EEG Based Biomarker prediction using Deep Learning
Explore the docs »
View Demo
·
Report Bug
This study aims to predict sleep-related hormone levels utilizing electroencephalogram (EEG) data. This could be an effective alternative, as it would provide a less invasive means of capturing vital biomarkers than blood sampling methods. Deep learning has proven to be very useful in means of EEG analysis, specifically by utilizing Convolutional Neural Networks (CNNs) we could correlate key hormones such as cortisol and testosterone. This approach eliminates the need for frequent blood draws as model prediction could be done in real-time, allowing for continuous hormonal sampling. Significant findings could have a meaningful effect for endocrine research. Hormone Regression in Sleep The first iteration of the model included all four sleep stages in its traning set, and a single output layer to predict regression values for all 12 hormones (TAC mmol/L, ADA U/L, ADA2 U/L, %ADA2, GLU mg/Dl, PHOS mg/Dl, CA mg/Dl, CHOL mg/Dl, TRI mg/Dl, HDL mg/dL, LDL-C mg/Dl, CPK U/L) at once. This was very challenging, as the model struggled to converge on all 12 labels at the same time. The model’s input being a 19x19 square matrix, originally it seemed that the model could not successfully optimise all of the hormones hormones due to the diverse scale of the hormone levels. This was resolved by normalising the values to 0-1. This change improved the models convergence during training, but still could not generalise well to the unseen data, with the best performance a mere: 0.175 r2. Due to the challenges in the regression task, the modelling approach was adapted to suit a classification approach, where hormone levels were categorised into three discrete groups to enhance generalisation. Sleep stages were also separated during training to remove possibility of frequency sensitive hormonal fluctuations. These changes are further demonstrated using the regression performance as outlined below.
Hormone Classification in N1 Sleep The N1 sleep stage is characterised by light sleep, only lasting for around 5 minutes; there is not much data available, so training is completed on the entire dataset. This is particularly noticeable in comparison to the other sleep stages, demonstrating a much larger amount of variance between feature extraction methodology. This does not seem to prevent results in N1 however, as notable findings include 0.8816 for ADA2 U/L (alpha) and 0.7829 for TAC mmol/L (alpha), it is to be expected that the delta wavelength appears to perform the worse, as the N1 sleep stage has very little delta wave activity. Low performing hormones such as CHOL, TRI and LDL-C may indicate limitations for EEG-based correlations in lipid biomarkers. Topographic visualisations reveal critical electrode channels Fp3 and Fp4 driving N1’s high performance. These frontal channels reveal coherence patterns relevant to early sleep. Future work that have access to larger computation and datasets could expand on this key pairing with further frequency analysis, to improve performance.
https://pytorch.org/Major frameworks/libraries
Docker compose is as located inside of /frontend/docker_compose.yaml. Deployment based upon mounted Caddyfile.
Infrence models are deployed on eeg.maxh.work, where .MAT files are currently supported.
Distributed under the MIT License. See LICENSE.txt for more information.
Max - 26584263
Project Link: Here
