Optimising mental health outreach for ex-inmates

In this project, I worked with the Johnson County Mental Health Services in Kansas to develop a machine learning pipeline that would optimise their outreach to people at risk of returning to jail. We received de-identified data from mental health services, court trials, jail bookings, and emergency services. Using best practices on dealing with the temporal auto-correlation of the data and bias and fairness issues, we were able to develop a tool that could be deployed monthly to find the top 100 ex-inmates most likely to return to prison within the next year. Our best-performing model, a Random Forest Classifier, had a precision of 61.5%. (Outperformed baseline of 21%.)

My main contribution was to help the analytical and computational progress of the team. I used “Triage”, a tool developed by the Center for Data Science for Social Good (DSSG), to do feature engineering, model training, and testing. I also drafted a field trial proposal for the deployment of the model.

Overseen by Dr. Rayid Ghani.