Adaptive EEG-based stroke diagnosis with a GRU-TCN classifier and deep Q-learning thresholding

TL;DR

The paper addresses rapid, bedside stroke triage using EEG by developing an adaptive multitask classifier that converts 32-channel EEG to spectral features and employs a GRU-TCN to predict stroke type, hemispheric lateralization, and severity. An accompanying deep Q-network thresholding module dynamically adjusts decision thresholds to optimize clinically relevant sensitivity–specificity trade-offs, improving macro-F1 for stroke-type classification from 92.8% to 97.7% and achieving high accuracy and AUC metrics on the primary dataset. Robustness tests on an independent, low-density EEG cohort showed performance declines due to domain shift, but adaptive thresholding continued to provide benefits over static thresholds. The approach is complemented by an interpretable GUI with scalp topographies and spectral views, supporting clinician review; overall, the work demonstrates that learned threshold adaptation can meaningfully enhance EEG-based diagnostic performance while preserving interpretability and calibration, albeit with limitations related to sample size and generalizability.

Abstract

Rapid triage of suspected stroke needs accurate, bedside-deployable tools; EEG is promising but underused at first contact. We present an adaptive multitask EEG classifier that converts 32-channel signals to power spectral density features (Welch), uses a recurrent-convolutional network (GRU-TCN) to predict stroke type (healthy, ischemic, hemorrhagic), hemispheric lateralization, and severity, and applies a deep Q-network (DQN) to tune decision thresholds in real time. Using a patient-wise split of the UCLH Stroke EIT/EEG data set (44 recordings; about 26 acute stroke, 10 controls), the primary outcome was stroke-type performance; secondary outcomes were severity and lateralization. The baseline GRU-TCN reached 89.3% accuracy (F1 92.8%) for stroke type, about 96.9% (F1 95.9%) for severity, and about 96.7% (F1 97.4%) for lateralization. With DQN threshold adaptation, stroke-type accuracy increased to about 98.0% (F1 97.7%). We also tested robustness on an independent, low-density EEG cohort (ZJU4H) and report paired patient-level statistics. Analyses follow STARD 2015 guidance for diagnostic accuracy studies (index test: GRU-TCN+DQN; reference standard: radiology/clinical diagnosis; patient-wise evaluation). Adaptive thresholding shifts the operating point to clinically preferred sensitivity-specificity trade-offs, while integrated scalp-map and spectral visualizations support interpretability.

Figures (6)

• Figure 1: Confusion matrices for (a) stroke-type (3 $\times$ 3), (b) severity, and (c) lateralization. Color scales are uniform across panels; numbers are counts on the test set.
• Figure 2: ROC and PR curves. (a) Stroke-type (one-vs-rest, micro and macro aggregation), (b) severity (binary), and (c) lateralization (binary). Shaded bands indicate 95% CIs via patient-level bootstrap efron1994.
• Figure 3: DQN adaptation. (a) Training curve of episodic return (mean $\pm$ s.e.m.). (b) Per-patient operating-point shifts (TPR vs. FPR) from static (open circles) to DQN-adapted (filled circles). (c) Reliability diagrams (ECE in legend with 95% CIs) before and after adaptation guo2017.
• Figure 4: Generalization to a naive EEG cohort. Stroke-type performance for static and DQN-adapted models on the original test set vs. the independent ZJU4H dataset tong2025eegfusion. Points show estimates; whiskers denote 95% CIs (bootstrap).
• Figure 5: GUI examples.(a) Main dashboard with file selection, preprocessing controls (0.5 Hz high-pass, 50 Hz notch filter), and analysis parameters including ICA and time window settings. (b) Scalp map showing Delta band (0.5-4 Hz) topographic activity with 32-channel electrode positions, anatomical landmarks, and interactive frequency band selection controls. (c) Diagnostic results displaying classification outputs: stroke type (Hemorrhage), severity (Severe), and localization (Right Hemisphere).