Abstracts

Rationale: Infantile spasms (IS) is a potentially catastrophic epilepsy syndrome of infancy. There are no established biomarkers that accurately predict response to specific treatments. Although hypsarrhythmia is clinically used as an EEG marker of IS disease activity, hypsarrhythmia does not predict response and identification of hypsarrhythmia is undermined by poor inter-rater reliability. If a robust biomarker were discovered which could predict response to specific treatments, treatment algorithms could be streamlined and duration of disease could be shortened in many patients, with the potential to dramatically improve long-term intellectual outcomes. We aim to develop an objective and robust EEG biomarker of IS disease activity. Utilizing a large EEG dataset, we performed a comparative analysis of different machine learning models to predict treatment response among children with IS. Methods: We retrospectively identified children with IS at our center who received extended EEG prior to initiation of any medical treatment. Treatment response was defined as resolution of both IS and hypsarrhythmia with no relapse for at least three months. Treatment non-response was defined as persistent IS despite treatment for at least three months. Predictive models evaluated included logistic regression, support vector machine, and random forest models trained to predict the likelihood of treatment response or nonresponse. For a reference model (RefMod), we selected two clinical features readily available at initial evaluation (age at time of EEG, age at IS onset) and two EEG features (relative delta power, averaged absolute amplitude) that capture core features of hypsarrhythmia (background slowing with high amplitudes) for inclusion into a logistic regression model. Models for comparison additionally included a variety of linear and non-linear quantitative EEG (qEEG) measures of correlation, connectivity, entropy, and fractal dimension. EEG features were calculated on subject EEG data segmented into 2.5 min epochs. Analyses was performed on a per sample basis by pooling samples across all subjects and per subject basis by averaging features per patient. Training and validation were performed with K-fold cross-validation with 10 folds. Accuracy and AUC ROC analysis results were calculated. An EEG feature exploration tool was developed using t-Stochastic Neighbor Embedding (tSNE) to inspect clustered EEG segment raw data and features (Figure 1). Results: We identified 58 children with IS who obtained extended video EEG prior to treatment. 64% were initial presentations without prior treatment. There were 25 treatment non-responders (NR) and 33 responders (R). The average age of spasms onset was 130 +- 72 d and 206 +- 90 d, and average latency to treatment was 61 +- 57 d and 21 +- 25 d for treatment NR and R respectively. An average of 14 hours of EEG was included per patient, yielding in total ~20,000 EEG segments each of length 2.5 minutes, with 47% and 53% NR and R respectively. The Random Forest model (RFM) demonstrated the best performance with per sample analysis accuracy 69% and AUC 0.77 +- 0.06 compared to reference model accuracy 59% and AUC 0.61 +- 0.24. For the per subject analysis, RFM had accuracy 68% and AUC 0.79 +- 0.09 and RefMod had accuracy 67% and AUC 0.67 +- 0.24 (Figure 2). tSNE showed more effective treatment outcome class separation with qEEG features compared to RefMod features (Figure 1). Conclusions: Using readily obtainable pre-treatment clinical variables (age and treatment latency) in combination with pre-treatment quantitative EEG features, the proposed machine learning approach predicts short-term treatment outcome in IS with good accuracy. Funding: No funding