CTX-FALLNET: A HYBRID CNN TRANSFORMER–XGBOOST FRAMEWORK FOR ELDERLY DAILY ACTIVITY CLASSIFICATION AND FALL DETECTION

Abstract

Wearable inertial sensors enable continuous monitoring of daily activities and falls, but separate deep models for human activity recognition (HAR) and fall detection duplicate feature learning and underutilize shared motion representations. This paper presents CTX-FallNet, a shared-representation framework coupling both tasks without end-to-end multi-task learning. A CNN–Transformer backbone is trained on activities-of-daily-living (ADL) windows and frozen to generate a 128-dimensional embedding for independently trained XGBoost classifiers performing multi-class HAR and binary fall detection. The contribution lies not in the individual components but in the shared-representation design and subject independent evaluation. On SisFall and KFall, subject-independent StratifiedGroupKFold cross-validation prevents participant overlap between training and evaluation. Fall detection achieved accuracies of 99.63 ± 0.21% and 99.13 ± 0.23%, ROC-AUC values of 0.998 ± 0.001 and 0.9995 ± 0.0003, sensitivities of 95.41 ± 1.75% and 98.60 ± 0.81%, and specificities of 99.62 ± 0.23% and 99.32 ± 0.38% on SisFall and KFall, respectively, with low Expected Calibration Error and Brier scores. HAR achieved accuracies of 84.92 ± 0.63% and 83.82 ± 3.32% and macro-F1 scores of 83.65 ± 0.69% and 83.61 ± 2.90% across 19 and 21 classes, respectively. Linear probing, PCA, and t-SNE support embedding discriminability. Cross-dataset, cross-device, and real-world generalization remain to be established.