Sensor failure in wearable stress detection: evaluating missing-modality robustness

This is the authors' abstract. We don't add key points for this paper.

Wearable stress detection systems are typically trained and evaluated under the assumption that every physiological sensor channel is available at inference time. In real deployments, however, sensors fail through electrode detachment, motion artifacts, battery exhaustion, or signal saturation. This paper presents a systematic benchmark of missing-modality robustness on the WESAD dataset. A handcrafted-feature pipeline (SVM, Random Forest, XGBoost on 11 features) and three deep multimodal models (a modality-wise CNN, a CNN-LSTM and a Transformer) are trained on four chest-worn modalities (ECG, EDA, respiration, skin temperature) and evaluated with Leave-One-Subject-Out cross-validation over 15 subjects. The robustness of the CNN-LSTM, the best full-modality model (Macro-F1 = 0.879 ± 0.105), is quantified with a Robustness Score (RS), the ratio of Macro-F1 under a failure condition to Macro-F1 with all sensors, over 4 single-sensor and 10 multi-sensor dropout scenarios without retraining. ECG is the most critical modality: its removal reduces Macro-F1 by 10.51% (RS = 0.895, paired t -test p = 0.0136), whereas removing respiration costs 1.07% (RS = 0.989, not significant). Corrupted signals are not equivalent to missing signals: heavy noise (0 dB SNR) on ECG lowers RS to 0.803, below complete ECG removal, while the same corruption of any other modality has little effect. Training with modality dropout gives a small, non-significant increase in Macro-F1 (0.879 to 0.897, p = 0.27) and does not significantly change sensitivity to ECG loss, and a Transformer baseline (0.862) does not outperform the CNN-LSTM. These findings provide an empirical basis for prioritizing ECG signal quality and sensor redundancy in wearable stress-monitoring hardware.