A GRU-based Model for Early Stroke Prediction Using Surface Electromyography Signals

Abstract

Stroke is a leading cause of mortality and long-term disability globally, necessitating early detection for timely intervention. This paper presents a comprehensive study on the development, training, and evaluation of a Gated Recurrent Unit (GRU) model for early stroke prediction using Surface Electromyography (sEMG) signals. The proposed model leverages the temporal modeling capabilities of recurrent neural networks to analyze sequential EMG data from 8 channels, intended to capture neuromuscular signal characteristics associated with stroke-induced motor impairment. The model architecture includes a GRU layer with 64 hidden units, followed by a fully connected layer with ReLU activation and a final softmax output layer for binary classification. We implement and compare two variants: a GRU-only model and a CNN-GRU hybrid. Training incorporates gradient clipping by norm and learning rate schedulers (step decay and exponential decay) to address vanishing/exploding gradient problems and optimize convergence. The model is trained and evaluated using the MUSED-I sEMG dataset, comprising 11 healthy subjects and 2 stroke patients, with a total of 11,146 windowed samples. Group K-Fold cross-validation (K=5) is used, though with only 5 resulting groups (3 healthy-file, 2 stroke-patient), some folds are single-class by construction rather than providing full subject-independent generalization evidence (see Section 3.7). The GRU-only model achieved a higher mean cross-validation accuracy (74.9%, SD 7.3%) than the CNN-GRU model (55.3%) and a Logistic Regression baseline (47.8%), though a paired comparison did not reach statistical significance (p = 0.125; see Section 5.1). An Integrated Gradients analysis on a single real test example showed differential channel importance concentrated on Channels 4 and 5; averaging across the test set is needed before generalizing this finding. The final model's Brier score (0.242) is reported for completeness, though it was computed on a single evaluation split containing no stroke-labeled test examples (see Section 4.3.2), so it reflects calibration on the negative class only; a supplementary, non-group-independent re-evaluation on a class-balanced split (Section 4.3.3) confirms the architecture can achieve genuine discrimination (ROC-AUC = 0.839, PR-AUC = 0.752) once both classes are present in the test set, at the cost of subject independence. Despite challenges related to dataset limitations and class imbalance, the results demonstrate the feasibility of GRU-based approaches for non-invasive, cost-effective stroke screening, providing a preliminary feasibility signal to guide future model refinement in a larger, prospectively validated cohort.