Predicting alcohol use disorder remission: a longitudinal multimodal multi-featured machine learning approach

Stacey Subbie‐Saenz De Viteri ,Bernice Porjesz,Ashwini K. Pandey,Danielle M. Dick,Grace Chan,Chella Kamarajan,Jessica E. Salvatore,Meredith W. Francis,Jacquelyn L. Meyers,Fazil Alıev ,Lance O. Bauer ,John I. Nürnberger ,Weipeng Kuang,Kathleen K. Bucholz,Sivan Kinreich,Jian Zhang,Marc A. Schuckit,David B. Chorlian,Tatiana Foroud ,Victor Hesselbrock,Jessica L. Bourdon,Vivia V. McCutcheon,Gayathri Pandey,Andrey P. Anokhin

doi:10.1038/s41398-021-01281-2

Abstract

Predictive models for recovering from alcohol use disorder (AUD) and identifying related predisposition biomarkers can have a tremendous impact on addiction treatment outcomes and cost reduction. Our sample (N = 1376) included individuals of European (EA) and African (AA) ancestry from the Collaborative Study on the Genetics of Alcoholism (COGA) who were initially assessed as having AUD (DSM-5) and reassessed years later as either having AUD or in remission. To predict this difference in AUD recovery status, we analyzed the initial data using multimodal, multi-features machine learning applications including EEG source-level functional brain connectivity, Polygenic Risk Scores (PRS), medications, and demographic information. Sex and ancestry age-matched stratified analyses were performed with supervised linear Support Vector Machine application and were calculated twice, once when the ancestry was defined by self-report and once defined by genetic data. Multifeatured prediction models achieved higher accuracy scores than models based on a single domain and higher scores in male models when the ancestry was based on genetic data. The AA male group model with PRS, EEG functional connectivity, marital and employment status features achieved the highest accuracy of 86.04%. Several discriminative features were identified, including collections of PRS related to neuroticism, depression, aggression, years of education, and alcohol consumption phenotypes. Other discriminated features included being married, employed, medication, lower default mode network and fusiform connectivity, and higher insula connectivity. Results highlight the importance of increasing genetic homogeneity of analyzed groups, identifying sex, and ancestry-specific features to increase prediction scores revealing biomarkers related to AUD remission.

Highlights

National surveys on alcohol use statistics and AlcoholUse Disorder (AUD) studies show that only one-third of it is of great importance to be able to identifyalcohol use disorder (AUD) resilience and readiness to recover features including predisposition characteristics that can predict a Kinreich et al Translational Psychiatry (2021)11:166 change in drinking behavior, impacting therapeutic approaches to AUD, helping individuals overcome addiction and overall reducing state, and federal associated financial burden
The AA male group combined feature model of Polygenic risk scores (PRS), especially resting-state functional connectivity (EEG-functional connectivity (FC)), marital status, and employment status achieved the highest accuracy of 86.04%
Results confirmed previous results showing that the combined feature model (e.g., EEG, PRS, medication, and demographic information) achieved a higher prediction score than models based on single domain suggesting that genetics prediction models will improve from the addition of phenotypes to the calculation

Summary

Introduction

National surveys on alcohol use statistics and AlcoholUse Disorder (AUD) studies show that only one-third of it is of great importance to be able to identifyAUD resilience and readiness to recover features including predisposition characteristics that can predict a Kinreich et al Translational Psychiatry (2021)11:166 change in drinking behavior, impacting therapeutic approaches to AUD, helping individuals overcome addiction and overall reducing state, and federal associated financial burden. Our own study[8] and others[7], have shown that the accuracy of ML models increases by using multimodal, multi-features approaches to describe complex disorders, permitting a variety of measurement domains that could be brought to bear on different aspects of disease pathology[7]. ML studies calculating AUD classifiers/predictive models have employed genetic loci[8], psychosocial[7], family history[8], and electrophysiological (EEG) measurements[8] as features in a multimodal analysis. In the current ML study, we have utilized EEG, genetics, medication intake, and demographic as predisposition characteristics to predict AUD remission. Polygenic risk scores (PRS), which summarize the effects of genome-wide association study (GWAS) markers to measure the genetic liability to a trait or a disorder, have shown promise in predicting human complex traits and diseases[11,12]. Current medication intake was added as a potential feature to the calculated

Methods

Results

Conclusion