Abstracts

Rationale: Prognostication for patients post cardiac arrest (CA) is key but may be unreliable and suffer from inter-rater variability, necessitating implementation of objective methods to predict recovery. Quantitative measures derived from EEG (qEEG) and clinical data both contain prognostic information. We use a unique combined dataset of qEEG and clinical features from patient health records to predict prognosis using machine learning (ML) models. To compare with our feature based ML models, we take a novel approach by using a spectrogram based convolutional neural network (CNN) to predict patient outcome post CA, avoiding the need for explicit feature extraction by exploiting the intrinsic ability of the network to learn from data.

Methods: In a retrospective cohort study, we analyzed 101 EEGs recorded within 8 days after CA (Jan 2015-Sept 2020). Functional outcome of the patients was evaluated using the Glasgow-Pittsburgh Cerebral Performance Category (CPC) scale assessed within 3-6 months. Outcome was dichotomized as 1-2 (good outcome) denoting patient survival post CA with no more than moderate disability and CPC 3-5 (poor outcome). 27 qEEG and 9 clinical features were obtained (Table I). We compared the use of (1) qEEG features only, (2) clinical features only, and (3) the combination of both qEEG and clinical features using ML models: random forest (RF) and support vector machine (SVM). To explore outcome prediction without feature extraction, raw EEGs were converted to time-frequency spectrograms and used as input to our CNN model consisting of 3 blocks each containing a convolutional layer, batch normalization, ReLU activation and average pooling, proceeded by 2 fully connected layers. To evaluate both ML and CNN approaches, we use a 5-fold cross validation with different folding configurations (20 total models) to examine performance. A holdout test set containing 20 patients was used for evaluation.

Results: Table II shows performance metrics for RF, SVM and CNN models. Both ML models using either feature set individually achieve high performance. However, performance is further improved using a combined set of qEEG and clinical features, SVM (AUC = 0.91) and RF (0.97). When using qEEG features only, the AUC of our CNN model (0.92) exceeds that of RF (0.86) and SVM (0.82).

Conclusions: ML models using combined qEEG and clinical data to predict patient outcome outperform those using individual feature types. This suggests that the use of clinical data in conjunction with qEEG could improve prediction of prognosis and provide a more reliable tool to assist prognostication in the Intensive Care Unit. We also show how the CNN architecture can be applied to spectrograms derived from EEGs to predict outcome, outperforming our proposed qEEG ML models. This system of automatic feature extraction is advantageous over manually crafted feature approaches and can be further analyzed through the integration of clinical data into this model.

Funding: Please list any funding that was received in support of this abstract.: Queen’s University Faculty of Health Sciences.