Paperdoi:10.1109/bmeicon70815.2026.11773001

Self-Supervised Fall Detection from IMU Signals Using Masked Reconstruction and Forecasting

Venue listed2026Added Nov 15, 2026

  • PH
  • PS
  • CL
  • CL

Authors: Ping-Hung Hsieh, Pi-Shan Sung, Chun-Yi Liu, Chih-Lung Lin

IEEE BMEiCON 2026 (2026-10-01)

TL;DR

Learns normal ADL motion with masked reconstruction + forecasting and flags falls by prediction-error anomaly score, so no labelled falls are needed for training. KFall: accuracy 98.4%, sensitivity 98.9%, false-alarm rate 1.0%; cross-dataset: accuracy 97.5%, sensitivity 94.2%, FAR 0.0% (second dataset not named in the abstract).

  • Form factor: waist/trunk (KFall)

Abstract

Falls can cause severe physical injuries, psychological trauma, and even death. To mitigate these risks, many fall detection methods have been proposed to provide timely rescues. However, most existing approaches rely on supervised learning and require large amounts of labeled fall data. This work presents a self-supervised fall prediction approach using IMU signals. The proposed model learns normal motion patterns from activities of daily living (ADLs) and identifies fall-related motions using an anomaly score derived from prediction errors. Experimental results on the KFall dataset demonstrate that the proposed method achieves an accuracy of 98.4%, a sensitivity of 98.9%, and a false alarm rate of 1.0%. In cross-dataset evaluation, the proposed approach achieves an accuracy of 97.5%, a sensitivity of 94.2%, and a false alarm rate of 0.0%. These results demonstrate the effectiveness and generalization capability of the proposed self-supervised approach for wearable fall detection.

Verbatim from Semantic Scholar (opens in a new tab) · Abstract via Semantic Scholar; © the authors/publisher

Results on ChatPPG leaderboards

All leaderboards

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

Tasks & signals

Tasks

Signal denoising

Signals

Keywords

Fall DetectionSelf SupervisedAnomaly DetectionMasked ReconstructionKFall

How it works

How Method works: Setup → Method → TargetSetup: Signals (IMU); Form factor (waist/trunk (KFall)). Method: Method (Masked reconstruction). Target: Signal denoising.SetupSignals: IMUxyzSignalsIMUForm factor: waist/trunk (KFall)Form factorwaist/trunk (KFall)MethodMASK → RECONSTRUCTMethodMasked reconstructionTargetSignal denoisingSignal denoising

Cite this paper

Cite this paper with the fields recorded here. Check it against the publisher before submitting.

Hsieh, P.-H., Sung, P.-S., Liu, C.-Y., & Lin, C.-L. (2026). Self-Supervised Fall Detection from IMU Signals Using Masked Reconstruction and Forecasting. IEEE BMEiCON 2026. https://doi.org/10.1109/bmeicon70815.2026.11773001
Hsieh, Ping-Hung, Pi-Shan Sung, Chun-Yi Liu, and Chih-Lung Lin. 2026. "Self-Supervised Fall Detection from IMU Signals Using Masked Reconstruction and Forecasting." IEEE BMEiCON 2026. https://doi.org/10.1109/bmeicon70815.2026.11773001.
Hsieh, P.-H., Sung, P.-S., Liu, C.-Y. and Lin, C.-L. (2026) 'Self-Supervised Fall Detection from IMU Signals Using Masked Reconstruction and Forecasting', IEEE BMEiCON 2026. doi: 10.1109/bmeicon70815.2026.11773001.
Hsieh, Ping-Hung, et al. "Self-Supervised Fall Detection from IMU Signals Using Masked Reconstruction and Forecasting." IEEE BMEiCON 2026, 2026. https://doi.org/10.1109/bmeicon70815.2026.11773001.
P.-H. Hsieh, P.-S. Sung, C.-Y. Liu, and C.-L. Lin, "Self-Supervised Fall Detection from IMU Signals Using Masked Reconstruction and Forecasting," IEEE BMEiCON 2026, 2026, doi: 10.1109/bmeicon70815.2026.11773001.
@article{hsieh2026selfsupervised,
  title         = {{Self-Supervised Fall Detection from IMU Signals Using Masked Reconstruction and Forecasting}},
  author        = {Hsieh, Ping-Hung and Sung, Pi-Shan and Liu, Chun-Yi and Lin, Chih-Lung},
  year          = {2026},
  journal       = {IEEE BMEiCON 2026},
  doi           = {10.1109/bmeicon70815.2026.11773001}
}

Added to the ChatPPG research hub