Skip to main content
Talk

Neurosymbolic Programming in Mental Health

“Neurosymbolic Programming in Mental Health” was presented by Morgan Talbot and Dr. Omar Costilla Reyes on November 1, 2023. The session was part of the Neurosymbolic Programming Reading Group.

Interpretability discussion session at MIT CSAIL, May 2023. Related photograph from Interpretability Discussion — Neurosymbolic Programming Reading Group. Credit: Image courtesy of Dr. Omar Costilla Reyes.
Methods graphic from the team’s December 2023 announcement: self-reports and passively recorded sensor measures are filtered by behavior and compared as networks in an n-of-1 analysis. Related photograph from Individual behavioral insights in schizophrenia: A network analysis and mobile sensing approach. Credit: Image courtesy of Dr. Omar Costilla Reyes.
Figure 1 shared with the team’s June 2023 announcement: one participant’s reported symptoms decrease from November 2015 and then gradually improve, and the authors focus on the clinical interventions that precede such decreases. Related photograph from Counterfactual explanations and predictive models to enhance clinical decision-making in schizophrenia using digital phenotyping. Credit: Image courtesy of Dr. Omar Costilla Reyes.

A session on mental-health research

On November 1, 2023, Morgan Talbot and Dr. Omar Costilla Reyes, both of MIT, presented “Neurosymbolic Programming in Mental Health” to the Neurosymbolic Programming Reading Group, the weekly paper-discussion series of the National Science Foundation–funded “Understanding the World Through Code” project. Their abstract began from an overburdened mental-healthcare system with a compelling need for new evidence-based approaches to diagnosis, prognosis, and treatment, and from the observation that machine learning offers vast potential for clinical applications in mental health while facing significant challenges.

Three directions and their obstacles

The session surveyed three directions. Digital phenotyping uses data from personal digital devices, such as phone sensors and brief repeated self-reports of symptoms (ecological momentary assessments), to predict mental states. Network analysis quantifies how symptoms and other factors relate to one another. Counterfactual analysis asks what would have changed had a factor been different; it aims to clarify the causal relationships that drive mental health and to identify appropriate treatments for individual patients. The abstract also named the obstacles: limited large-scale datasets, noise and missing data in longitudinal records from patients’ smartphones and symptom self-reports, and the complexity of the many inter-related processes in patients’ lives.

Program synthesis, which searches for readable programs or equations that explain data, and neurosymbolic methods, which pair such explicit structure with learning, were explored as research directions under those conditions. Because the models stay readable, their reasoning can be examined by the clinicians and researchers who would rely on them. Together they point to a route for bringing programs that clinicians can read into mental-health research.

Grounding in his research

Related work includes Dr. Costilla Reyes’s digital mental-health studies “Counterfactual explanations and predictive models to enhance clinical decision-making in schizophrenia using digital phenotyping” (2023) and “Individual behavioral insights in schizophrenia: A network analysis and mobile sensing approach” (2024), and his interpretable machine-learning work on treatment response in body dysmorphic disorder with Massachusetts General Hospital. It was the third reading-group session he led in 2023, after an interpretability discussion in May and the “Neurosym Introduction” that opened the fall series on September 6.

Topics