A pilot study of single machine learning classifier pipeline to detect infantile spasms syndrome from clinical data

Background: Diagnosis of Developmental and Epileptic Encephalopathies requires an EEG for confirmation and often delayed due to multiple factors. In collaboration with Patient Groups, Engineers, Machine Learning Experts and Clinical Team, we aim to develop a digital solution to improve early detection. In this study Infantile Spasms Syndrome (ISS), well defined clinical syndrome with characteristic EEG abnormalities was selected. This study assesses if a machine learning (ML) pipeline can accurately detect ISS using existing specialist confirmed EEG recordings.

Method: The pipeline explores a dataset (diagnosed with ISS) from a Children's Hospital EEG database. This dataset is annotated by experienced physiologists for EEG characteristics. The Pipeline evaluates the impact of applying Current Source Density (CSD) to mitigate volume conductance, comparing results with and without CSD. A data augmentation module quadruples the dataset size. The pipeline processes the data in two ways: first, by calculating temporal, frequency, correlation, and graph theory-derived features; and second, by using the unprocessed data.

Results: A total of 256 seizures (median duration 3 seconds) were used in this ongoing work. The pipeline examines nine different ML and DL models, incorporating both feature-engineered and unprocessed data approaches. Logistic Regression (LR) Leave One Out Cross Validation (LOOCV) and Recursive Feature Elimination Cross Validation (RFECV) are applied, exploring both 18 and 6 bipolar EEG channel configurations. The best performance was achieved by LR LOOCV for the 18-channel montage with CSD (ROC AUC of 0.73 ± 0.12).

Conclusions: Developing effective ML classifiers for ISS identification is possible. Our results encourage further research in the area of more advanced machine learning algorithms and deep learning methodologies as well as reducing the number of channels to facilitate the ML-based IS detection. Furthermore, the refinement of our work will be enabled through the incorporation of additional carefully annotated clinical data, which is currently underway.