Counterfactual explanations and predictive models to enhance clinical decision-making in schizophrenia using digital phenotyping

A 2023 preprint tests whether phone-sensor data can forecast changes in schizophrenia symptoms and explain them with counterfactual explanations. It is evaluated in a simulated monitoring setting.

Posted to arXiv in June 2023 by Juan Sebastián Cañas, Francisco Gómez, and Omar Costilla-Reyes, the paper developed from Cañas’s master’s research at the Universidad Nacional de Colombia, co-advised by Gómez and Dr. Omar Costilla Reyes. That research received the Best Student Paper Award at MAPI 2 in June 2022. Dr. Costilla Reyes announced the paper as his research team’s first on counterfactual explanations, applied to mobile phone sensor data from people with schizophrenia.

Digital phenotyping means using the data a person’s phone records during daily life, such as movement and activity, to describe behavior and health over time. Here, behavioral measures collected this way feed a three-part pipeline: forecasting symptom changes, detecting shifts with changepoint algorithms, which flag moments when a measured series changes level, and generating counterfactual explanations. A counterfactual explanation asks what would have had to be different for a model to give another answer, so that a clinician can see which inputs shaped a prediction.

In the reported evaluation, forecasts of individual symptom changes had a low error rate, and changepoint methods detected symptom decreases. The authors saw the explanations as a way to move beyond predictive models toward meaningful accounts of the daily experiences of people with schizophrenia, and toward insights that could inform personalized treatment rather than benchmark metrics alone. The arXiv abstract describes the pipeline as a basis for a future decision-support system that combines clinical assessments with sensor-based inputs.

The evaluation ran the pipeline as a simulated continuous-monitoring workflow, which mirrors how such a tool would follow a person’s data as it arrives over time. Explanations matter in this setting because a forecast that a clinician cannot interrogate is hard to act on, while a counterfactual shows which behaviors would have to change for the forecast to change. The authors released the preprint to invite feedback from the scientific community. The same research line continues in the individual-level mobile-sensing study of schizophrenia presented at PervasiveHealth 2023.