Predicting Future Service Use in Dutch Mental Healthcare: A Machine Learning Approach

Area of Science:

• Health Services Research
• Applied Machine Learning
• Mental Healthcare Management

Background:

• Equitable and efficient allocation of scarce mental healthcare resources is crucial.
• Predictive modeling for service use is vital for resource management but challenged by service use skewness.

Purpose of the Study:

• To apply a machine learning approach to forecast expected mental healthcare service use.
• To provide a foundation for agreements between mental healthcare financiers and suppliers.

Main Methods:

• Utilized administrative data from a large Dutch mental healthcare organization.
• Developed a baseline model and three random forest models using 2017 training data.
• Validated models on 2018 data (N=10,201) to predict individual treatment hours.

Main Results:

• The best random forest model achieved a mean error of 21 minutes at the insurance group level.
• An average absolute error of 28 hours was observed at the patient level.
• A systematic under-prediction of service use occurred for high-service users.

Conclusions:

• Machine learning is effective for predicting group-level mental healthcare service use.
• These predictive models can aid financiers and suppliers in resource planning and allocation.
• Accurate prediction for high-cost patients remains an area for future development.