Predicting Substance Use Treatment Failure with Transfer Learning.

Area of Science:

• Machine Learning
• Data Science
• Addiction Research

Background:

• Transfer learning offers potential for improved predictions in substance use data, especially with limited clinical datasets.
• Repurposing trained models on related tasks can enhance predictive performance compared to models trained solely on target data.

Purpose of the Study:

• To evaluate a transfer learning method for classifying substance use treatment success.
• To determine if model weights transfer effectively across related substances and from large to small datasets.

Main Methods:

• A convolutional neural network was trained on a large heroin use treatment dataset.
• The trained model was then adapted and tested on a smaller opioid use treatment dataset.
• Performance was compared against a baseline model and a tuned random forest (RF) model.

Main Results:

• The transfer learning model demonstrated superior performance compared to both the RF and baseline models.
• Findings indicate successful transfer of model weights across related substances and from large to small datasets.
• The transfer model achieved a higher predictive score than the RF model.

Conclusions:

• Leveraging large datasets through transfer learning is an effective strategy for predicting substance use disorder treatment outcomes.
• Transfer learning presents a promising approach for enhancing the accuracy of SUD treatment success prediction models.
• This methodology offers a viable alternative to traditional modeling techniques for analyzing clinical substance use data.