Logo image
Back
Developing clinical prediction models for post-treatment substance use relapse with explainable artificial intelligence
Dissertation   Open access

Developing clinical prediction models for post-treatment substance use relapse with explainable artificial intelligence

Ou Stella Liang
Doctor of Philosophy (Ph.D.), Drexel University
Sep 2022
DOI:
https://doi.org/10.17918/00001328
pdf
Liang_Ou_202212.58 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Alcoholism--Relapse Substance abuse--Relapse Opioid abuse Prediction theory--Mathematical models Clinical prediction model Machine Learning Quantitative Psychology
Substance use disorder is a chronic mental health condition that devastates the lives of many. Substance use relapse is common during treatment, which is not only demoralizing but also potentially dangerous as it may lead to life-threatening overdose. Predicting substance use relapse is an urgent yet challenging problem as evident in the high rate of post-treatment relapse despite four decades of research, owing to the multitude of neuro-psycho-social risk factors and the complexity of their interaction, many of which are hard to reproduce due to methodological limitations of legacy studies. A universal clinical prediction model (CPM) for substance use relapse does not currently exist. In this dissertation, we developed substance-specific, machine learning-based CPMs using electronic health records from the nation's largest nonprofit addiction treatment provider. The calibrated, cross-site validated CPMs achieved good discrimination exceeding the performance of human experts and were capable of generating individualized, realistic risk estimates. This work identified discriminative features and their effects from a large library of clinical questionnaires to inform clinical practice with influential risk and protective factors of substance use relapse. Greater participation in group and individual sessions, as well as a longer overall length of stay during residential treatment were shared protective factors for patients with a primary diagnosis of alcohol use disorder (AUD) or opioid use disorder (OUD), while greater withdrawal craving and a problematic recovery environment were common risk factors. We also proposed a novel approach and demonstrated preliminary success in predicting one-year readmission with multivariate, longitudinal treatment data using deep neural networks.

Metrics

140 File views/ downloads
173 Record Views

Details

Logo image