Mealtime Prediction | Baiying Lu

Post-meal hyperglycemia is one of the most common failure modes in type 1 diabetes management, and it is usually caused by something mundane: an insulin dose that was missed, or taken too late. A system that knew when someone was about to eat could nudge them to bolus beforehand.

This project asked whether that is possible using data people already generate — the logs from their insulin pump — with no additional sensors, no manual input, and no burden on the user.

Approach

Using two independent datasets containing over 45,000 meal logs from 82 patients, we first characterized how regular real-world mealtime patterns actually are, then trained personalized LSTM models to predict upcoming meals for each individual.

Findings

Roughly 60% of participants had irregular and inconsistent mealtime patterns that shifted both across the day and across months of their own history — undermining the assumption that a population-level or fixed-schedule model would work.
Despite that irregularity, personalized models reached an average F1 score above 95% with fewer than 0.25 false positives per day, a false-alarm rate low enough to be tolerable in a real nudging system.

Together these results lay the groundwork for a meal prediction system that prompts patients to administer bolus insulin before eating, reducing post-meal highs (Lu et al., 2024).

References

2024

Mealtime prediction using wearable insulin pump data to support diabetes management

Baiying Lu, Yanjun Cui, Prajakta Belsare, and 3 more authors

Scientific Reports 2024 14:1, Sep 2024

Abs DOI

Many patients with diabetes struggle with post-meal high blood glucose due to missed or untimely meal-related insulin doses. To address this challenge, our research aims to: (1) study mealtime patterns in patients with type 1 diabetes using wearable insulin pump data, and (2) develop personalized models for predicting future mealtimes to support timely insulin dose administration. Using two independent datasets with over 45,000 meal logs from 82 patients with diabetes, we find that the majority of people ( $∼$ 60%) have irregular and inconsistent mealtime patterns that change notably through the course of each day and across months in their own historical data. We also show the feasibility of predicting future mealtimes with personalized LSTM-based models that achieve an average F1 score of > 95% with less than 0.25 false positives per day. Our research lays the groundwork for developing a meal prediction system that can nudge patients with diabetes to administer bolus insulin doses before meal consumption to reduce the occurrence of post-meal high blood glucose.