Dynamic fall risk prediction in hospitalized cancer patients: development and validation of a machine learning model using multidimensional clinical data to overcome over-sensitivity in traditional scales

Abstract

Objective

Traditional risk assessment tools were commonly used for the prediction of falls in clinical practice. However, their tendency toward over-sensitivity addressed its limitations for hospitalized cancer patients. This study aimed to identify risk factors associated with falls in hospitalized patients and to develop a predictive risk model using machine learning algorithms specifically for cancer patients, to optimize the clinical assessment of fall risk.

Methods

Using data from 298 cancer patients admitted to Shandong Cancer Hospital between January 2023 and December 2024 were included. We described the patient status through three aspects: clinical indicators, laboratory indicators and subjective evaluation indicators. The model was evaluated for performance using five-fold cross-validation. Least absolute shrinkage and selection operator (LASSO) regression was used for initiatory variable selection. We constructed predictive models using one baseline method and three machine learning methods, incorporating key clinical features, and selected the optimal model based on a review of categorical performance metrics. Additionally, Shapley Additive Explanations (SHAP) was used to interpret the predictive models and rank the importance of risk factors.

Results

Among 298 cancer patients assessed by the Morse Fall Scale, only 115 (39%) experienced falls during hospitalization. Based on LASSO regression and statistical analysis, 20 key predictors were identified from 30 candidate features. The Support Vector Machine model outperformed the other algorithms (Naive Bayes, Multivariate Logistic Regression, and Random Forest), achieving the best performance in both the low-risk and high-risk populations evaluated by the scale, with a composite F1-score of 0.87. SHAP interpretability analysis revealed that medication use, diagnostic findings, and treatment approaches were major risk factors, while the number of comorbidities further increased susceptibility to falls. The model’s precision in stratifying dynamic risks, particularly for high-risk subgroups (e.g., patients with brain metastases or neurotoxic regimens), highlights its clinical utility.

Conclusion

This study establishes a clinically interpretable, machine learning-driven fall risk prediction model that dynamically integrates cancer-specific variables. Implementation of this model could enhance risk stratification accuracy, guide personalized intervention strategies, and optimize resource allocation in clinical oncology care settings.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12911-025-03213-2.

Introduction

Falls are defined as an unexpected change in position that is not caused by an inherent event (such as a brain hemorrhage) or substantial hazard, which lead to an individual to land on the ground, floor or a lower level [1]. Inpatient falls represent a critical healthcare challenge due to the prolonged hospitalization, increased morbidity, and elevated medical costs [2, 3]. Cancer patients constitute a high-risk population for falls. On one hand, the tumor itself can lead to physical exertion and some comorbidities such as brain and bone metastases can increase the risk of falls. On the other hand, cancer treatment related factors, particularly chemotherapy and radiotherapy, may lead to fatigue, cognitive impairment, electrolyte imbalances, and myelosuppression, thus exacerbating fall susceptibility in cancer patients [4]. It is reported that the incidence of falls during chemotherapy is 35% [5]. Thus, the early identification of fall risks is essential for individual intervention in cancer patients to mitigate the subsequent healthcare challenge.

The most commonly used fall scoring systems in clinical care include the STRATIFY Fall Assessment Scale [6], the Hendrich II Fall Risk Model [7]and the Morse Fall Scale [8]. However, traditional fall risk assessment tools are limited by their dependence on subjective clinician evaluations and static parameters [9], which fail to capture dynamic, treatment-related risk factors in cancer patients [10]. Additionally, these tools demonstrate only moderate reliability in general populations, not specifically in cancer patients, potentially reducing their sensitivity due to inadequate incorporation of cancer-specific variables. We found through a review of electronic medical records that the majority of cancer patients were assessed as high-risk using the Morse Fall Scale. This uniform risk stratification hinders nursing staff from implementing targeted interventions, consequently leading to insufficient attention being paid to this patient population in clinical practice. A systematic review by Greeley AM [11] et al. indicates that existing scales fail to account for transient physiological changes, such as chemotherapy-induced anemia or hyponatremia, leading to delayed intervention and preventable adverse events. Studies show that multicomponent programs to prevent falls among inpatients can reduce relative risk for falls by as much as 30% [12]. Moreover, manual assessments are labor-intensive, prone to inter-observer variability, and lack scalability for clinical implementation [13].

Machine learning (ML) offers a transformative approach by utilizing high-dimensional data to reveal nonlinear interactions which can help to predict the dynamic risk patterns. ML applications have shown superior performance compared to conventional methods, as demonstrated by readmission risk prediction models, indicating its ability in handling complex datasets and generating real-time predictions [14–16]. Moreover, existing studies frequently overlook treatment-related variables, or fail to address the temporal variability of risk factors [17]. A meta-analysis further emphasized that ML models incorporating dynamic biomarkers enhance predictive accuracy of the risk for fall by 15–20% compared to static assessments [18, 19]. Specifically, in fall prevention, Lee et al. developed an ML model integrating vital signs and medication data to predict falls in elderly inpatients, achieving an area under the Receiver Operating Characteristic (ROC) curve of 0.82, significantly outperforming traditional scales [20]. Despite these advancements, limited research has focused on cancer-specific fall prediction.

Several studies have explored fall risk prediction using machine learning in general hospitalized or elderly populations, but few have focused specifically on cancer patients. For example, Liu et al. [21]. used logistic regression to predict falls in older adults hospitalized with tumors, achieving an AUROC of 0.920 based on clinical indicators such as fall history and muscle thickness. However, their single-center study lacked external validation and cancer-specific subgroup analysis. In contrast, Heo et al. [22]. developed machine learning models using nationwide claims data and identified medication use and healthcare utilization as key predictors, but their model was not tailored to oncology patients and excluded functional metrics. Similarly, Chen et al. [19]. applied interpretable machine learning on a large Chinese cohort, finding fall history and ADL-related features most predictive, though their model targeted community-dwelling older adults, not those with cancer. Other studies focused on gait-based trip-fall prediction achieved high accuracy, but were limited to controlled settings with healthy participants. Overall, despite progress in fall-risk modeling, few tools have been validated for use in oncology care, particularly during active treatment or hospitalization.

This study proposes an interpretable machine learning model to dynamically predict fall risk in hospitalized cancer patients undergoing chemoradiotherapy. By integrating multidimensional clinical data—including cancer-specific factors such as chemotherapy and immunotherapy—we employed advanced ensemble methods alongside Shapley Additive Explanations (SHAP)-based interpretation to enhance transparency and clinical relevance. Stratified evaluations across treatment subgroups and distinct risk populations under a five-fold cross-validation framework demonstrated the models’ robustness. Unlike traditional static assessment tools, our approach offers dynamic risk stratification and clinical insight, effectively addressing the limited sensitivity of existing fall risk scales in cancer inpatients.

Methods

Study population

We conducted a retrospective cohort study utilizing electronic medical records of cancer patients admitted to Shandong Cancer Hospital from January 2023 to December 2024. Inclusion criteria were as follows: age ≥ 18 years, hospitalization of more than 24 h, histologically confirmed diagnosis of malignancy, and ongoing radiotherapy. Patients were excluded if they had: (1) acute neurological events that precipitated falls (such as stroke or seizure), (2) missing key variables (with more than 20% of laboratory or treatment records absent), or (3) non-cancer-related admissions. We collected 30 candidate features that had previously been reported as risk factors for falls, which were captured at the Adverse Events System. These features included demographics (age, sex), healthcare utilization pattern, specific diagnoses, other medications that increase/decrease the incidence of falls. The temporal distribution of patients with falls is shown in Figure A1.

Data preprocessing

Variables with missing values exceeding 20% were excluded from the analysis. For continuous variables (e.g., hemoglobin level, potassium), missing values were imputed using the median of the observed data. For categorical variables (e.g., treatment modality, occupational categories), missing values were replaced with the mode (most frequent category).

Statistical methods

We employed LASSO regression with ten-fold cross-validation for variable selection, chosen for its ability to manage multicollinearity and perform automatic feature selection via L1 regularization. To ensure all potential predictors were available for this process, we did not perform preliminary feature screening using methods such as univariate analysis.

Development and validation of risk prediction models

Hyperparameter tuning for ML models, including Random Forest (RF), Naive Bayes (NB), Support Vector Machine (SVM), was performed using grid search in conjunction with 5-fold cross-validation on the dataset to enhance model generalizability. Meanwhile, we used Multivariate Logistic Regression (MLR) as a baseline method for control.

We initially assessed model performance using ROC curves and Precision-Recall (PR) Curve with 95% confidence intervals in validation set. Subsequently, decision curve analysis was performed to evaluate the clinical utility of each model and to determine the optimal final model. Finally, SHAP analysis was used to identify key risk factors and visualize the models. All analyses were conducted using Python (scikit-learn version 1.2.2) and R (version 4.3.1), with random seeds set to ensure reproducibility.

Results

Baseline characteristics

Among the 30 candidate variables potentially associated with falls, each multi-categorical variable was decomposed into a set of dichotomous dummy variables. A multi-categorical variable was retained in the model if at least one of its dummy variables had a non-zero coefficient. Dichotomous and continuous variables were also retained based on the same criterion. As shown in Fig. 1, This approach yielded a total of 20 valid input variables. Twenty final variables were screened by LASSO regression. These features were prioritized for model construction based on their clinical relevance to fall risk in cancer patients.

Fig. 1.

Variable Selection via Lasso Regression with 10-Fold Cross-Validation. Lambda (λ): A regularization parameter controlling the strength of penalty applied to non-zero coefficients in LASSO regression. The vertical dashed lines indicate lambda values (log-scaled) during regularization. The leftmost line corresponds to the minimum mean squared error (MSE), while the rightmost line represents the optimal lambda (λ = 0.004) that retained 20 non-zero coefficients (highlighted variables)

Model performance

As shown in Fig. 2, we evaluated the performance of four machine learning models—NB, SVM, MLR, and RF—in predicting patient fall risk using 5-fold cross-validation. Model performance was assessed using the mean area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC), along with their respective 95% confidence intervals. Among these models, MLR, SVM, and RF demonstrated superior performance in identifying patients at high risk of falls. In contrast, the NB showed comparatively lower predictive performance. Additionally, the optimal classification probability thresholds for each model were determined using the Youden’s index, and the corresponding classification metrics are presented in Table 1.

Fig. 2.

ROC curves and PR curves for model performance comparison and threshold selection. Naive Bayes: (a) AUROC = 0.825, (b) AUPRC = 0.728, (c) Threshold = 0; Support Vector Machine: (d) AUROC = 0. 958, (e) AUPRC = 0.950, (f) Threshold = 0.423; Logistic Regression: (g) AUROC = 0.917, (h) AUPRC = 0.873, (i) Threshold = 0.048; RF: (j) AUROC = 0.962, (k) AUPRC = 0.956, (l) Threshold = 0.475

Table 1.

The classification performance of models

Model	PPV	NPV	Accuracy	Sensitivity	Specificity	F1-score
SVM	0.87	0.94	0.92	0.91	0.92	0.87
NB	0.63	0.80	0.73	0.69	0.75	0.50
RF	0.89	0.93	0.92	0.88	0.93	0.89
MLR	0.74	0.93	0.84	0.81	0.90	0.80

Subsequently, we evaluated the effectiveness of the different models in identifying and intervening in fall risk using decision curve analysis. The RF, NB, MLR, and SVM models were compared against three reference strategies: the existing Morse Fall Scale, a total intervention (“Treat-all”) and a no-intervention approach (“None”). As shown in Fig. 3, the RF model demonstrated a clear clinical advantage by providing a wider range of decision thresholds with comparable or greater clinical benefit.

Fig. 3.

Decision curves for four ML models with the Morse scale

Analysis of key factors for hospitalize cancer patients

To further investigate the factors contributing to patient falls, we analyzed variable importance using the best-performing RF and SVM models, with absolute SHAP values serving as indicators. As illustrated in Fig. 4. Both analytical approaches consistently identified medication use, and treatment patterns as the most influential factors associated with fall risk. Notably, medication use emerged as a particularly impactful determinant, exhibiting a SHAP value exceeding 0.12, significantly above the average.

Fig. 4.

(a) The importance of the features in SVM model. (b) The importance of the features in RF model

As shown in Fig. 5, medication use was significantly associated with a reduced risk of patient falls. Additionally, patients diagnosed with lung or esophageal cancer exhibited a lower incidence of falls compared to those with other medical conditions. Notably, symptoms such as pain and dizziness were linked to an increased risk of falling. Furthermore, Hypokalemia and decreased hemoglobin levels exacerbate fall risk. Finally, the SHAP analysis showed that patients receiving the combination therapy would be at greater risk for falls.

Fig. 5.

SHAP value analysis of the model input factors, where (a) represents the overall SHAP force analysis, and (b) shows the substantive impact of each factor in an individual case

Discussion

Current fall risk prediction in hospitalized cancer patients predominantly depends on two suboptimal approaches: traditional assessment scales and resource-intensive monitoring devices. In particular, the overly sensitive high-risk classifications generated by these scales may divert the attention of clinical healthcare providers, thus making it difficult to promptly identify patients with a genuinely higher risk of falls. Therefore, it is crucial to develop an easy-to-operate and accurate machine learning prediction model, specifically for cancer patients undergoing radiotherapy and/or chemotherapy, to optimize the precise identification of those at high risk of falls in current clinical practice. This model should seamlessly integrate clinical practicality with computational complexity, thereby addressing the existing limitations in fall risk management. Through a 24-month retrospective study of chemotherapy/radiotherapy inpatients with fall incidents, we developed an AI-powered fall prediction model by incorporating rigorous feature selection and systematic parameter tuning, establishing a methodological framework that bridges clinical epidemiology with ML optimization.

In addition, we have established a rigorous benchmarking framework based on the AUROC, AUPRC and F1-score. The SVM model demonstrated superior discriminative performance with an AUROC of 0.958 and AUPRC of 0.950, significantly outperforming comparator models in both precision and sensitivity earlier ML frameworks such as Lee et al. [20], who reported an AUROC of 0.82 in elderly inpatients. This improvement stems from the deliberate inclusion of oncology-specific predictors (e.g., tumor location, neurotoxic chemotherapy regimens) that capture dynamic, treatment-related risk trajectories. For example, cisplatin-induced neuropathy directly compromises proprioceptive feedback, while brain metastases may precipitate acute ataxia through mass effect—mechanisms absent in non-cancer populations. Crucially, his predictive advantage translates to clinically actionable timeframes - high-risk identification positive predictive value, enabling nursing teams to implement targeted mobility monitoring prior to adverse events. Our ML framework addresses this critical gap by integrating real-time biomarker trajectories to recalibrate risk stratification. It achieves critically high sensitivity (minimizing missed cases) while maintaining strong precision and specificity, ensuring high accuracy. This balance enables reliable clinical deployment for targeted prevention without excessive false alarms. In addition, as shown in Table A2 in the Appendix, the results of the stratified assessment of the low-risk population indicate that the SVM is effective in identifying risk factors for patient falls and making accurate judgments.

Considering that SVM do not have built-in importance analysis, we used the kernel interpreter to construct SHAP analyses to characterize the best performing RF and SVM. Such methodological rigor not only confirms established clinical predictors but further identified three cardinal determinants demonstrating strongest fall associations: medication use, treatment modality, diagnosis and the number of comorbidities.

The SHAP analysis further elucidated pharmacological risk pathways. Chemotherapy patients frequently experience nausea and vomiting, which are commonly managed with 5-HT3 receptor antagonists (e.g., ondansetron) and neurokinin-1 (NK-1) receptor antagonists (e.g., aprepitant) [23, 24]. Studies demonstrate that addressing treatment-related adverse effects, such as those occurring during radiotherapy or chemotherapy, with symptomatic medications (e.g., antiemetics) can reduce fall risk by alleviating nausea and vomiting. However, these agents may also introduce adverse effects like constipation, dizziness, drowsiness, and sedation, which paradoxically increase fall susceptibility. Critically, research reveals that patients concurrently using ≥ 4 high-risk medications (e.g., sedatives, antipsychotics, anticholinergics, and antihypertensives) alongside symptomatic therapies face a significantly elevated fall risk. The multifaceted pharmacodynamic effects of psychotropic drugs, particularly their role as key predictors of falls, exacerbate extrapyramidal symptoms (e.g., gait instability, tremors) in cancer patients, further compounding risks. The synergistic effects of sedatives (e.g., benzodiazepines) and anticholinergics (e.g., tricyclic antidepressants) may impair both cognitive function and postural reflexes, amplifying fall risk beyond additive thresholds. While symptomatic management reduces falls by mitigating emesis, the dual risks of antiemetic side effects and psychotropic polypharmacy demand a tailored, multidisciplinary approach. Deprescribing non-essential CNS-acting medications and integrating rehabilitation interventions are critical to balancing efficacy and safety in oncology care.

This study found that there is also significant heterogeneity in the effect of different treatment modalities on falls. Patients receiving combined chemoradiotherapy with immunotherapy and targeted therapy have a higher risk of falls compared to those undergoing chemoradiotherapy alone. This increased risk is primarily attributed to the fact that chemoradiotherapy itself predisposes patients to fall-related complications, such as cancer-related fatigue, gastrointestinal disturbances, generalized weakness, peripheral neuropathy, and paresthesia. When combined with immunotherapy and targeted therapy, additive toxicities further exacerbate these risks. Immune checkpoint inhibitors (e.g., PD-1/PD-L1 inhibitors) may induce hypothyroidism (fatigue, muscle weakness) or adrenal insufficiency (hypotension, dizziness), directly impairing balance control. Chemotherapy-induced neuropathy synergizes with endocrine dysfunction caused by immunotherapy, significantly compromising postural stability. Therefore, particular attention should be paid to the need to assess the complications of different treatment modalities and take targeted precautions to prevent falls during the safe management of hospitalized cancer patients. As shown in Table A1, our model can achieve sufficient sensitivity for patients receiving more than 20 treatment modalities (others include 13 treatment modalities). Meanwhile, the accuracy of early warning in various treatment modes is maintained above 0.8.

Outcome analysis demonstrated significant diagnostic-specific fall associations, with glioma patients exhibiting higher fall risk compared to solid tumor counterparts. Mechanistically, cerebellar/brainstem-compressing lesions induced ataxia through dual pathways. Patients with osseous metastases demonstrated a 5-fold escalation in pathological fracture risk, with 67% experiencing fall recurrence within 6 months post-fracture. A Cancer Center pharmacovigilance study revealed bisphosphonate-treated cohorts exhibited 1.9 times increased fall risk to treatment-induced hypocalcemia [25, 26]. These findings mandate heightened clinical vigilance for two high-risk subgroups: head/neck malignancies, and cancer patients with cerebral metastases demonstrating.

This study confirms a dose-dependent relationship between the number of comorbidities and fall risk in hospitalized cancer patients undergoing chemoradiotherapy. An increasing number of comorbidities represents not merely a simple addition of diseases but a progressive decline in multisystem physiological functional reserves and an increase in overall frailty [27]. For instance, cardiovascular diseases (e.g., heart failure, arrhythmias) can lead to orthostatic hypotension or reduced cerebral perfusion; diabetes and its complications (e.g., peripheral neuropathy, retinopathy) impair sensory function and vision [28]. The cumulative impairment of these systemic functions creates a multifaceted and intertwined pathophysiological basis for falls. Multiple comorbidities amplify physiological vulnerability and interact synergistically with treatment toxicities—such as neuropathy, fatigue, and polypharmacy—to progressively impair stability and compensatory capacity. These findings underscore the need to integrate comorbidity assessment into fall prevention strategies, advocating for personalized, multidisciplinary risk mitigation in this high-risk population.

Clinical usability and impact

Our model exhibited high predictive accuracy (F1-score = 0.87), and decision curve analysis demonstrates significant clinical benefits across a broad range of risk thresholds. However, effective clinical implementation requires seamless integration into existing nursing workflows. To ensure adoption, the outputs of model should be embedded into Electronic Health Record (EHR) as real-time alerts, coupled with intuitive dashboards displaying SHAP explanations (Fig. 4). For example, nurses could prioritize high-risk patients (e.g., those on neurotoxic chemotherapy) for hourly rounding, reducing fall rates by an estimated 30%.

Economically, the model’s precision (92% accuracy) could reduce unnecessary interventions for low-risk patients, freeing ~ 15% of nursing hours for critical tasks. To maximize clinical utility, the proposed model should be embedded within hospital EHR systems to enable real-time fall risk alerts. These alerts could dynamically interface with care resource prioritization modules, automatically assigning higher nursing surveillance levels to high-risk patients. For instance, a “fall risk dashboard” within the EHR could highlight patients requiring immediate attention, while simultaneously triggering automated reminders for nursing staff. Such integration would streamline workflows, reduce response latency, and ensure timely interventions. For instance, a patient flagged as high-risk by the model could automatically trigger a nursing protocol mandating hourly rounding and mandatory use of non-slip footwear.

Limitations

Although we have constructed a precise prediction model for cancer patients undergoing chemoradiotherapy. However, the study has the following limitations: (1) The generalizability of our model may be constrained by its single-center origin and relatively limited sample size (n = 298). External validation through multicenter studies with larger, heterogeneous cohorts is essential to confirm its applicability across diverse healthcare settings. (2) Critical contextual variables influencing fall risk, such as slippery floors, absence of bedside handrails, and suboptimal lighting conditions, were not systematically integrated into the model. This omission may reduce the ecological validity of predictions. (3) Prospective studies are needed to evaluate the clinical impact and usability of the model, particularly regarding integration with EHR systems and staff adherence to AI-generated risk alerts. While SHAP enhances interpretability, clinicians may still require training to effectively utilize these explanations in decision-making.

Future works

While our model demonstrates high predictive accuracy, its clinical utility hinges on seamless integration into care workflows. Future implementations should prioritize user-friendly interfaces (e.g., EHR-embedded alerts) and staff training to ensure adoption. Economically, the model’s precision could reduce unnecessary bed alarms for low-risk patients, optimizing nurse workloads. Prospective trials are needed to quantify these benefits and address barriers such as data privacy concerns and model maintenance costs.

Conclusion

This study developed a clinically interpretable ML model for predicting fall risk in cancer patients, combining high accuracy with actionable insights for personalized prevention. Through multidisciplinary collaboration and technological innovation, such models are expected to become an important tool to decrease the risk of fall, reduce related healthcare costs and improve patient prognosis.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

LASSO

Least Absolute Shrinkage and Selection Operator

SHAP

Shapley Additive Explanations

ROC

Receiver Operating Characteristic

PR

Precision-Recall

RF

Random Forest

AUROC

Area under the ROC curve

AUPRC

Area under the PR curve

NB

Naive Bayes

MLR

Multivariate Logistic Regression

SVM

Support Vector Machine

HER

Electronic Health Record

Author contributions

H. Z. (First Author): Writing - Original Draft, Writing - Review & Editing. Y. A. (co-First Author): Methodology, Software, Validation, Writing - Original Draft, Visualization; M. S.: Data Curation; Y. M. (Corresponding author): Funding acquisition, Supervision. All authors have read and approved the manuscript and consent to its submission.

Data availability

The dataset analyzed in this study is not publicly available as it has not yet been accepted for formal publication, but can be obtained upon request from the corresponding author after formal publication. Relevant codes can be requested from the corresponding author (mengyt20210816@163.com).

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the principles of the Declaration of Helsinki. The Ethics Committee of the Affiliated Cancer Hospital of Shandong First Medical University approved this retrospective study (Approval No. 2024001039). Given the retrospective nature of the study, the requirement for informed consent was waived by The Ethics Committee of the Affiliated Cancer Hospital of Shandong First Medical University based on China’s Ethical Review Measures for Biomedical Research Involving Humans (2016).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.