Explaining predictions in digital mental-health studies
Dr. Costilla Reyes’s work examines prediction and explanation across several research settings: treatment response in body dysmorphic disorder, outcomes from smartphone-delivered interventions, and behavioral patterns studied with mobile sensing in schizophrenia. The methods include interpretable models, which a researcher can read and question; feature analysis, which ranks the measurements a model relies on; counterfactual explanations, which show how an input would have to change for a model’s answer to change; and individual-level network analysis, which maps how one person’s reported experiences relate to each other.
These projects ask different questions. Some estimate treatment response from clinical or self-reported measures; others use smartphone data to study behavior over time. Reading them together shows how the study design and available data shape which predictions can be made and which explanations are useful to review.
What the published analyses show
In a secondary analysis of a trial with several hundred participants across three smartphone-delivered interventions for major depressive disorder, higher baseline anxiety scores were associated with lower odds of improvement. In body dysmorphic disorder studies, treatment credibility emerged among the strongest predictors of response, and decision-tree models exposed thresholds that researchers could inspect.
Mobile-sensing studies have also examined how patterns in location and other behavioral measures relate to psychological-network measures, linking what a phone records with what a person reports. Each study pairs its participants, intervention, and measurements with a matching evaluation method, so the explanations stay anchored to the data they describe.
Interpretable models for digital mental-health research
Feature attributions, counterfactual explanations, and individual-level network analysis give researchers tools to examine model outputs and patterns in digital mental-health studies. The publications apply these methods to defined participant groups, measures, and research questions, and each publication page describes its methods and findings.
Across this work, interpretable modeling connects predictive analysis with domain expertise, making the factors and patterns considered in each study available for examination.