Paperdoi:10.21203/rs.3.rs-10746145/v1
Sensor Modality and Placement Optimization for Wearable Fall Detection: A Systematic Evaluation of Accelerometer, Gyroscope, and Magnetometer Contributions Using the UMAFall Dataset
Research Square preprint (2026-08-20)
Alex Kubiak et al., "Sensor Modality and Placement Optimization for Wearable Fall Detection: A Systematic Evaluation of Accelerometer, Gyroscope, and Magnetometer Contributions Using the UMAFall Dataset", doi:10.21203/rs.3.rs-10746145/v1, CC BY 4.0
TL;DR
Secondary analysis of UMAFall: 4 sensor configs (acc; acc+gyro; acc+mag; all) × 5 positions (chest, waist, wrist, ankle, all) × 4 classifiers (SVM, RF, kNN, NB). 74/80 combos met Se and Sp ≥ 0.90; all 6 failures were kNN. Best single position: waist acc+gyro SVM, Se 0.979 / Sp 0.995.
- Form factor: chest / waist / wrist / ankle
Abstract
Falls are a leading cause of injury-related morbidity and mortality among older adults, and wearable inertial measurement unit-based fall detection systems offer a promising approach to continuous ambulatory monitoring. Despite theoretical justification for multi-modal sensing, the independent and combined contributions of gyroscopes and magnetometers to fall detection accuracy remain poorly characterized across body positions and classifier types. We present a comprehensive secondary analysis of the publicly available UMAFall dataset, evaluating four sensor modality configurations—accelerometer only, accelerometer plus gyroscope, accelerometer plus magnetometer, and all three sensors combined—across five body positions (chest, waist, wrist, ankle, and all positions combined) using four supervised machine learning classifiers: support vector machine, random forest, k-nearest neighbor, and naïve Bayes. Configurations were judged against a pre-specified dual acceptance rule requiring sensitivity and specificity to be at least 0.90 simultaneously; overall balanced performance is summarized by the composite SpSe, defined as the arithmetic mean of sensitivity and specificity. Among 80 unique configuration–position–classifier combinations, 74 (92.5%) met the dual acceptance rule, with all six failures involving the k-nearest neighbor classifier. The best single-position result was obtained with an accelerometer–gyroscope configuration at the waist using a support vector machine (sensitivity = 0.979, specificity = 0.995, composite = 0.987). Gyroscope addition improved the composite in 17 of 20 position–classifier comparisons relative to accelerometry alone. Magnetometer inclusion produced consistent degradation with the k-nearest neighbor classifier at all five positions (composite range: 0.538–0.707) while providing marginal gains with support vector machine and random forest. Fourteen of 15 tested sensor placement combinations—including a single chest-mounted configuration—met the dual acceptance rule. Frequency-domain feature augmentation provided position-specific benefits only at the wrist. These results offer empirical guidance for sensor architecture decisions in wearable fall detection system design.
Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by/4.0/
Results on ChatPPG leaderboards
All leaderboardsNo results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.
Tasks & signals
Signals
Keywords
Datasets used
How it works
Cite this paper
Sensor Modality and Placement Optimization for Wearable Fall Detection: A Systematic Evaluation of Accelerometer, Gyroscope, and Magnetometer Contributions Using the UMAFall Dataset. (2026). Research Square preprint. https://doi.org/10.21203/rs.3.rs-10746145/v1
@misc{kubiakumafallsen2026sensor,
title = {{Sensor Modality and Placement Optimization for Wearable Fall Detection: A Systematic Evaluation of Accelerometer, Gyroscope, and Magnetometer Contributions Using the UMAFall Dataset}},
year = {2026},
howpublished = {Research Square preprint},
doi = {10.21203/rs.3.rs-10746145/v1}
}Added to the ChatPPG research hub