Predicting who drops out of behavioral health treatment
A state health department was spending heavily on beds and still losing patients from treatment. Clinical data couldn't predict who would drop out, so I added census data, built a risk model in Python, and paired it with recommended next steps.
- Organization
- EY, for a state health department
- My role
- Staff Technology Consultant, Data and Platform Analysis
- When
- 2021 to 2023
- Result
- Non-clinical factors drove about 60% of readmissions; the state shifted investment toward housing and transportation support
The problem
The department was spending a lot on hospital beds and still seeing poor outcomes, because patients kept dropping out of behavioral health treatment and nobody could predict who. Clinical data on its own wasn’t explaining it.
What I did
I brought in US Census data and merged it with 12 years of patient history in Python, then built features around social determinants of health like housing instability and access to transportation.
A risk score alone doesn’t tell anyone what to do, so I paired the model with a next-best-action recommender. If a patient was flagged as high risk because of transportation, the recommendation was a rideshare voucher rather than another appointment.
Result
Non-clinical factors drove about 60% of readmissions. The state used that to shift investment from emergency beds toward housing and transportation support, and the model became a roadmap for future funding decisions.
What I took from it
The answer was outside the dataset everyone started with. Combining sources is often where the useful model comes from.