Machine Learning Reads ECT Seizure Adequacy From Raw EEG Signals
Electroconvulsive therapy remains one of psychiatry’s most effective interventions, but the electrical dose still depends on a clinician’s judgment of seizure quality. In this preprint, investigators tested whether machine learning could classify seizure adequacy directly from raw electroencephalogram data, turning a subjective bedside assessment into a data-driven signal. The question matters because dosing too low risks ineffective treatment, while dosing too high can add unnecessary adverse effects without improving benefit.
The team assembled 1,000 ECT session EEGs from 130 patients and labeled each tracing into three actionable categories: null-response, threshold, and adequate. Those categories reflect a practical clinical sequence in ECT titration. A null-response implies no useful seizure activity, threshold indicates the minimum response, and adequate marks the level most likely to support effective therapy. In other words, the model was not asked to make a vague judgment about “good” or “bad” EEGs; it was trained to separate treatment-relevant stages that can influence the next stimulus dose.
Strong Separation at the Bedside-Meaningful Cut Points
The best-performing algorithm achieved an area under the receiver operating characteristic curve above 0.9 for both null-response versus threshold and threshold versus adequate. In a classification task, that level of discrimination indicates excellent performance. The signal was strongest where clinical decisions are most immediate: determining whether the seizure has crossed from absent to minimally adequate, and whether it has advanced from threshold to adequate.
One sentence captures the main result plainly: the raw EEG contained enough structure for machine learning to distinguish seizure adequacy with high accuracy at the key decision boundaries.
The biology behind that result is plausible. ECT seizures generate patterned cortical activity that varies in organization, amplitude, and evolution across the stimulus response. If those features are captured in the raw waveform, an algorithm can learn the signatures of a meaningful seizure better than a human observer scanning a complex tracing in real time. That does not eliminate clinical interpretation, but it suggests that the signal itself may carry more information than the manual grade currently used to guide dosing.
Why Added Clinical Variables Did Not Help
The investigators also tested whether demographic data, anesthesia exposure, and anticonvulsant exposure could improve model performance. They did not. In practical terms, the raw EEG already seemed to contain the dominant predictive information, leaving little room for baseline clinical variables to add discrimination.
That finding is clinically interesting because it suggests seizure adequacy may be encoded more directly in the electrophysiologic response than in broad patient characteristics. Anesthetic agents and anticonvulsants are relevant to seizure threshold and expression, but in this dataset they did not improve the algorithm once the EEG itself was available. For a decision-support tool, that is a useful design lesson: the most informative input may be the most immediate one, not the most easily tabulated one.
Prediction of Later Treatments Was Less Reliable
Performance dropped when the model tried to predict seizure quality in subsequent treatments. Accuracy was only moderate. That matters because serial ECT is not static; seizure threshold can shift across sessions, medications can change, anesthetic choices can vary, and the physiologic response may drift over time. A model that performs well on the current EEG does not automatically generalize to the next session if the clinical context has moved.
This is where the proof-of-concept nature of the work becomes clearest. The algorithm can classify a recorded seizure with impressive precision, yet forecasting future seizure quality is a harder problem. The underlying reasons are likely familiar to ECT clinicians: treatment courses are dynamic, and the seizure response is influenced by variables that evolve from session to session. A static model trained on one trace may struggle when the next trace reflects a different threshold state.
Objective Dosing Could Reduce Variability in ECT Practice
ECT dosing remains heterogeneous across clinical settings because seizure adequacy is often judged manually. That subjectivity can introduce inconsistency between operators, hospitals, and even individual treatment sessions. A machine learning tool built on raw EEG could provide a more standardized reference point, especially during titration when small dosing decisions have outsized consequences.
There is also a safety angle. Better recognition of inadequate or borderline seizures could reduce repeated underdosing, while clearer recognition of adequate seizures could help avoid unnecessary escalation. The concept is not to replace clinicians, but to give them a more reproducible signal at the moment when the stimulus dose is being adjusted.
The authors’ provisional patent application signals that the group sees translational potential in the method. At the same time, the preprint status keeps expectations appropriately restrained. This is early-stage work, not a ready-to-deploy clinical tool, and the dataset came from a defined cohort rather than a broad multicenter implementation.
What This Proof-of-Concept Does and Does Not Show
The study shows that raw EEG can support automated classification of ECT seizure adequacy with high discrimination at the most important thresholds. It does not show that the model improves patient outcomes, reduces adverse effects, or can be dropped into routine practice without further validation. It also does not answer how the algorithm would perform across different machines, institutions, electrode placements, or treatment protocols.
The data cannot be made publicly available, which may further limit external replication. Even so, the work establishes a technical foundation for more objective ECT support systems. If future validation confirms that this approach generalizes across settings, clinicians could gain a tool that reduces reliance on subjective seizure grading and makes dosing decisions more consistent.
For now, the result is narrower but concrete: raw EEG appears capable of separating null-response, threshold, and adequate seizures with strong accuracy, while session-to-session prediction remains only moderate and still needs a better solution.