Overview

In this study, we leverage multi-modal, longitudinal data from ABCD study and multi-kernel machine learning approaches to accurately classify ambiguous OCD cases and increase power to downstream genetic analyses. We leverage logistic regression model with elastic net penalty to train each kernel independently, followed by accuracy-weighted case probability score summed across kernels

DataSets

ABCD Study data were accessed through National Institute of Mental Health (NIMH) 60 Data Archives (NDA). We are using different types of Datasets for building our prediction model.

• Genetic Data : Genetic data for participants were derived from blood and saliva samples and 66 genotyping on National Institute for Drugs and Addiction (NIDA)
• NeuroImaging Data: Tabulated summaries of post-processed neuroimaging data obtained via structural 84 magnetic resonance imaging (MRI) and diffusion MRI scans were downloaded from 85 NDA for imaging kernels.
• Questionarrie and phenotype Data: Tabulated questionnaire and phenotype data were downloaded from NDA for 93 questionnaire kernel and label definition.

A total of 6 processed data sets were defined for all samples after processing the above datasets:

1) Congnitive and Behavioural Phenotype
2) Genotypes
3) Diffusion MRI
4) Structural MRI
5) Rest State FMRI
6) Task FMRI

Machine Learning Model:

Method

We built 6 different kernels based on these datasets categories and are combined using weighted probabilistic accuracy kernel to make prediction on final test data-set. We trained each kernel using Logistic Regression model with elastic net as the penalty for every batch.

Support Method

We compared model performance using Cohen’s kappa score to Cohen’s kappa scores reported in psychiatry and psychiatric epidemiology literature.

Success Measurement

The aim of this project is to develop a machine learning model to classify individuals with OCD status. In order to measure our success, we have utilized confusion-matrix based assessments like accuracy, specificity and sensitivity. At the baseline, we expect accuracy of 0.5, sensitivity of 0.5, and specificity of 0.94, assuming prevalence of 0.5.

Languages and Tools used

• R-language for data preprocessing and cleaning the data.
• Machine learning Model: Logistic Regression, Elastic Net
• Python and Scikit learn used for computation and development of model.

Results

Performance of each kernel in terms of accuracy has been visually represented. Phenotype kernel vastly outperformed all other kernels in terms of accuracy. Additionally, plots of receiver operative curves and precision-recall curves are shown for each individual kernel. Based on AUC metrics and F1 scores, we can see that phenotype kernel outperforms all other kernels. The kernels based on imaging and genetic data performed similarly, very close to expected baseline performance.

Contributors

• Franjo Ivankovic
• Aryan Singh
• Shashank Kumar
• Richa Gupta