An Explainable Risk-Calibrated Stacked-Ensemble Framework for Anticipatory Multimodal Fall Analytics in the Elderly Using Wearable IMU and IoMT Assistive-Stick Sensing
Falls in geriatric patients continue to pose a significant source of harm, hospitalization, worsening of function, and impaired autonomy, prompting the development of smart technologies that no longer rely on reactive fall detection methods to move towards proactive predictions and prevention. A new application-oriented multimodal fall analytics framework that utilizes two supplementary sets of data, namely, IMU-based wearable data to detect binary falls and smart cStick IoMT data to distinguish between no fall, fall predicted, and definite fall is introduced in the paper. The suggested system incorporates motion-derived statistical measures like the maximum acceleration, the maximum gyroscopes, kurtosis, skew and after-impact with contextual and physiological data, including grip pressure, variations in heart rates, blood oxygen saturation levels, sugar level, and distance of proximity. This combination allows an event-based detection of risks and in-the-field anticipatory risk analysis in real time. Rather than selecting a single classifier, the framework introduces a heterogeneous stacked-generalization ensemble that combines Random Forest, Extremely Randomized Trees, and gradient boosting as base learners with a regularized logistic-regression meta-learner trained on out-of-fold predictions. The ensemble outputs are probability-calibrated (isotonic regression) so that a subject-specific, risk-aware adaptive decision threshold operates on trustworthy probabilities, and model behaviour is made transparent through SHapley Additive exPlanations (SHAP) for both global and local interpretability. A salient preprocessing, feature-engineering, and evaluation pipeline supports the comparison against conventional baselines including Logistic Regression, Naive Bayes, K-Nearest Neighbors, Support Vector Machine, and Decision Tree. Experimental analysis of IMU data shows that discrimination is very reliable and the proposed calibrated stacked ensemble attains a high value of 97.472% test accuracy, matching or marginally exceeding its strongest base learner, with good macro-F1 performance and good generalization. Multimodal characteristics are highly applicable in the practice on the cStick dataset since pressure, distance and HRV turned out to be the central predictors of instability and fall progression. The main innovation of the work is that it is risk-conscientious, multimodal and anticipatory constructed in such a manner that it can be applied in real-time to smart assistive devices, elderly care monitoring, home safety system, and remote healthcare applications. The developed framework provides a scalable base on the next-generation AI-empowered elderly fall prevention and intervention technologies