Machine learning based seizure classification and digital biosignal analysis of ECT seizures

Abstract

While artificial intelligence has received considerable attention in various medical fields, its application in the field of electroconvulsive therapy (ECT) remains rather limited. With the advent of digital seizure collection systems, the development of novel ECT seizure quality metrics and treatment guidance systems in particular will require cutting-edge digital seizure analysis. Using artificial intelligence will offer more analytical degrees of freedom and could play a key role in enhancing the precision of currently available procedures. To this end, we developed the first machine learning (ML) framework that can classify ictal and non-ictal EEG segments, accurately identifying seizure endpoints-a critical step in deriving seizure quality parameters-and computing these metrics at least as reliable as existing precomputed scores. The ML model retained in this study effectively discriminated ictal from non-ictal EEG segments with 89% accuracy, precision, and sensitivity. The reproduced ECT quality parameters showed correlations up to ϱ = 0.99 (p < 0.01) with the pre-calculated values from the stimulation device and did not significantly differ from the reference values. Mean seizure duration differences were 0.23 ± 15.59 s compared to the expert rater and 0.28 ± 16.19 s compared to the stimulation device. The study highlights the potential of integrating ML into the field of ECT and emphasizes the critical role of a highly sensitive seizure detection method in reliably determining seizure duration and deriving subsequent quality indices, paving the way for more individualized treatment strategies and novel approaches to determine seizure quality.

Fig. 1

Simplified flowchart including machine learning pipeline for ECT-induced seizure detection, seizure quality index extraction, and resulting research opportunities.

Fig. 2

Classification performance and exemplary EEG processing framework. (a) Validation metrics of tuned machine learning models. (b) ROC curve of optimized classifiers, (c) Single channel EEG with ECT inducted seizure, segmented into equally sized time windows with corresponding seizure probabilities. Additionally, were added the experimental (Framework), the precomputed (Thymatron®) as well as the seizure endpoint provided by the expert rater. The subplot below depicts various segments associated with the extraction of seizure quality parameters. The last subplot demonstrates the application of power spectral analysis along the EEG.

Fig. 3

Spearman correlations (left columns) and distributions (right columns) for seizure quality indices calculated by the framework, by the device and rated by the expert. In the subset, rated by the expert: The relationships between the framework, ECT device (Thymatron), and expert ratings for seizure duration (A) and PSI (B) are shown; the relationship between the expert and the stimulation device is shown for seizure duration (E) and for PSI (H). Relationships between the expert and the framework are plotted for the seizure duration (C) and PSI (F). Relationships between the framework and ECT device are plotted for seizure duration (D), PSI (G), ASEI (I), EIA (J), MIA (K), MSC (L), MSP (M), and TTPP (N).