Fall Risk Classification in Community-Dwelling Older Adults: A Cross-Sectional Machine Learning Study

Falls are a leading cause of injury and functional decline among older adults, yet existing screening tools often struggle to capture the full complexity of fall risk or reliably differentiate those at higher risk.. Through this thesis, an interpretable machine learning model was developed to classify fall risk using multimodal data collected during Denmark’s preventive home visit (PHV) program. A retrospective cross-sectional dataset of 382 community-dwelling adults aged 75+ from the HANC study was analyzed. Input features included physical performance (SPPB, grip strength), physical activity from wrist-worn accelerometry (mean counts per minute), health-related quality of life (SF-12, EQ-5D), cognition (DSST, MMSE), fatigue (Avlund Mobility-Tiredness, Pittsburgh Fatiguability Scale), activities of daily living (ADLs), and fear of falling (FES-I). The outcome was self-reported falls in the past year.

An XGBoost classifier was trained and evaluated using five-fold cross-validation with hyperparameters selected using Bayesian optimization. The model achieved moderate = 0.193). Using the Youden-selected threshold (≥ 0.29 to classify fallers), sensitivity was 0.64, specificity 0.60, precision 0.40, F1 score 0.49, and balanced accuracy 0.62. Overall, these results are comparable to those reported in studies using multimodal datasets of similar sample sizes and feature inputs. SHAP analysis highlighted physical activity, chair rise time, fear of falling, grip strength, and health-related quality of life as key predictors of fall risk.

The model’s design aligns with Denmark’s PHV framework and emphasizes transparent reporting of evaluation metrics, including calibration and threshold selection, addressing key gaps in practical decision-making and reproducibility within the fall-risk classification literature. Future research should focus on expanded datasets, temporal validation, and investigating three-tiered classification schemes, such as by including a recurrent faller group to enhance clinical utility. With further development, this approach could support scalable, data-driven fall-risk stratification within municipal public health systems, improving early identification and targeted preventive interventions for older adults.