Exploring Associations Between Infant EEG and Genetic Risk for Schizophrenia
Beheshti Dehkordi, Mohammadamin (2023)
Beheshti Dehkordi, Mohammadamin
2023
Master's Programme in Biomedical Technology
Lääketieteen ja terveysteknologian tiedekunta - Faculty of Medicine and Health Technology
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2023-12-18
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2023112610239
https://urn.fi/URN:NBN:fi:tuni-2023112610239
Tiivistelmä
Schizophrenia is a complex mental disorder that is typically identified in adulthood, but early diagnosis remains a significant challenge. In this research study, our primary goal was to investigate the relationship between brain activity and the genetic risk of schizophrenia. We utilized Electroencephalogram (EEG) signals as a critical investigative tool to explore this connection. To achieve this, we conducted an extensive analysis involving 56 infants, whose genetic risk scores for schizophrenia had been calculated.
The dataset consisted of sleep EEG of 56 children. Sleep EEG was classified into sleep stages. After that EEG of each child was sampled with an extensive array of 96 EEG features across several signal channels. After calculating all combinations of channels, sleep stages of interest, and the EEG features, we ended up with 5184 numerical scores for each participant. We collectively referred to these as "Feature-Channel-Sleep-Stage Combinations", or FCSCs. The FCSCs included a wide range of EEG characteristics, EEG channels, and sleep stages. The key objective of the study was to identify specific FCSCs, either on an individual basis, or in groups, that could be linked to the genetic risk scores for schizophrenia. Unlike previous studies that have typically began with a limited number of EEG features, our research encompassed 96 EEG features. Besides statistical testing and correlational analysis, the primary methodology for relevant FCSC identification was based on machine learning -based regression from FCSCs to genetic risk scores, as paired with an automatic feature selection algorithm. Although the present dataset was small relative to the number of FCSCs, leading to a potential issue of overfitting, we took meticulous methodological steps mitigate this problem through a combination of machine learning and statistical techniques.
The data analysis yielded a number of findings. Notably, we discovered that sleep stages S2 and S4 and bipolar EEG channels were statistically linked to the genetic risk for schizophrenia. The data analysis revealed a total of 14 individual FCSCs, along with two groups of FCSCs, that displayed both linear and nonlinear statistical dependencies with the genetic risk scores for schizophrenia. Finally, by utilizing the most promising FCSCs detected by the data analysis pipeline, and as manually further filtered to a subset of the most robust signal descriptors by an experienced EEG-clinician, a substantially above-chance genetic risk prediction accuracy was obtained in a cross-validation experimental setup.
Overall, this study identified a set of promising features for schizophrenia genetic risk estimation from infant EEG. Given the exploratory nature of the work, future studies should aim validate these findings on an independent dataset.
The dataset consisted of sleep EEG of 56 children. Sleep EEG was classified into sleep stages. After that EEG of each child was sampled with an extensive array of 96 EEG features across several signal channels. After calculating all combinations of channels, sleep stages of interest, and the EEG features, we ended up with 5184 numerical scores for each participant. We collectively referred to these as "Feature-Channel-Sleep-Stage Combinations", or FCSCs. The FCSCs included a wide range of EEG characteristics, EEG channels, and sleep stages. The key objective of the study was to identify specific FCSCs, either on an individual basis, or in groups, that could be linked to the genetic risk scores for schizophrenia. Unlike previous studies that have typically began with a limited number of EEG features, our research encompassed 96 EEG features. Besides statistical testing and correlational analysis, the primary methodology for relevant FCSC identification was based on machine learning -based regression from FCSCs to genetic risk scores, as paired with an automatic feature selection algorithm. Although the present dataset was small relative to the number of FCSCs, leading to a potential issue of overfitting, we took meticulous methodological steps mitigate this problem through a combination of machine learning and statistical techniques.
The data analysis yielded a number of findings. Notably, we discovered that sleep stages S2 and S4 and bipolar EEG channels were statistically linked to the genetic risk for schizophrenia. The data analysis revealed a total of 14 individual FCSCs, along with two groups of FCSCs, that displayed both linear and nonlinear statistical dependencies with the genetic risk scores for schizophrenia. Finally, by utilizing the most promising FCSCs detected by the data analysis pipeline, and as manually further filtered to a subset of the most robust signal descriptors by an experienced EEG-clinician, a substantially above-chance genetic risk prediction accuracy was obtained in a cross-validation experimental setup.
Overall, this study identified a set of promising features for schizophrenia genetic risk estimation from infant EEG. Given the exploratory nature of the work, future studies should aim validate these findings on an independent dataset.