محمدمهدي ميرزاعلي محمدي

Accurate prediction of recovery in post-cardiac arrest comatose patients can be improved by incorporating EEG signals an‎d cardiac arrest characteristics. This study aims to leverage machine learning an‎d deep learning techniques to develop an‎d compare predictive models for patient outcomes, also identifying effective input features. Outcomes were classified into good an‎d poor categories based on the Cerebral Performance Category (CPC). Two commonly used deep learning architectures, ConvNeXt an‎d EEGNet, were implemented. To reduce data dimensionality, four key features were selec‎ted based on their relevance to neurological outcomes. Eight machine learning models including Naive Bayes, Logistic Regression, Decision Tree, Ran‎dom Forest, XGBoost, LightGBM, Support Vector Machine, an‎d K-Nearest Neighbors were trained. To enhance performance, five ensemble methods including bagging, boosting, stacking, voting, an‎d a rule-based were applied. The study included 350 ran‎domly selec‎ted patients for training an‎d validation, an‎d an independent group of 80 patients for testing. The accuracy of deep learning models in validation did not exceed 0.68, experiencing overfitting. Statistical analysis identified overall continuity, recovery progress, ROSC time, an‎d shockable rhythm as important features. Using the selec‎ted features, Logistic Regression achieved an accuracy of 0.79 an‎d an F1-score of 0.844. The majority voting ensemble combining NB, LR, an‎d RF had the highest accuracy of 0.83, F1-score of 0.83, an‎d AUC-ROC of 0.86, requiring 4.65 seconds of prediction time per patient. Our findings highlight the value of long temporal an‎d sequential EEG features such as continuity, an‎d support the use of ensemble learning to improve the reliability in the prediction of neurological recovery. Although deep learning models can detect various patterns, traditional machine learning classifiers can outperform them when overfitting is a critical problem. In these cases, dimensionality reduction plays a key role in predictive systems.

يادگيري ماشين , سيگنال الكتروانسفالوگرافي , پيشبيني بازيابي عصبي , بيماران كما , ايست قلبي

machine learning , electroencephalography signal , neurological recovery prediction , comatose patients , cardiac arrest

Mohammad Mahdi Mirza Ali Mohammadi

Dr. Ali Sadr