@inproceedings{xueswimcoachaimultimodalwearableswimming2026,
  title = {{SwimCoachAI: A Multimodal Wearable System for Swimming Rhythm Sensing and Coaching}},
  author = {Xue, Xiaochan and Cai, Haofan and Pan, Yanjun and Wu, Shuqiong and Zheng, Yao},
  booktitle = {{2026 IEEE Annual Congress on Artificial Intelligence of Things (AIoT)}},
  year = {2026},
  month = dec,
  publisher = {IEEE},
  note = {Accepted for publication; DOI pending},
  abstract = {Swimming performance depends on coordinated breathing, arm strokes, kicking, and physiological response. Existing wearable systems primarily report isolated metrics such as stroke count or heart rate, while camera-based biomechanical analysis requires dedicated infrastructure and is typically performed offline. This paper presents the design of SwimCoachAI, a three-point wearable sensing and coaching system that integrates smart goggles, a wrist-worn hub, and an ankle sensor to capture breathing, arm-stroke, and kick events together with heart rate and SpO2. The nodes communicate over a 915 MHz LoRa link using event-driven transmission, in which each node detects and timestamps semantic events locally, decoupling event time from packet arrival time. A lightweight classifier at the wrist hub is designed to recognize representative breathing–stroke–kick rhythms and to drive goggle LED and ankle-vibration cues toward a configurable target pattern. As an initial feasibility study, synthetic analyses show that event-driven transmission reduces nominal communication load by more than an order of magnitude relative to raw IMU streaming. Under the stated synthetic packet-loss model, buffering locally timestamped events improves four-class rhythm classification from a macro-F1 of 0.757 to 0.996 at an 80% initial packet delivery ratio. A retrospective Garmin case study of one 800-unit freestyle workout further examines wrist-derived heart rate, stroke rate, and SWOLF across manually annotated 2-beat, 4-beat, and 4–6-beat kick regimes, illustrating both measurable within-session variation and the inability of a wrist-only record to directly resolve breathing–kick coordination. Prototype implementation, in-water communication measurements, and controlled swimmer evaluation remain planned validation.}
}