Paperdoi:10.3390/s26227197

Wearable Activity Recognition for Swimming Education: Feature Engineering, Sensor Placement, and Machine Learning Evaluation

Venue listedDated Nov 14, 2026Added Nov 17, 2026

  • LH

Authors: Larry C. Powell; Tracy Anne Hammond (Texas A&M Computer Science & Engineering); Selina Stasi (Texas A&M School of Public Health)

Sensors 26(22):7197 (2026), MDPI

TL;DR

Wrist and lower-back wearables recorded 45 children during ordinary swimming lessons. Instead of a 21-class problem that the data could not support, the authors ran 12 binary tasks under strict participant-held-out nested cross-validation. Results ranged from macro-F1 0.42 to 0.84 with a decision tree; the best task (backstroke vs jump entry and recovery, 3-second wrist windows) reached 0.874 with a random forest.

Abstract

Wearable activity recognition in swimming education differs from competitive-stroke analysis because children’s lessons include brief, transitional, and orientation-dependent skills. We evaluated engineered geometric, statistical, and spectral descriptors extracted from wrist and lower-back wearable recordings during natural lessons. The dataset comprised 45 children, with differing sensor availability; 21 activity labels were considered in historical exploratory analyses. A provenance audit identified variable historical F1 conventions and incomplete participant coverage for several activities. Twelve screened binary tasks were therefore evaluated using outer participant-held-out nested leave-one-subject-out (LOSO) validation with inner-fold Decision Tree feature selection. Across 201 outer-fold instances, the Decision Tree achieved pooled macro-F1 scores ranging from 0.422 to 0.839; the strongest result distinguished Backstroke from Jump Entry and Recovery using three-second wrist windows (40 participants, 294 windows). Four additional classifiers were evaluated on the same saved fold-specific selected features and splits: Random Forest reached macro-F1 0.874 on that task. Signal-processing features were selected in all 201 outer folds, including features labeled Welch PSD in 197 folds (98.0%). Coverage auditing documented why the full 21-class task did not meet participant-support criteria; no fully nested reduced multiclass experiment was performed. A small retrospective agreement check of separately labeled student video clips yielded 18/18 agreements (κ=1.00) within one recording and is not representative of study-wide reliability. Sampling-rate provenance and historical fixed-range scaling parameters remain limitations. The results support selected context-restricted activity discrimination, not comprehensive 21-class recognition or universal classifier superiority.

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 leaderboards

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

Tasks & signals

Tasks

Activity recognition

Signals

Keywords

HarSwimmingChildrenSensor PlacementWrist IMULower Back IMUFeature EngineeringWelch Psd

How it works

How Method works: Setup → Method → TargetSetup: Signals (accel, imu). Method: Method. Target: Activity recognition.SetupSignals: accel, imuxyzSignalsaccel, imuMethodMethodTargetActivity recognitionActivity recognition

Cite this paper

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

Powell, L. C., Hammond, T. A., & Stasi, S. (2026). Wearable Activity Recognition for Swimming Education: Feature Engineering, Sensor Placement, and Machine Learning Evaluation. Sensors , MDPI, 26(22), 7197. https://doi.org/10.3390/s26227197
Powell, Larry C., Tracy Anne Hammond, and Selina Stasi. 2026. "Wearable Activity Recognition for Swimming Education: Feature Engineering, Sensor Placement, and Machine Learning Evaluation." Sensors , MDPI 26 (22): 7197. https://doi.org/10.3390/s26227197.
Powell, L. C., Hammond, T. A. and Stasi, S. (2026) 'Wearable Activity Recognition for Swimming Education: Feature Engineering, Sensor Placement, and Machine Learning Evaluation', Sensors , MDPI, 26(22), p. 7197. doi: 10.3390/s26227197.
Powell, Larry C., et al. "Wearable Activity Recognition for Swimming Education: Feature Engineering, Sensor Placement, and Machine Learning Evaluation." Sensors , MDPI, vol. 26, no. 22, 2026, p. 7197. https://doi.org/10.3390/s26227197.
L. C. Powell, T. A. Hammond, and S. Stasi, "Wearable Activity Recognition for Swimming Education: Feature Engineering, Sensor Placement, and Machine Learning Evaluation," Sensors , MDPI, vol. 26, no. 22, p. 7197, 2026, doi: 10.3390/s26227197.
@article{powell2026wearable,
  title         = {{Wearable Activity Recognition for Swimming Education: Feature Engineering, Sensor Placement, and Machine Learning Evaluation}},
  author        = {Powell, Larry C. and Hammond, Tracy Anne and Stasi, Selina},
  year          = {2026},
  journal       = {Sensors , MDPI},
  volume        = {26},
  number        = {22},
  pages         = {7197},
  doi           = {10.3390/s26227197}
}

Added to the ChatPPG research hub