This study investigates the effectiveness of brain-state-specific autoencoder models in reconstructing neural activity in the motor cortex, focusing on beta (13-30 Hz) and gamma (30-100 Hz) oscillations. Using data from macaque monkeys performing a motor task, each trial from this dataset was given a label of ‘beta’ or ‘gamma’, found using each trial’s respective local field potential recording. We trained autoencoder models separately on beta-only, gamma-only, and a mix of these trials of labeled spiking activity. We hypothesized that models tailored to specific brain-wave states would yield better reconstruction of neural activity of the same label during task execution. Our results showed no significant difference in reconstruction error between the brain-state-specific models and the combined model. However, when comparing models trained on mismatched brain states, significant differences in reconstruction loss were observed, highlighting the importance of using correctly labeled data. Based on these findings, a single combined model trained on all data was found to be more effective than separate models for distinct brain states.

This project utilized the International Cardiac Arrest Research Consortium (I-CARE) database which contained physiological data about individuals who had suffered cardiac arrest (CA). In a previous study, a model predicting Cerebral Performance Category (CPC) score was developed by primarily using smaller amounts of electroencephalogram (EEG) data. This project's target was to develop a classifier model that incorporates both EEG and electrocardiogram (ECG) data features as a predictor of binary and CPC scores post-CA. 

The novelty and unique contributions of this study are listed below: 

• The combination of electrocardiography and electroencephalography pre-processed features to create machine learning input vectors 

• Utilized Principal Component Analysis to create a dimensionally reduced dataset with two components for machine learning 

• Finding the most relevant features across the defined EEG and ECG datasets for machine learning 

To achieve the desired outcome, raw ECG data had to be filtered from 250-500 Hz, and feature extraction (e.g. QRS amplitude and Heart Rate) performed on the one-hour long signals. EEG data underwent a robust pre-processing pipeline including digitization and a 19-channel net- utilized to better extract the frequency specific features of neural oscillations with a multi-taper frequency transformation. PCA led to the discovery of two components contributing to 93% of the variation in data. Three sets of data were used to develop several classifier models, each further split into two identical features, except the CPC classifications (which were left as-is in one copy of each dataset and binarized in the other). Thus, six total datasets were produced (randomly divided into 80/20 split for model testing and training). The best overall model and dataset combination proved to be the Random Forest architecture (built with 100 estimators), trained on binary EEG/ECG data. One apparent limitation in this project includes the limited use of data in the I-CARE database (due to time constraints). 

A study published in 2021 used the same ICARE dataset to develop a deep neural network (DNN) to leverage only the EEG features of the data to predict CPC scores [21]. This study achieved a time-dependent AUC of 0.83 using a cohort of 1,038 cardiac patients from the dataset, also using a binarized version of classification. Considering both the sample size and complexities of the 2021 model versus those of the model presented here, it can be extrapolated that with further refinement, the current model can out preform the 2021 model with the addition of pre-processed ECG features.