AWS • PyTorch • Serverless inference
Predicting premature dropout from MOUD treatment
An interactive research demonstration of a hard-sharing multi-task neural network with state-specific prediction heads, trained on 2015–2019 TEDS-D treatment episodes.
Interactive demo
Run a single treatment-episode prediction
Research use only.
This demo is not clinical decision support and must not be used to make treatment decisions.
The model reflects historical administrative data and does not establish causal effects or an individual's prognosis.
Model design
Shared representation, state-specific heads
The PyTorch model uses a shared multilayer perceptron to learn common structure across states, followed by a separate linear prediction head for each of the 20 highest-volume states in the cohort.
25 predictors→
128→
64→
32→
state head
Frozen evaluation
- Test ROC-AUC
- 0.710339
- Test F1 @ 0.50
- 0.721275
- Macro state ROC-AUC
- 0.622735
- Training years
- 2015–2019
Interpretation
Important limitations
Episodes, not unique people
TEDS records represent treatment episodes; the same person can appear in more than one episode.
Not causal
The model estimates statistical associations in historical administrative data. It does not identify causes of dropout.
State-specific performance varies
Overall ROC-AUC does not imply equal performance in every state. State-level discrimination differs materially.
Research demonstration
This website is an engineering and research portfolio project, not a medical device or clinical recommendation system.