Predicting neurological recovery in comatose patients after cardiac arrest remains a critical challenge in clinical practice, with limited accuracy of current prognostic tools. This study aims to develop and validate a transfer learning-based machine learning model for predicting neurological recovery in patients following cardiac arrest as part of the George B. Moody PhysioNet Challenge 2023. We used the I-CARE dataset of 485 patients who experienced cardiac arrest, with 412 in the development set and 73 in the validation set. To analyze the spectrogram, we utilized the multi-tapering method to calculate power spectral density by sliding a 30-second window within a 5-minute EEG epoch at 5-second intervals and estimated the mean of alpha, beta, delta, and theta bands. EEG quality score, age, sex, return of spontaneous circulation (ROSC), out-of-hospital cardiac arrest, ventricular fibrillation, and targeted temperature management were also included in the random forest model, currently tested for leaderboard score. We plan to implement transfer learning techniques by pre-training a 1D-CNN model on VitalDB, an open dataset of surgical patients. We assume that the EEG patterns of patients recovering from anesthesia share similarities with those of comatose patients recovering from cardiac arrest, as both populations transition from unconsciousness to consciousness. Leveraging these similarities, the pre-trained model will be fine-tuned on the I-CARE dataset to improve accuracy in predicting neurological recovery. The evaluation metric demonstrated scores of 0.05, 0.18, 0.08, and 0.24 at 12, 24, 48, and 72 hours from ROSC for the random forest model in the testing set. Ongoing work with the 1D-CNN model showed scores of 0.11, 0.24, 0.51, and 0.68 at corresponding time points in the validation set and will be tested in the future.
In conclusion, our study aims to improve the prediction of neurological recovery in comatose patients after cardiac arrest by leveraging transfer learning techniques.