

Table 7.

Mean ± standard deviation of overall model performance metrics with SMOTENC across five-fold cross-validation for training, validation, and test sets.

Model	Mode	Accuracy	AUC_ROC	Specificity	Sensitivity
Linear Regression	Train	63.08% ± 0.38%	68.50% ± 0.37%	67.45% ± 0.12%	58.71% ± 0.68%
	Validation	67.39% ± 0.15%	67.49% ± 1.30%	67.44% ± 0.15%	57.92% ± 2.41%
	Test	67.35% ± 0.14%	68.32% ± 0.30%	67.38% ± 0.13%	61.29% ± 0.87%
Random Forest	Train	73.88% ± 0.07%	81.70% ± 0.20%	72.94% ± 0.74%	74.83% ± 0.65%
	Validation	72.76% ± 0.80%	67.08% ± 1.33%	72.88% ± 0.82%	50.15% ± 2.69%
	Test	72.82% ± 0.75%	67.60% ± 0.37%	72.93% ± 0.76%	50.96% ± 0.95%
XGBoost	Train	70.10% ± 0.29%	77.32% ± 0.16%	70.55% ± 0.68%	69.66% ± 0.48%
	Validation	70.48% ± 0.78%	69.19% ± 1.23%	70.55% ± 0.79%	56.74% ± 2.72%
	Test	70.54% ± 0.70%	70.03% ± 0.34%	70.61% ± 0.71%	57.65% ± 1.07%
SVM	Train	63.12% ± 0.33%	68.49% ± 0.37%	67.22% ± 0.16%	59.02% ± 0.60%
	Validation	67.14% ± 0.17%	67.51% ± 1.28%	67.19% ± 0.17%	58.16% ± 2.75%
	Test	67.12% ± 0.19%	68.32% ± 0.30%	67.14% ± 0.19%	61.76% ± 0.49%
MLP	Train	82.17% ± 0.37%	90.78% ± 0.37%	82.72% ± 1.37%	81.62% ± 1.29%
	Validation	82.16% ± 1.33%	50.23% ± 2.81%	82.45% ± 1.33%	26.29% ± 2.73%
	Test	82.13% ± 1.42%	48.71% ± 1.64%	82.43% ± 1.43%	25.48% ± 2.19%

Table 8.

Mean ± standard deviation of class-specific performance metrics with SMOTENC for fall and non-fall predictions across five-fold cross-validation.

Model	Mode	Precision_Fall	Recall_Fall	F1_Fall	Precision_NonFall	Recall_NonFall	F1_NonFall
Linear Regression	Train	64.33% ± 0.32%	58.71% ± 0.68%	61.39% ± 0.51%	62.03% ± 0.41%	67.45% ± 0.12%	64.62% ± 0.26%
	Validation	0.94% ± 0.04%	57.92% ± 2.41%	1.85% ± 0.08%	99.67% ± 0.02%	67.44% ± 0.15%	80.44% ± 0.11%
	Test	0.99% ± 0.02%	61.29% ± 0.87%	1.95% ± 0.03%	99.70% ± 0.01%	67.38% ± 0.13%	80.41% ± 0.10%
Random Forest	Train	73.44% ± 0.37%	74.83% ± 0.65%	74.13% ± 0.14%	74.35% ± 0.30%	72.94% ± 0.74%	73.63% ± 0.24%
	Validation	0.97% ± 0.03%	50.15% ± 2.69%	1.91% ± 0.05%	99.64% ± 0.02%	72.88% ± 0.82%	84.18% ± 0.54%
	Test	0.99% ± 0.03%	50.96% ± 0.95%	1.95% ± 0.06%	99.64% ± 0.01%	72.93% ± 0.76%	84.22% ± 0.51%
XGBoost	Train	70.29% ± 0.42%	69.66% ± 0.48%	69.97% ± 0.27%	69.93% ± 0.28%	70.55% ± 0.68%	70.23% ± 0.37%
	Validation	1.01% ± 0.04%	56.74% ± 2.72%	1.99% ± 0.08%	99.67% ± 0.02%	70.55% ± 0.79%	82.62% ± 0.54%
	Test	1.03% ± 0.02%	57.65% ± 1.07%	2.03% ± 0.04%	99.68% ± 0.01%	70.61% ± 0.71%	82.67% ± 0.48%
SVM	Train	64.29% ± 0.28%	59.02% ± 0.60%	61.54% ± 0.45%	62.13% ± 0.36%	67.22% ± 0.16%	64.57% ± 0.24%
	Validation	0.93% ± 0.05%	58.16% ± 2.75%	1.84% ± 0.09%	99.67% ± 0.02%	67.19% ± 0.17%	80.27% ± 0.12%
	Test	0.99% ± 0.01%	61.76% ± 0.49%	1.95% ± 0.02%	99.70% ± 0.00%	67.14% ± 0.19%	80.24% ± 0.13%
MLP	Train	82.55% ± 0.96%	81.62% ± 1.29%	82.07% ± 0.43%	81.83% ± 0.84%	82.72% ± 1.37%	82.27% ± 0.45%
	Validation	0.80% ± 0.11%	26.29% ± 2.73%	1.55% ± 0.22%	99.53% ± 0.02%	82.45% ± 1.33%	90.18% ± 0.80%
	Test	0.77% ± 0.07%	25.48% ± 2.19%	1.49% ± 0.13%	99.52% ± 0.01%	82.43% ± 1.43%	90.17% ± 0.86%

Table 9.

Mean ± standard deviation of confusion matrix components with SMOTENC across five-fold cross-validation for training, validation, and test sets.

Mode	Mode	TN	FP	FN	TP
Linear Regression	Train	207463.80 ± 360.93	100130.60 ± 360.76	127005.60 ± 2080.38	180588.80 ± 2080.10
	Validation	51857.20 ± 114.47	25041.40 ± 114.40	172.20 ± 9.71	237.00 ± 10.10
	Test	64771.00 ± 128.12	31353.00 ± 128.12	197.80 ± 4.44	313.20 ± 4.44
Random Forest	Train	224352.00 ± 2288.34	83242.40 ± 2288.79	77421.80 ± 1993.31	230172.60 ± 1993.78
	Validation	56041.00 ± 633.24	20857.60 ± 632.78	204.00 ± 10.93	205.20 ± 11.12
	Test	70105.40 ± 728.80	26018.60 ± 728.80	250.60 ± 4.88	260.40 ± 4.88
XGBoost	Train	216995.40 ± 2080.15	90599.00 ± 2080.54	93323.40 ± 1463.06	214271.00 ± 1463.14
	Validation	54254.40 ± 609.21	22644.20 ± 608.86	177.00 ± 11.11	232.20 ± 11.19
	Test	67876.40 ± 678.31	28247.60 ± 678.31	216.40 ± 5.46	294.60 ± 5.46
SVM	Train	206749.80 ± 482.30	100844.60 ± 482.02	126042.80 ± 1852.18	181551.60 ± 1851.85
	Validation	51665.40 ± 129.58	25233.20 ± 129.59	171.20 ± 11.12	238.00 ± 11.51
	Test	64541.40 ± 178.82	31582.60 ± 178.82	195.40 ± 2.51	315.60 ± 2.51
MLP	Train	254455.80 ± 4213.24	53138.60 ± 4213.05	56541.40 ± 3976.21	251053.00 ± 3975.87
	Validation	63404.80 ± 1023.83	13493.80 ± 1023.97	301.60 ± 10.92	107.60 ± 11.28
	Test	79238.80 ± 1379.36	16885.20 ± 1379.36	380.80 ± 11.17	130.20 ± 11.17

The SHAP summary plot ranks features according to their overall contribution to the model and illustrates the direction of their effects as shown in Figure 2. The Barthel Index total score was identified as the most important predictor. Lower functional scores were associated with positive SHAP values, indicating an increased risk of falls, whereas higher functional independence contributed to reduced risk. Age was the second most influential factor, demonstrating a clear positive relationship with fall risk, as higher age values were associated with increased SHAP values. Among functional variables, transfer ability, mobility, and activity level showed substantial contributions. Impairment in these functions was associated with higher SHAP values, indicating elevated fall risk. In addition, history of falls within one year exhibited a strong positive effect, with higher values contributing to increased risk. Other clinically relevant factors, including dizziness or vertigo, sensory perception, medication use, and caregiver status, also contributed to the model, although their overall impact was smaller compared to functional and mobility-related variables.

Figure 2.

SHAP summary plot illustrating feature importance and directional effects of predictors in the Random Forest model for inpatient fall risk prediction.

The SHAP summary plot illustrates both the relative importance of each predictor and the direction of its effect on inpatient fall risk as shown in Figure 3. The Barthel Index total score is identified as the most influential predictor. Lower values, indicating reduced functional independence, are associated with higher SHAP values and thus a higher predicted fall risk. This finding emphasizes that functional dependency is the dominant driver of inpatient falls. Age is the second most important feature and shows a clear monotonic pattern, where increasing age corresponds to progressively higher SHAP values. This indicates that older patients have a substantially increased risk of falls, consistent with clinical evidence. Among functional variables, transfer ability, mobility, and activity level demonstrate strong contributions. Lower functional performance in these domains is associated with positive SHAP values, indicating increased fall risk. Similarly, sensory perception shows a meaningful effect, suggesting that impaired perception contributes to instability and fall occurrence.

Figure 3.

SHAP summary plot illustrating feature importance and direction of impact on inpatient fall risk prediction using the XGBoost model.

Discussion

This study compared the performance of multiple machine learning methods for predicting fall risk among hospitalized patients and further examined the impact of SMOTENC on addressing class imbalance. The results indicate that model performance varied substantially depending on the evaluation metric used. In the absence of SMOTENC, Random Forest and XGBoost demonstrated relatively stable discrimination performance, with AUC-ROC values of approximately 0.62–0.72. However, despite moderate discrimination, most models exhibited extremely limited ability to identify fall events, as reflected by near-zero recall and F1-scores. This suggests that high overall accuracy in imbalanced datasets may be misleading and does not necessarily reflect clinically meaningful predictive performance. Similar findings have been reported in previous machine learning–based fall prediction studies, where severe class imbalance frequently resulted in high overall accuracy but poor minority-class detection performance. In clinical fall prediction, optimizing sensitivity for rare fall events often leads to a substantial increase in false-positive predictions, reflecting the inherent recall–precision trade-off commonly observed in imbalanced healthcare datasets. Therefore, evaluation of fall prediction models should not rely solely on overall accuracy or AUC-ROC, but also incorporate clinically relevant metrics such as recall, precision, F1-score, and precision–recall curves to better assess minority-event prediction performance. After applying SMOTENC, model behavior changed notably. Recall for fall events increased substantially across Logistic Regression, Random Forest, XGBoost, and SVM, reaching approximately 50–60%, indicating improved sensitivity to minority cases. This improvement was accompanied by a pronounced reduction in precision, which remained below 1% in the test and validation sets. Consequently, F1-scores for fall prediction remained low, reflecting the trade-off between detecting minority events and increased false positive predictions. These findings suggest that while SMOTENC effectively mitigates prediction collapse, it does not fully resolve the challenges associated with imbalanced clinical data. This observation suggests that algorithmic resampling alone may be insufficient for achieving clinically deployable fall prediction performance in extremely imbalanced inpatient datasets, highlighting the need for multimodal feature integration and more clinically informative representations of patient mobility and functional decline. Random Forest and XGBoost demonstrated relatively consistent performance across training, validation, and test sets, indicating better robustness compared with other approaches. In contrast, MLP exhibited strong performance in the training set but showed reduced AUC-ROC and recall in the test and validation sets, suggesting limited generalization under resampled conditions. Logistic Regression and SVM showed improved recall after SMOTENC but remained constrained by low precision, limiting their practical applicability.

To further enhance model interpretability, SHAP analysis was conducted for tree-based models. The results consistently demonstrated that functional status, particularly the Barthel Index, was the most influential predictor, followed by age and mobility-related variables such as transfer ability, activity level, and mobility. Lower functional independence and impaired mobility were strongly associated with increased fall risk, highlighting that functional decline is the primary driver of inpatient falls. In addition, the Random Forest model emphasized the importance of prior fall history, suggesting that historical events contribute additional predictive value. In contrast, variables such as medication use, dizziness, and caregiver status showed comparatively smaller contributions in the SHAP analysis. These findings indicate that, although such clinical factors are relevant, their predictive impact is secondary to functional and mobility-related impairments in this dataset. The SHAP results provide clinically interpretable evidence that fall risk is predominantly driven by functional dependency and mobility limitations rather than isolated clinical conditions. This aligns with the model performance findings and reinforces the importance of incorporating functional assessments into fall risk prediction frameworks. These findings are consistent with recent studies demonstrating that multidimensional functional assessments, including functional balance tests and computerized dynamic posturography, can substantially improve both predictive performance and clinical interpretability in fall risk prediction models. Previous machine learning studies have shown that combining physical characteristics, co-morbidities, and objective balance measurements enables more comprehensive characterization of mobility impairment and balance dysfunction. Importantly, although advanced balance systems may enhance predictive accuracy, prior studies have also suggested that simpler functional assessments combined with routinely available clinical variables may still achieve strong predictive capability without requiring specialized equipment or additional bedside procedures.¹⁹ In this context, the present study further supports the clinical value of functional and mobility-related variables derived from routine inpatient assessment workflows.

This study compared the characteristics of patients with and without falls, demonstrating that multiple demographic, clinical, and functional factors were significantly associated with fall events. Among demographic variables, the proportion of males was significantly higher in the fall group, consistent with prior studies suggesting that male inpatients may be at increased fall risk in certain populations due to physical or functional conditions.²⁰ The higher prevalence of smoking among fall patients also aligns with evidence linking smoking to an increased likelihood of falls.²¹ Clinically, a history of falls within the past year emerged as a strong predictor, in agreement with previous findings.²² The use of sedatives, psychiatric medications, antidiabetic drugs, antiepileptics, and hypnotics was significantly more frequent in the fall group, supporting established evidence that these medications elevate fall risk.^23,24 Furthermore, the higher prevalence of polypharmacy highlights medication burden as another potential risk factor. In terms of functional and sensory status, dizziness, abnormal foot sensation, and impairments in vision and hearing were all more common in the fall group, underscoring the role of sensory deficits as important fall risk factors.²⁵ Significant differences were also observed in activities of daily living (ADL), including personal hygiene, bathing, feeding, toileting, and dressing, with fall patients demonstrating greater dependency. This finding is consistent with prior studies showing that ADL assessments are valuable tools for identifying high-risk inpatients.²⁶ Mobility-related factors, such as unstable gait, the need for walking aids, and greater difficulty with stair climbing and level walking, were also markedly higher in the fall group, further confirming the strong link between reduced mobility and fall risk. Additionally, patients in the fall group were significantly older, reaffirming advanced age as a critical determinant of fall risk. Age-related declines in muscle strength, balance, and the presence of multiple comorbidities are likely contributors to this association.²⁷ Lower body weight and BMI in the fall group further suggest that malnutrition and reduced muscle mass may play a role in increasing fall susceptibility. Low BMI and weight loss have been closely associated with sarcopenia, impaired balance, and elevated fall risk.²⁸ Finally, the fall group had significantly lower Barthel Index scores compared with the non-fall group, indicating greater dependence in daily activities. As a widely used clinical measure, the Barthel Index effectively reflects patient self-care ability and further emphasizes the connection between functional limitations and fall events. The findings suggest that routinely collected functional assessments, particularly measures related to mobility and activities of daily living, may provide substantial value for early inpatient fall risk stratification. Because these variables are already embedded within standard nursing assessment workflows, they may facilitate scalable implementation of automated fall risk screening systems without imposing additional burden on healthcare providers. Compared with studies incorporating specialized balance assessment systems or sensor-based monitoring technologies, the present study focused on routinely collected electronic medical record variables obtained during standard inpatient care. Although this approach may provide lower predictive precision than highly instrumented assessment frameworks, it offers important advantages in scalability, automation, and feasibility for large-scale hospital deployment. These findings suggest that clinically accessible functional and mobility assessments remain highly valuable for practical inpatient fall risk prediction in real-world healthcare settings.

Table 10 compares the present study with recent machine learning–based studies for inpatient fall risk prediction. Previous studies generally reported moderate to high discrimination performance (AUC or C-index ≈ 0.70–0.80) and consistently identified functional status, mobility, and medication-related variables as key predictors; most did not explicitly address class imbalance or report class-specific performance metrics. The present study, based on a large-scale and highly imbalanced inpatient dataset (N = 483,174), demonstrates that models without resampling exhibit severe degradation in fall detection despite moderate AUC-ROC performance (∼0.62–0.72), with recall approaching zero, indicating prediction collapse toward the majority class. The application of SMOTENC substantially improved recall to approximately 50–60%, but at the cost of extremely low precision (<1%) and persistently low F1-scores, reflecting a pronounced trade-off between sensitivity and false positives. Further application of SMOTENC improved model stability and achieved a more balanced performance by accounting for mixed-type clinical features, particularly in tree-based models such as Random Forest and XGBoost. Unlike many previous studies that primarily emphasized discrimination performance alone, the present study specifically investigated the impact of severe class imbalance on clinically meaningful minority-event prediction performance. These findings highlight that model performance in inpatient fall prediction is primarily constrained by class imbalance rather than model selection, and underscore the necessity of evaluating models using both AUC-ROC and class-specific metrics, as well as adopting feature-aware resampling strategies to achieve clinically meaningful performance.

Table 10.

Comparison of recent machine learning studies and the present study for inpatient fall risk prediction.

Study	Sample size	Key variables	Model	Discrimination (AUC/C-index)	Key insight
Yang et al. (2026)²⁹	—	Barthel Index, dizziness, fall history, mobility	GBM, RF, SVM, LR	C-index ≈ 0.74	Functional status and dizziness are key predictors
Jahandideh et al. (2024)³⁰	—	Clinical + functional	DNN, RF	High (reported)	Deep learning captures nonlinear patterns
Kim et al. (2025)¹³	—	Frailty, function, comorbidities	ML models	Moderate–high	Frailty strongly associated with falls
Kang et al. (2025)³¹	—	Medications, demographics	Gradient boosting	AUC ∼0.75–0.80	Medication-related risk important
Nishino et al. (2025)³²	83,917	Clinical + ADL + medications	LR, XGB, LGBM	High discrimination	Multivariable integration improves prediction
Jahangiri et al. (2024) ¹²	—	Mobility, ADL, clinical	ML models	AUC ∼0.70+	Mobility and ADL are dominant predictors
This study (No SMOTE)	483,174	Age, Barthel Index, ADL, meds, sensory, fall history	LR, RF, SVM, XGB, MLP	AUC ∼0.62–0.72	Severe imbalance leads to prediction collapse
This study (SMOTENC)	483,174	Same	ML models	AUC ∼0.68–0.72	Better stability and feature-aware resampling

Despite the promising results, several limitations merit consideration. The use of retrospective clinical data introduces a potential risk of data leakage, including both temporal and target leakage, as certain variables may implicitly encode information related to the outcome, thereby leading to overestimated model performance. Although efforts were made to appropriately structure the dataset, the possibility of unintended information leakage cannot be entirely excluded. The application of the SMOTENC algorithm may also introduce bias if not rigorously controlled, as synthetically generated samples may not fully capture the underlying clinical heterogeneity, and inappropriate implementation before dataset partitioning may further exacerbate information leakage between training and validation sets. Although accuracy and AUC-ROC remain widely used metrics in machine learning studies, these measures may inadequately reflect minority-event prediction performance in highly imbalanced clinical datasets because they are strongly influenced by majority-class dominance.

Future investigations should incorporate more discriminative evaluation metrics, including precision–recall curves, F1-score, and class-specific recall, to more accurately reflect minority class performance in highly imbalanced fall prediction datasets. In addition, the limited interpretability of the developed models may constrain their clinical applicability because the underlying decision-making processes remain insufficiently transparent; therefore, the integration of explainable artificial intelligence techniques is warranted to improve model transparency, enhance clinician trust, and support real-world deployment. Future studies may also further improve predictive performance by integrating wearable sensor data, computerized dynamic posturography, or longitudinal mobility monitoring, as suggested in previous fall prediction research.¹⁹ Combining routinely collected electronic medical record data with objective balance assessments and real-time movement analysis may enable more comprehensive characterization of patient mobility, balance dysfunction, and functional decline, thereby enhancing both predictive performance and clinical interpretability. Furthermore, the incorporation of continuous monitoring technologies and explainable artificial intelligence frameworks may facilitate earlier intervention, individualized risk stratification, and more effective clinical decision-making in hospital settings. Future multicenter prospective studies with external validation are warranted to further evaluate the generalizability, robustness, and clinical applicability of machine learning–based fall prediction systems across diverse patient populations and healthcare environments.

Conclusions

This study integrated multidimensional clinical characteristics with multiple machine learning models to develop a predictive framework for inpatient fall risk. The findings confirmed that prior fall history, medication use, reduced functional independence, impaired mobility, and sensory deficits are key determinants of fall events. Model comparisons showed that Random Forest and XGBoost achieved relatively stable performance, with moderate AUC-ROC values of approximately 0.62–0.72, indicating acceptable but limited discrimination ability. Under severe class imbalance, most models demonstrated limited capability in identifying fall events, as reflected by extremely low recall and F1-scores despite moderate AUC-ROC performance. After applying SMOTENC, recall for fall events improved substantially to approximately 50–60% across several models; however, precision remained below 1%, resulting in persistently low F1-scores. In addition, certain models, such as MLP, exhibited unstable generalization performance, with reduced AUC-ROC on the test and validation sets. These findings indicate that, although machine learning models can capture clinically relevant risk factors, their effectiveness in detecting rare events is strongly constrained by class imbalance. While SMOTENC alleviates prediction collapse and improves sensitivity, it introduces a notable trade-off with precision. Therefore, model performance in imbalanced clinical datasets should be evaluated using both discrimination metrics, such as AUC-ROC, and class-specific measures, with appropriate imbalance handling strategies considered essential for improving clinical applicability.

Supplemental material

Supplemental material - Admission clinical and functional assessments for predicting inpatient fall risk using machine learning models

Supplemental material for Admission clinical and functional assessments for predicting inpatient fall risk using machine learning models by Ya-Wen Lee, Ying-Lin Hsu, Shu-Mei Lai, Chih-Hao Lin, Chih-Ming Lin, Jia-Lang Xu in DIGITAL HEALTH

Footnotes

Acknowledgments

This study acknowledge the use of ChatGPT for language editing support during manuscript preparation. The authors retained full responsibility for the content and accuracy of the manuscript. The authors would like to thank the referees and the editors for their comments and valuable suggestions.

ORCID iD

Jia-Lang Xu

Author Contributions

Conceptualization, Y.-W.L. and J.-L.X.; methodology, Y.-L.H.; validation, C.-M.L. and S.-M.L.; formal analysis, Y.-L.H.; investigation, S.-M.L.; data curation, J.-L.X. and C.-H.L.; writing—original draft preparation, Y.-L.S. and J.-L.X.; writing—review and editing, Y.-W.L., C.-M.L., Y.-L.S. and J.-L.X.; supervision, C.-M.L. All authors have read and agreed to the published version of the manuscript.

Funding

The authors wish to thank the National Science and Technology Council (NSTC) of the Republic of China for financially supporting this research under Contract Grants (No. NSTC-115-2222-E-025-001).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Supplemental material

Supplemental material for this article is available online.

References

Cameron

Dyer

Panagoda

, et al. Interventions for preventing falls in older people in care facilities and hospitals. Cochrane Database Syst Rev 2018; 9: CD005465.

Rubenstein

. Falls in older people: epidemiology, risk factors and strategies for prevention. Age Ageing 2006; 35(Suppl 2): ii37–ii41. https://doi.org/10.1093/ageing/afl084

Bouldin

Andresen

Dunton

, et al. Falls among adult patients hospitalized in the United States: prevalence and trends. J Patient Saf 2013; 9(1): 13–17. https://doi.org/10.1097/PTS.0b013e3182699b64

Wong

Recktenwald

Jones

, et al. The cost of serious fall-related injuries at three Midwestern hospitals. Jt Comm J Qual Patient Saf 2011; 37(2): 81–87. https://doi.org/10.1016/s1553-7250(11)37010-9

Ganz

Huang

Saliba

, et al. Preventing falls in hospitals: a toolkit for improving quality of care. Ann Intern Med 2013; 158(5 Pt 2): 390–396.

Andrade

. Sedative hypnotics and the risk of falls and fractures in the elderly. J Clin Psychiatry 2018; 79(3): 19106. https://doi.org/10.4088/JCP.18f12340

Ladios-Martin

Cabañero-Martínez

Fernández-de-Maya

, et al. Development of a predictive inpatient falls risk model using machine learning. J Nurs Manag 2022; 30(8): 3777–3786. https://doi.org/10.1111/jonm.13760

Liu

Lin

. A machine learning–based fall risk assessment model for inpatients. Comput Inform Nurs 2021; 39(8): 450–459. https://doi.org/10.1097/CIN.0000000000000727

Chen

. Applying artificial intelligence to predict falls for inpatient. Front Med 2023; 10: 1285192. https://doi.org/10.3389/fmed.2023.1285192

10.

Chen

Luo

Chang

, et al. Post-surgical fall risk prediction: a machine learning approach for spine and lower extremity procedures. Front Med 2025; 12: 1574305. https://doi.org/10.3389/fmed.2025.1574305

11.

Shao

Wang

Xie

, et al. Development and external validation of a machine learning–based fall prediction model for nursing home residents: a prospective cohort study. J Am Med Dir Assoc 2024; 25(9): 105169. https://doi.org/10.1016/j.jamda.2024.105169

12.

Jahangiri

Abdollahi

Patil

, et al. An inpatient fall risk assessment tool: application of machine learning models on intrinsic and extrinsic risk factors. Mach Learn Appl 2024; 15: 100519. https://doi.org/10.1016/j.mlwa.2023.100519

13.

Kim

Seo

Kwon

, 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(1): e13713. https://doi.org/10.1002/jcsm.13713

14.

LaValley

. Logistic regression. Circulation 2008; 117(18): 2395–2399. https://doi.org/10.1161/CIRCULATIONAHA.106.682658

15.

Breiman

. Random forests. Mach Learn 2001; 45(1): 5–32. https://doi.org/10.1023/a:1010933404324

16.

Chen

Guestrin

. XGBoost: a scalable tree boosting system. Proc 22nd ACM SIGKDD Int Conf Knowl Discov Data Min, 2016, pp. 785–794.

17.

Suthaharan

. Support vector machine. Machine learning models and algorithms for big data classification. Boston: Springer, 2016, pp. 207–235.

18.

Popescu

Balas

Perescu-Popescu

, et al. Multilayer perceptron and neural networks. WSEAS Trans Circuits Syst 2009; 8(7): 579–588.

19.

Soylemez

Tokgoz-Yilmaz

. Predicting fall risk in elderly individuals: a comparative analysis of machine learning models using patient characteristics, functional balance tests and computerized dynamic posturography. J Laryngol Otol 2025; 139(6): 464–472. https://doi.org/10.1017/S0022215124002160

20.

Kubo

Fujii

Hayashi

, et al. Sex differences in modifiable fall risk factors. J Nurse Pract 2021; 17(9): 1098–1102. https://doi.org/10.1016/j.nurpra.2021.06.016

21.

Kimura

Suzuki

Kitamura

, et al. Smoking, alcohol consumption, and risk of recurrent falls in community-dwelling Japanese people aged 40–74 years: the Murakami cohort study. Geriatr Gerontol Int 2025; 25(1): 67–74. https://doi.org/10.1111/ggi.15040

22.

Deandrea

Lucenteforte

Bravi

, et al. Risk factors for falls in community-dwelling older people: a systematic review and meta-analysis. Epidemiology 2010; 21(5): 658–668. https://doi.org/10.1097/EDE.0b013e3181e89905

23.

Woolcott

Richardson

Wiens

, et al. Meta-analysis of the impact of nine medication classes on falls in elderly persons. Arch Intern Med 2009; 169(21): 1952–1960. https://doi.org/10.1001/archinternmed.2009.357

24.

Seppala

Wermelink

de Vries

, et al. Fall-risk-increasing drugs: a systematic review and meta-analysis II. Psychotropics. J Am Med Dir Assoc 2018; 19(4): 371.e11–371.e17. https://doi.org/10.1016/j.jamda.2017.12.098

25.

Lord

. Visual risk factors for falls in older people. Age Ageing 2006; 35(Suppl 2): ii42–ii45. https://doi.org/10.1093/ageing/afl085

26.

Oliver

Connelly

Victor

, et al. Strategies to prevent falls and fractures in hospitals and care homes and effect of cognitive impairment: systematic review and meta-analyses. BMJ 2007; 334(7584): 82. https://doi.org/10.1136/bmj.39049.706493.55

27.

Ambrose

Paul

Hausdorff

. Risk factors for falls among older adults: a review of the literature. Maturitas 2013; 75(1): 51–61. https://doi.org/10.1016/j.maturitas.2013.02.009

28.

Clynes

Edwards

Buehring

, et al. Definitions of sarcopenia: associations with previous falls and fracture in a population sample. Calcif Tissue Int 2015; 97(5): 445–452. https://doi.org/10.1007/s00223-015-0044-z

29.

Yang

Ren

, et al. Development and validation of machine learning models for predicting falls among hospitalized older adults: retrospective cross-sectional study. JMIR Aging 2026; 9(1): e80602. https://doi.org/10.2196/80602

30.

Jahandideh

Hutchinson

Bucknall

, et al. Using machine learning models to predict falls in hospitalised adults. Int J Med Inform 2024; 187: 105436. https://doi.org/10.1016/j.ijmedinf.2024.105436

31.

Kang

Yan

Tian

, et al. Constructing a fall risk prediction model for hospitalized patients using machine learning. BMC Public Health 2025; 25(1): 242. https://doi.org/10.1186/s12889-025-21284-8

32.

Nishino

Matsuyama

Miyagi

, et al. Machine learning–based prediction of in-hospital falls in adult inpatients: retrospective observational multicenter study. JMIR Med Inform 2025; 13: e75958. https://doi.org/10.2196/75958

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.07 MB

0.00 MB