Multimodal EMG–IMU sensor fusion with dual-output LSTM for fatigue estimation during neonatal chest compressions

Abstract

Background

During neonatal cardiopulmonary resuscitation (NCPR), rescuer fatigue develops rapidly and compromises compression quality. Conventional feedback systems infer fatigue indirectly from mechanics (depth/rate) and may miss early neuromuscular changes.

Objective

To develop and evaluate a multimodal framework that fuses surface EMG (physiology) and IMU (biomechanics) to improve the accuracy of (i) fatigue level classification and (ii) prediction of fatigue-onset time during neonatal chest compressions (NCPR).

Methods

Twenty trained providers performed simulated neonatal compressions on a manikin while synchronized EMG (deltoid, triceps, upper trapezius) and 3-axis IMU signals were recorded and windowed (2 s, 50% overlap). Features included EMG RMS, MAV, median frequency (MF), and IMU depth dynamics. A dual-output Long Short-Term Memory (LSTM) jointly produced 3-class fatigue labels and onset-time regression.

Results

Fusion outperformed unimodal models: 98.3% accuracy, macro-F1 0.982, AUC 0.99; onset prediction RMSE 38.3 s, R² 0.68. EMG-only: 69.4% accuracy; IMU-only: 96.7%. EMG provided early physiological fatigue signatures, complementing IMU mechanical degradation.

Conclusion

EMG–IMU fusion with temporal deep learning improves fatigue estimation during NCPR and is suitable for real-time feedback to support optimal rescuer rotation. Earlier, physiology-aware fatigue detection enables proactive team management before compression quality declines. Lightweight LSTM fusion runs in real time and generalizes across rescuers.