Using machine learning to match clients and therapy providers: Evaluating clinical quality and cost of care

Jennifer L Lee; Chris Billovits; Shih-Yin Chen; Robert E Wickham; Bob Kocher; Connie E Chen; Anita Lungu

doi:10.1016/j.jval.2025.06.002

Using machine learning to match clients and therapy providers: Evaluating clinical quality and cost of care

Value Health. 2025 Jun 26:S1098-3015(25)02405-2. doi: 10.1016/j.jval.2025.06.002. Online ahead of print.

Authors

Jennifer L Lee¹, Chris Billovits², Shih-Yin Chen², Robert E Wickham³, Bob Kocher⁴, Connie E Chen², Anita Lungu²

Affiliations

¹ Lyra Health, Burlingame, CA, USA; Adjunct Assistant Professor, Emory University, Atlanta, GA, USA. Electronic address: lee.jennifer.phd@gmail.com.
² Lyra Health, Burlingame, CA, USA.
³ Associate Professor, Northern Arizona University, Flagstaff, AZ, USA.
⁴ Adjunct Professor, Stanford University, Palo Alto, CA, USA; Lyra Clinical Associates, Burlingame, CA, USA.

PMID: 40581153
DOI: 10.1016/j.jval.2025.06.002

Abstract

Objectives: Matching clients in need of mental health care with providers who will deliver high quality treatment presents a substantial challenge. Machine learning models hold potential for predicting the best pairings from a multitude of data points, leveraging relevant characteristics to recommend providers.

Methods: Propensity score matching was used to match individuals who searched for a psychotherapy providers using either a Pragmatic algorithm (leveraging logistical and clinical relevance features) or a Value-based algorithm (adding provider-specific clinical outcomes and cost features). Post-matching cohorts included on average 1677 pairs with clinically elevated symptoms of anxiety. Symptom improvement from pre- to post-treatment was calculated. Total costs of care were compared between algorithm cohorts.

Results: After matching, participants were on average 34 years of age, 54-55% White, and 63-66% female. Mean level of anxiety symptom change from pre- to post-treatment was statistically significant for both groups (Pragmatic: -5.82; Value-based: -5.57, Ps < .001) with large effect sizes. People searching for therapy providers with either algorithm had similar rates of reliable improvement or recovery in anxiety (Pragmatic: 71.74%, Value-based: 70.02%). Participants using the Value-based care algorithm group had 20% lower total cost of care, using 2.08 fewer therapy sessions. Depression outcomes were similar to those for anxiety, thus are presented in the supplement.

Conclusions: Results indicate that a Value-based machine learning matching algorithm integrating historical provider performance and cost metrics may result in better provider-client pairings that reduce the total cost of care with no effect on outcomes. Further research is needed to establish the generalizability of these algorithms.

Keywords: anxiety; depression; machine learning; mental health; value-based care.