AMASS 1 - Harnessing Innovative Precision Medicine Methods to Improve Digital Mental Health Interventions: Introduction to Applied Examples and Tutorials

Perennial clinical science questions to inform treatment targets and targeted treatments involve identifying risk factors and prescriptive predictors. Multivariate models are central to these questions. However, typical general linear models are limited in their ability to test intricate interactions among variables with diverse distributional patterns, identify relevant predictors for specific subgroups, and deduce complex relations (Shatte et al., 2019). To tackle these challenges, the ascent of machine learning (ML) is evident as it develops algorithms capable of categorizing persons with a wide array of unevenly distributed risk factors (Dwyer & Krishnadas, 2022). Predictive ML uses algorithms to gauge the predictive strength of multiple variables for future unseen data. Employing flexible non-linear and higher-order interaction ML algorithms can enhance model development and optimize predictive accuracy (Deisenhofer et al., 2023). Further, the Shapley additive explanations (SHAP) method helps to intuitively visualize predictor-outcome relations (Lundberg & Lee, 2017) while adjusting for other predictors (Molnar, 2022). However, precision medicine techniques are often overlooked in psychology.

This workshop thus introduces ML methods with digital health. First, we examine the clinical uses of ML. In the process, we provide examples of how ML has been used to identify who benefits most from internet-delivered cognitive-behavioral therapy vs. treatment-as-usual in a randomized controlled trial among undergraduates with anxiety or depression from Colombia and Mexico (Benjet et al., 2023) and who engages optimally with i-CBT. Second, we will show how to conduct nested cross-validated ML analysis. Examples include performing predictive ML analyses with nestedcv (Lewis et al., 2023) and visualizing predictor-outcome relations in a multivariate model using the fastshap R package (Greenwell, 2021). Third, we discuss considerations of using predictive models to develop clinical decision support tools in healthcare systems, with real-world examples of how predictive ML models can be applied to healthcare data records to predict relapse after substance use treatment. Finally, we discuss the ethics of using ML on clinical data.