Wearable diagnosis stops treating every sensor stream alike: a gated capsule-Transformer lifts accuracy to 0.9348
Synopsis
For clinical diagnosis from multi-device wearable sensors, a cross-platform framework chains stream selection, class-imbalance correction and spatiotemporal feature learning into one pipeline. Previously, sensor modalities, sampling rates and transmission properties were handled separately, rare abnormal samples were scarce, and inference shifted with the device. On two public datasets the framework reaches AUC 0.966 and 0.947; at 24 streams it reports accuracy 0.9348, precision 0.9305, sensitivity 0.9615 and false-positive rate 0.0915.
Interpretation
Wearable sensor data is filtered and rebalanced before it reaches the diagnostic model, rather than sending every physiological stream into the network with equal weight. Earlier practice treated all incoming wearable streams alike and gave them the same computational resources, which invited misjudgment under the class imbalance caused by rare physiological abnormalities. Ablations ran on two public datasets: a cardiovascular set of 48K samples from ECG, PPG, SpO2 and motion sensors at 250 Hz, and an IoMT set of 62.5K samples from glucose, EEG, respiratory and temperature sensors at 50 to 500 Hz.
Three parts carry the framework: an optimizer that ranks streams by physiological importance, a module that generates minority-class samples, and a gated network fusing capsule layers with a Transformer. Selective meta-heuristic optimization (SMHO) orders streams with a multi-dimensional utility vector combining anomaly contribution, signal relevance, transmission reliability and temporal urgency; integrated generative-synthetic data balancing (IG-SDB) fills minority classes through context-aware synthetic reconstruction. The gated capsule-Transformer network (GCT-Net) extracts spatial-hierarchical features with capsule layers, models long-range temporal dependencies with customized Transformer blocks, and gates out irrelevant activations; the setup uses 15 layers, learning rate 0.5, 1200 iterations and 30 epochs.
At 24 sensor streams the framework beats four comparison methods on accuracy, precision, sensitivity and false-positive rate at once. The comparison methods LWEN, DDIH, GAN-I and MMFL report accuracies of 0.844, 0.863, 0.888 and 0.889, with false-positive rates up to 0.184. Under the same setting the framework reports accuracy 0.9348, precision 0.9305, sensitivity 0.9615, false-positive rate 0.0915 and uncertainty factor 0.0502; the stated gains are 12.76% accuracy, 12.95% precision and 11.3% sensitivity, with 11.35% lower false positives and 12.51% lower uncertainty.
Perspective
Evaluation is confined to two public datasets: the cardiovascular set separates normal from abnormal, and the IoMT set separates good from variable status, at fixed sampling of 250 Hz and 50 to 500 Hz respectively. The framework uses 15 layers, learning rate 0.5, 1200 iterations and 30 epochs, with comparison at 24 sensor streams. The authors propose extending it to federated, privacy-preserving distributed training and to digital-twin predictive modeling with explainable AI.
The reported gains are relative to four comparison methods at a single stream setting and have not been validated in an independent cohort or a real clinical deployment. The stream-selection sensitivity analysis notes that removing too many streams can suppress clinically relevant difference signals, while keeping too many raises inference time, so the two must be traded off. The interoperability conclusion rests on distribution alignment across two datasets and does not cover devices from more vendors.