A PHYSIOLOGICALLY-WEIGHTED ENSEMBLE LEARNING FRAMEWORK FOR CONTINUUM-BASED FALL RISK STRATIFICATION IN OLDER ADULTS USING WEARABLE IMU AND IOMT SMART-CANE SENSOR STREAMS

Abstract

Falls remain among the most consequential events in geriatric health, contributing to fracture, hospitalization, loss of independence, and downstream psychological harm such as fear of falling. Most deployed fall-safety technologies still operate reactively, flagging an event only once it has already occurred. This paper develops and empirically evaluates a continuum-based fall risk stratification framework that fuses two complementary, clearly-disclosed synthetic sensor streams: a wearable inertial measurement unit (IMU) stream used for binary fall/no-fall discrimination, and a multimodal Internet-of-Medical-Things (IoMT) smart-cane (“cStick”) stream that adds grip pressure, proximity, heart-rate variability (HRV), blood-oxygen saturation (SpO₂), and glucose readings to support a three-state classification of No Fall, Fall Predicted, and Definite Fall. Seven classical machine learning algorithms — logistic regression, Gaussian Naive Bayes, k-nearest neighbours (KNN), support vector machines, decision trees, random forest, and gradient boosting — are trained and compared under an identical preprocessing and evaluation protocol. A distance-weighted KNN classifier (k = 9) is found empirically to give the best held-out performance on the IMU stream, reaching 91.11% test accuracy, 91.67% recall, 90.56% specificity, and a Matthews correlation coefficient of 0.822, while a gradient boosting classifier reaches 84.38% accuracy and a macro-F1 of 84.3% on the harder three-class cStick problem. Feature-importance analysis identifies post-event gyroscopic dynamics, root-mean-square acceleration, and peak linear acceleration as the dominant IMU predictors, and proximity distance, wrist/cane acceleration, and grip pressure as the dominant cStick predictors. A physiologically-weighted adaptive thresholding mechanism is also introduced and tested: subject-level baseline risk, modelled through a logistic combination of vulnerability covariates, shifts the operating decision threshold so that higher-risk individuals are evaluated at a more sensitive cut-off. This adaptive scheme recovers a markedly better recall than a single fixed threshold (98.33% recall in the high-risk stratum versus 92.62% under a fixed threshold) at a deliberate, quantified cost in false-alarm rate. All numerical results in this paper are produced by an executed Python pipeline (scikit-learn) on transparently simulated data, and every reported figure is generated directly from that pipeline’s output rather than asserted. The paper situates these findings against 25 peer-reviewed studies published between 2023 and 2026, and discusses the framework’s relevance to smart assistive canes, wearable elder-care monitoring, and edge-deployable IoMT healthcare.