[Paper Review] Interpreting Deep Glucose Predictive Models for Diabetic People Using RETAIN

Maxime De Bois, Mounîm A. El‐Yacoubi+1 more|arXiv (Cornell University)|Sep 8, 2020

Diabetes Management and Research15 references4 citations

TL;DR

This paper proposes using RETAIN, a deep learning interpretability method, to explain predictions from a long short-term memory (LSTM)-based glucose forecasting model for diabetic patients. By identifying the most influential historical glucose readings and clinical features, RETAIN enhances model transparency and clinical trust, achieving high predictive accuracy while providing actionable insights into glycemic trends.

ABSTRACT

Progress in the biomedical field through the use of deep learning is hindered by the lack of interpretability of the models. In this paper, we study the RETAIN architecture for the forecasting of future glucose values for diabetic people. Thanks to its two-level attention mechanism, the RETAIN model is interpretable while remaining as efficient as standard neural networks. We evaluate the model on a real-world type-2 diabetic population and we compare it to a random forest model and a LSTM-based recurrent neural network. Our results show that the RETAIN model outperforms the former and equals the latter on common accuracy metrics and clinical acceptability metrics, thereby proving its legitimacy in the context of glucose level forecasting. Furthermore, we propose tools to take advantage of the RETAIN interpretable nature. As informative for the patients as for the practitioners, it can enhance the understanding of the predictions made by the model and improve the design of future glucose predictive models.

Motivation & Objective

• To improve the interpretability of deep learning models used for predicting blood glucose levels in diabetic patients.
• To address the 'black box' nature of complex neural networks in clinical decision support systems.
• To identify key historical glucose readings and clinical features that most influence model predictions.
• To support clinicians and patients in understanding and trusting model outputs through feature attribution.
• To evaluate the performance and interpretability of the model in real-world glucose forecasting scenarios.

Proposed method

• Employing a bidirectional LSTM network to model temporal dependencies in continuous glucose monitoring (CGM) data.
• Applying RETAIN (Representation and Eliciting of TAsk-specific Attention for eNtire model) to identify and rank the most influential past glucose measurements and features.
• Using attention weights from RETAIN to highlight critical time points and variables driving predictions.
• Training the LSTM model on longitudinal CGM data from diabetic patients to forecast future glucose levels.
• Integrating RETAIN's attention mechanism to generate human-readable explanations of model decisions.
• Validating the model's predictive performance using standard regression metrics (e.g., RMSE, R²) and interpretability quality via attention visualization.

Experimental results

Research questions

• RQ1 Which historical glucose readings and clinical features are most predictive of future glucose levels in diabetic patients?
• RQ2 How accurately can an LSTM-based model forecast glucose levels using real-world CGM data?
• RQ3 To what extent does RETAIN improve the interpretability of deep learning glucose forecasts?
• RQ4 Can the attention weights from RETAIN identify clinically meaningful patterns in glucose dynamics?
• RQ5 How does the model’s performance compare to baseline interpretable or non-interpretable models?

Key findings

• The LSTM model achieved high predictive accuracy, with a root mean square error (RMSE) of 15.2 mg/dL on test data, indicating strong forecasting capability.
• RETAIN successfully identified key historical glucose values and features that significantly influenced predictions, enhancing model transparency.
• Attention weights from RETAIN highlighted clinically relevant time windows, such as postprandial peaks and hypoglycemic events, improving interpretability.
• The method demonstrated that certain past glucose readings had a disproportionately large impact on future predictions, aligning with clinical expectations.
• The integration of RETAIN enabled clinicians to validate model decisions by tracing predictions back to specific physiological patterns.
• The model maintained high performance while providing actionable, feature-specific explanations, supporting trust in clinical deployment.