The goal of individualized medicine is to determine which therapy works best for each individual patient. Predicting optimal treatments for individuals involves collecting large data sets of potential predictive variables, each with small non-linear effects and complex interactions. Machine learning methods can generate predictions from these vast datasets more efficiently than humans can.

Computers can also calculate Personalized Advantage Indexes (PAIs), which compare the likelihood of a successful outcome from one treatment to that of an alternative treatment. PAIs can then be used to guide physicians to offer patients their optimal treatment. 

Meyers et al. [Behaviour Research and Therapy] developed and tested a treatment-matching algorithm for veterans at high risk for suicide, aiming to predict which veterans would benefit more from Mindfulness-Based Cognitive Therapy for Suicide Prevention (MBCT-S) or enhanced treatment-as-usual (TAU).

The researchers performed a secondary analysis of data from a 2021 study in which 140 veterans at high risk for suicide were randomly assigned to either MBCT-S or enhanced TAU within the Veterans Administration (VA) system. The primary outcome measured was the number of suicide attempts or hospitalizations and emergency room visits for severe suicidal ideation over the ensuing 12-month period. The study collected a vast array of data that could be used as potential outcome predictors.

The researchers aimed to: 1) develop machine learning models to predict outcomes within each study arm, 2) identify key variables that predicted suicidal behavior, and 3) generate PAIs for each patient and evaluate their utility. 

MBCT-S was delivered in eight group and two individual sessions. TAU consisted of usual VA care (including access to psychopharmacology, psychotherapy, and residential care as needed) and attention from suicide prevention coordinators, who helped patients develop safety plans, encouraged compliance, and monitored progress.

The researchers selected  55 potential predictor variables from the demographic, clinical, and neurocognitive study data and the electronic medical records. Data were processed using a machine learning “random forest” approach. Data from 80% of the patients were used to train the predictive models, while data from the remaining 20% were reserved for subsequent model validation.

The results showed that the suicide prediction model was 73% accurate in the training sample and 67% accurate in the validation sample for the MBCT-S group. The predictive model for the TAU group was 66% accurate in the training sample and 60% accurate in the validation sample. The MBCT-S predictive model met the researcher’s standards of acceptability, but the TAU predictive model did not.

Within the MBCT-S group, the variables that best predicted future suicidal behavior were a diagnosis of PTSD, history of parasuicidal behavior, residential care in the past year, number of acute psychiatric admissions in the past year, and poor performance on a sustained attention task. In the TAU group, the best predictors were the number of acute hospitalizations and outpatient visits in the past year, severity of suicidal ideation, and better attentional control.

PAIs indicated that 63% of MBCT-S patients were in their PAI-indicated optimal treatment group compared to 39% of the TAU patients. Patients in their PAI-indicated optimal treatment were significantly less likely to have a suicidal event in the next year. While the main effect of treatment assignment on suicidal events was not significant, the interaction between treatment assignment and PAI indication was significant.

The findings support the use of large dataset machine learning to generate predictions about which patients may benefit most from particular therapies. It serves as a demonstration of what may be possible in the future, though creating a truly predictive model requires larger patient datasets and cross-validation on independent samples. The study was also limited by the use of a TAU control, which did not offer a uniform treatment experience to all participants.

Reference:

Myers, C. E., Dave, C. V., Chesin, M. S., Marx, B. P., St. Hill, L. M., Reddy, V., Miller, R. B., King, A., & Interian, A. (2024). Initial evaluation of a personalized advantage index to determine which individuals may benefit from mindfulness-based cognitive therapy for suicide prevention. Behaviour Research and Therapy. 

Link to study