autoSCORE

autoSCORE is a software-only device utilizing a deep learning Convolutional Neural Network to assist clinicians by analyzing scalp EEG recordings for both epileptiform and non-epileptiform abnormalities, providing probabilistic assessments and marker suggestions. Its methodology aims to enhance the efficiency and consistency of EEG review, with clinical studies demonstrating its ability to identify abnormalities with varying degrees of accuracy compared to human experts and predicate technologies.

autoSCORE is a software-only decision support tool intended to assist qualified medical practitioners in reviewing, monitoring, and analyzing scalp EEG recordings from patients aged three months and older, excluding neonatal and ICU settings. It employs a deep learning algorithm, specifically a Convolutional Neural Network trained on a large dataset, to process previously acquired EEG data. The software identifies and marks potential spikes corresponding to epileptiform abnormalities (Focal and Generalized) and also detects and classifies non-epileptiform abnormalities (Focal and Diffuse Non-epileptiform). It assesses the probability of these abnormalities and sends this information, along with suggested marker placements, to compatible EEG review software, without altering the original EEG data or providing a diagnostic conclusion. Clinical validation across five retrospective studies demonstrated autoSCORE's capabilities. When classifying EEGs as normal or abnormal at the recording level, it achieved high agreement with human experts (HEs), with sensitivities often above 80% and specificities generally between 84% and 96%. For specific abnormality types like Focal Epileptiform and Generalized Epileptiform, performance varied, with notable strengths in negative predictive value and correlation with HE assessments; for instance, Generalized Epileptiform detection showed sensitivities from 71.2% to 100%. While marker-level agreement was more modest, it improved with higher autoSCORE-assigned probability, and in comparisons for epileptiform detection, autoSCORE generally outperformed predicate devices in accuracy and specificity.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the performance of autoSCORE's non-epileptiform abnormality detection compare, in terms of specific metrics (sensitivity, specificity, PPV, NPV), to established methods or expert consensus for these less commonly automated findings?; What is the variability in autoSCORE's performance across different EEG recording montages, durations, or signal quality levels commonly encountered in routine clinical practice?; Given that marker-level agreement with human experts is lower than recording-level agreement, what are the primary types of discrepancies (e.g., over-marking, under-marking, misclassification of marker type), and how might these impact clinical workflow efficiency?; The validation studies used human expert assessments as a reference; how does inter-rater reliability among the human experts themselves compare to the agreement levels observed between autoSCORE and the expert consensus, particularly for nuanced or rare abnormalities?; How does the 'probability' output for abnormalities correlate with specific electrographic features or clinical contexts, and what is the optimal threshold for these probabilities to balance sensitivity and specificity for different clinical decision-making scenarios?