Stagewise Optimization Framework for Fall Direction Recognition From Wearable Sensor Data Based on Machine Learning

Accurate fall direction recognition is essential for developing effective fall prevention and intervention systems, yet it remains challenging due to the subtle differences in motion patterns across fall types. This research proposes a Stagewise Optimization Framework for Fall Type Recognition (SOFFDR), which systematically enhances classification performance through four sequential stages: (1) classifier selection via K-fold crossvalidation over ten candidate algorithms, (2) superior filtering method determination, (3)optimal window time tracking for segment-based feature extraction with Shapley Additive Explanations (SHAP) analysis, and (4) final classification using the best parameter combination from all previous stages. The framework was evaluated on wearable inertial measurement unit (IMU) data and compared against a traditional feature vector approach in which each recording is treated as a single instance. This feature vector method achieved an accuracy of 71% (macro F1-score = 0.72), with significant misclassifications between similar fall types. In contrast, the proposed SOFFDR system achieved 100% accuracy and perfect precision, recall, and F1-scores across all fall categories. These results highlight the critical role of systematic stagewise optimization, temporal segmentation, and filtering in enhancing fall type recognition performance from wearable sensor data. The proposed framework demonstrates its potential for high-precision fall monitoring applications in healthcare and assisted living environments.