Will EEG AI work on the next ward?

The task behind the number

Testing older Persyst software versions in continuous inpatient EEG, Ganguly and colleagues reported 50% sensitivity for individual seizures in a pilot cohort and 91% for seizure-positive recordings in a larger analysis. These answer different questions: identifying a recording that contains seizures does not mean finding every seizure within it. [1]

Settings and software versions matter too. Koren and colleagues found different balances between seizure detection and false positives when comparing updated commercial packages in an epilepsy-monitoring dataset. [2] For Sentinel, fewer alerts would be valuable only alongside acceptable detection and timely warnings.

Evidence of transfer needs boundaries

There is encouraging evidence. Mansilla and colleagues tested a fixed routine-EEG model on 104 patients at a new centre using different equipment. However, the deliberately balanced sample excluded ICU patients and assessed routine interpretation. Its findings cannot establish performance for continuous intensive-care alarms. [3]

Newer versions also need fresh tests. AccuRASE II compared an updated seizure-burden estimator on the earlier study’s 235 recordings, which the authors report were excluded from training. This supports version comparison, while leaving replication in new patients as a separate question. [4]

What this means for Sentinel

Our proposed evaluation should fix the model and alert settings before testing on new hospitals and patients. It should assess the whole monitoring pathway: the signal acquired, seizures detected, false alerts, warning delay and periods when the recording cannot be interpreted.

The SzCORE framework offers a starting point for consistent scoring, including rules for matching detections to seizure events. [5] Our interpretation is that a credited event match and a timely bedside warning need separate assessment. We would also measure the review burden on clinicians.

The useful test is whether Sentinel delivers dependable information in time to act. Demonstrating that will require evidence from Sentinel itself.

Sources

1. Ganguly et al. (2022). Seizure Detection in Continuous Inpatient EEG: A Comparison of Human vs Automated Review. Neurology. Results checked against the primary abstract and publisher supplement.
2. Koren et al. (2025). Systematic comparison of Commercial seizure detection Software: Update equals Upgrade? Clinical Neurophysiology. Comparison summarised from the authors’ institutional abstract.
3. Mansilla et al. (2024). Generalizability of electroencephalographic interpretation using artificial intelligence: An external validation study. Epilepsia. External routine-EEG evaluation.
4. Sheikh et al. (2026). Accuracy of the rapid-response electroencephalography’s Automated Seizure Burden Estimator: A follow-up validation study of version 8 (AccuRASE II). Epilepsia. Retrospective version comparison.
5. Dan et al. (2025; online 2024). SzCORE: Seizure Community Open-Source Research Evaluation framework for the validation of electroencephalography-based automated seizure detection algorithms. Epilepsia. Evaluation framework.