Inducing Causal Structure for Interpretable Neural Networks Applied to Glucose Prediction for T1DM Patients

TL;DR

This work investigates imposing causal structure on neural networks for blood glucose forecasting in Type 1 Diabetes via Interchange Intervention Training (IIT). By adapting the FDA-approved simglucose model into a time-compressed, acyclic causal graph and training an MLP with IIT, the authors demonstrate improved predictive accuracy and enhanced interpretability through counterfactual loss analysis. Key findings show IIT reduces prediction error across short horizons and provides module-level insight into which causal mechanisms are learned well, though gains diminish at longer horizons due to unmodeled factors. The study suggests IIT can yield interpretable, real-time glucose predictions suitable for on-device deployment, while highlighting avenues for future work, including temporal extensions and validation on real patient data.

Abstract

Causal abstraction techniques such as Interchange Intervention Training (IIT) have been proposed to infuse neural network with expert knowledge encoded in causal models, but their application to real-world problems remains limited. This article explores the application of IIT in predicting blood glucose levels in Type 1 Diabetes Mellitus (T1DM) patients. The study utilizes an acyclic version of the simglucose simulator approved by the FDA to train a Multi-Layer Perceptron (MLP) model, employing IIT to impose causal relationships. Results show that the model trained with IIT effectively abstracted the causal structure and outperformed the standard one in terms of predictive performance across different prediction horizons (PHs) post-meal. Furthermore, the breakdown of the counterfactual loss can be leveraged to explain which part of the causal mechanism are more or less effectively captured by the model. These preliminary results suggest the potential of IIT in enhancing predictive models in healthcare by effectively complying with expert knowledge.

Figures (10)

• Figure 1: Simglucose specific kinetic constants and details on the model
• Figure 2: NN architecture diagram for (a) MLP parallel, (b) MLP tree, and (c) MLP joint
• Figure 3: RMSE (mg/dL) prediction error of the model MLP tree (256) IIT using amended simglucose across the four prediction horizons (PH) for the test (n=30) in-silico T1DM patients and 10 different random seeds. The solid bars refer to IIT trainings while the striped bars refer to standard training. The mean is indicated by the cross.
• Figure 4: $L_{INT}$ tracking for the MLP tree model for PH 30. The $L_{INT}$ is grouped by modules. a) during training and b) during testing.
• Figure 5: RMSE (mg/dL) prediction error of the different MLP architectures (4 models) across the four PHs for the test (n=30) in-silico T1DM patients. The causal model used is the regular simglucose and the boxplots are produced with 10 different random seed. The solid bars refer to the models being trained through interchange intervention training while the striped bars refer to conventional training; without IIT.