# Incorporating Uncertainty Estimation and Interpretability in Personalized Glucose Prediction Using the Temporal Fusion Transformer

## TL;DR

This paper introduces a new AI method for predicting blood glucose levels in diabetes patients that is more accurate, interpretable, and includes uncertainty estimation.

## Contribution

The novel contribution is applying the Temporal Fusion Transformer with uncertainty estimation and interpretability for personalized glucose prediction.

## Key findings

- The Temporal Fusion Transformer outperformed existing methods in glucose prediction accuracy by at least 13% in RMSE.
- The model's interpretability was assessed through feature importance and attention mechanisms.
- The approach was tested on two datasets, showing consistent performance improvements.

## Abstract

More than 14% of the world’s population suffered from diabetes mellitus in 2022. This metabolic condition is defined by increased blood glucose concentrations. Among the different types of diabetes, type 1 diabetes, caused by a lack of insulin secretion, is particularly challenging to treat. In this regard, automatic glucose level estimation implements Continuous Glucose Monitoring (CGM) devices, showing positive therapeutic outcomes. AI-based glucose prediction has commonly followed a deterministic approach, usually with a lack of interpretability. Therefore, these AI-based methods do not provide enough information in critical decision-making scenarios, like in the medical field. This work intends to provide accurate, interpretable, and personalized glucose prediction using the Temporal Fusion Transformer (TFT), and also includes an uncertainty estimation. The TFT was trained using two databases, an in-house-collected dataset and the OhioT1DM dataset, commonly used for glucose forecasting benchmarking. For both datasets, the set of input features to train the model was varied to assess their impact on model interpretability and prediction performance. Models were evaluated using common prediction metrics, diabetes-specific metrics, uncertainty estimation, and interpretability of the model, including feature importance and attention. The obtained results showed that TFT outperforms existing methods in terms of RMSE by at least 13% for both datasets.