TAI-based Smart Wearable System for Continuous Patient Health Monitoring

Abstract

Nowadays, it is a normal practice to monitor patients following discharge not a novelty in research. But most wearables still measure and summarize a wealth of physiology and report it as an average over time, and much is lost in the measurement and reduction. This paper introduces a new wearable we call TAI, which preserves the temporal form of the signal and infers on the signal within the patient's body, rather than in a data centre. The channels recorded concurrently are photoplethysmography, 3-lead electrocardiography, skin temperature, inertial motion and electrodermal activity. We break them down into 10 second windows and input those windows into a small transformer encoder, four heads, and three stacked blocks. The trained encoder was then compressed using post-training integer quantisation to fit within the power budget of a microcontroller with a Cortex-M55 architecture. On a held-out subject-wise test partition of 41,860 windows, the model achieved an accuracy of 96.1% and a macro F1 score of 95.4%. These values are 3.8 and 4.2 points higher than those of a CNN-LSTM baseline tuned under the same protocol. The quantised build answers in 23ms and uses 0.51mJ an inference. Each alert also includes its attention map, to give a clinician a visual representation of which portion of the waveform contributed to the increase in risk rather than relying on a number. In summary, attention-based inference does not require server-class hardware, and always-on monitoring at the edge is achievable with present-day components.