Degree Name

Department

Mathematics

College

Physical and Mathematical Sciences

Defense Date

2025-02-27

Publication Date

2025-03-14

First Faculty Advisor

Zachary Boyd

First Faculty Reader

David Erekson

Honors Coordinator

Davi Obata

Keywords

Machine learning, time series forecasting, distress, psychotherapy, prediction

Abstract

The need for quality mental health care is great, but many clinics face limited resources as they work to help clients. Predictions of a client’s future mental state during a course of therapy have the potential to inform decisions by individual therapists and counseling clinics. In this paper, we use archival OQ-45 data from a large university counseling center to examine the ability of many simple, rule-of-thumb predictive methods, as well as more complex methods (ridge regression, random forests, and LSTMs), to predict the next value in a time series of distress scores based on the previous scores. Learned inference from data was more accurate than rules of thumb with 95% credibility, with random forests being the most effective predictors (MAE ~8.62 OQ), although the other models performed nearly as well. Most recent distress score was a significant factor in all of the most successful models. ML and baseline performance was also compared to that of practicing psychotherapists at the same task, with therapists found to be comparable to ML models and the most accurate baselines.