Contrastive Learning for Predicting Neurological Outcome in Comatose Pediatric Patients After Cardiac Arrest

Abstract

Accurate early prognostication after pediatric cardiac arrest is clinically crucial yet remains methodologically challenging. This thesis frames the problem as a representation learning task on raw electroencephalography (EEG) and introduces a fully self-supervised pipeline followed by a supervised classifier. A Time-Series-to-Vector (TS2Vec) encoder is trained on 6,669 unlabeled twenty-second EEG epochs (T = 2000 time points, C = 4 channels) from 84 comatose children treated at Erasmus MC Sophia Children’s Hospital. The encoder learns to map each EEG epoch to a 320-dimensional vector by contrasting transformed views of the same signal from unrelated signals, without using outcome labels.

Downstream classification is performed with a k-nearest neighbors (k-NN) model. Probabilities are assigned at the epoch level and aggregated across all epochs of a patient to generate a final patientlevel prognosis. Five-fold patient-level cross-validation gives an area under the ROC curve of 0.861 ± 0.092, an accuracy of 0.774±0.045 and an F1 score of 0.733±0.057. Critically, specificity and precision are both 1.000 ± 0.000. Thus the model never predicts death for a child who survives, satisfying a stringent clinical safety requirement. Sensitivity is 0.581 ± 0.070, reflecting a deliberately conservative decision threshold.

Qualitative analyses offers tentative support for the physiological relevance of the learned features. t-SNE projections show clustering by patient identity and good separation between the two extreme EEG background patterns (continuous normal activity and electrocerebral silence) without supervision. Saliency analysis did not yield clinically interpretable patterns.

Compared to the leading feature-engineered qEEG baseline, the proposed approach achieves equivalent specificity and a slightly lower AUC, while removing the need for manual feature design and expert annotation. The modular architecture invites several extensions, including graph-based encoders that respect electrode topology, integration of auxiliary data such as ECG and clinical metadata, retrieval-based reasoning using large historical EEG archives, and more interpretable or end-to-end trainable models. Overall, this work demonstrates that contrastive learning can be effectively applied to raw pediatric EEG for conservative, label-efficient outcome prediction after cardiac arrest, and offers a mathematically grounded proof of concept for future clinical applications.