Towards Personalized Allocation of Patients to Therapists

Objective: Psychotherapy outcomes vary between therapists, but it is unclear how such information can be used for treatment planning or practice development. This proof-of-concept study aimed to develop a data-driven method to match patients to therapists. Method: We analyzed data from N = 4,849 patients who accessed cognitive-behavioral therapy in U.K. primary care services. The main outcome was posttreatment reliable and clinically significant improvement (RCSI) on the Patient Health Questionnaire-9 (PHQ-9) depression measure. Machine-learning analyses were applied in a training sample (N = 2,425 patients treated by 68 therapists in Year 1), including a chi-squared automatic interaction detector (CHAID) algorithm and a random forest (RF) algorithm. The predictive models were cross-validated in a statistically independent test sample (N = 2,424 patients treated by the same therapists in Year 2) and evaluated using odds ratios (ORs) adjusted for baseline depression severity. Results: We identified subgroups of therapists that were differentially effective for highly specific subgroups of patients, yielding 17 classes of patient-to-therapist matches. The overall base rate of RCSI in the sample was 40.4%, but this varied from 10.5% to 69.9% across classes. Cases classed by the prediction algorithms as expected responders in the test sample were ∼60% more likely to attain posttreatment RCSI compared with those classed as nonresponders (adjusted ORs = 1.59, 1.60; p < .001). Conclusions: Machine-learning approaches could help to improve treatment outcomes by enabling the strategic allocation of patients to therapists and therapists to supervisors. What is the public health significance of this article? It is well known that, even when they apply the same treatment model, some therapists attain better clinical outcomes compared with others. Using data from a large (N = 4,849) naturalistic cohort of patients who accessed highly standardized cognitive-behavioral therapy for common mental disorders, the present article shows that specific therapists are more or less able to help specific subgroups of patients. We developed machine-learning algorithms that are able to pinpoint the profiles of patients that could be matched to specific therapists in order to improve treatment outcomes. An additional possibility is to use this model to support practice development by matching therapists to peer supervisors who evidently attain better outcomes with specific profiles of