Personalizing fall fear prevention in knee osteoarthritis: an interpretable prediction framework via IGKSO synchronous optimization and SHAP-based risk stratification.

OBJECTIVE

To construct a concern about falling (CAF) prediction model for patients with knee osteoarthritis (KOA) based on synchronous optimization.

METHODS

A total of 541 patients with KOA admitted to two hospital from September 2021 to September 2023 were selected. CAF was evaluated using the Falls Efficacy Scale-International (FES-I). Patients were divided into a CAF group (n = 360, FES-I ≥ 28 points) and a no CAF group (n = 181, FES-I < 28 points). 80% of the data (433 cases) were used as the training set, and the remaining 20% (108 cases) were used as the test set. An improved swarm intelligence algorithm was used for feature selection and hyperparameter optimization. The selected variables were further analyzed by Shapley Additive exPlanation (SHAP) interpretable method.

RESULTS

In the training set, the maximum F1 score of the improved synchronous optimization machine learning model was 0.8842, and the area under the curve reached 0.9451. In the test set, the maximum F1 score of the improved synchronous optimization machine learning model was 0.8589, and the area under the curve reached 0.9315. Eight variables were selected based on the improved synchronous optimization machine learning model, including Timed Up-and-Go (TUG) time, Western Ontario and McMaster Universities Osteoarthritis (WOMAC) pain score, Hospital Anxiety and Depression Scale (HADS) anxiety score, knee extensor moment, age, sex, Kellgren-Lawrence (KL) grade, and Body mass index (BMI). Critically, SHAP analysis demonstrated triadic interactive effects among key risk indicators, revealing that older adult female patients with concurrent TUG time >14 s, HADS-anxiety scores >10, and high WOMAC pain scores constituted the peak-risk cohort amplified through bio-psycho-social interactions.

CONCLUSION

This study validated a multimodal predictor model for CAF in knee osteoarthritis (KOA) patients through a machine learning framework, revealing synergistic mechanisms among biomechanical, psychological, and social dynamics, with specific risk stratification for high-risk populations such as older adult females to guide clinical practice.