AI Ensemble Predicts Blood Glucose in Type 1 Diabetes with Minimal Data Cleanup

For the millions of people living with Type 1 Diabetes, every day is a tightrope walk between too much glucose and too little. Insulin dosing decisions hinge on knowing where blood sugar is heading, not just where it is now, and even a half-hour of foresight can mean the difference between a smooth correction and a dangerous hypoglycemic episode. A new study published in Neural Computing and Applications by Doaa Mahmoud Abdelaziz, Amr M. Refaat and Sayed T. Muhammad of Fayoum University in Egypt takes a deliberately pragmatic approach to this prediction problem, showing that competitive forecasts of future blood glucose can be achieved with surprisingly light data preparation, and that explainable artificial intelligence can reveal which signals matter most for each individual patient.

The research team built a patient-specific prediction framework and evaluated it on the OhioT1DM dataset, a widely used benchmark collected from people with Type 1 Diabetes wearing continuous glucose monitors. What sets the study apart is its commitment to minimal preprocessing. Many machine learning pipelines in diabetes research rely on elaborate cleaning, imputation and feature engineering stages that can be fragile in real-world deployment, where sensor dropouts and irregular meal logging are the norm. By keeping data preparation light, the Egyptian team aimed to test how well modern machine learning models perform when fed something much closer to the raw stream of clinical measurements that a real monitoring system would produce.

The experimental design was systematic. The researchers compared several machine learning models using standard regression metrics: Mean Absolute Error, which captures the average size of prediction mistakes; Root Mean Square Error, which penalizes large errors more heavily; and the coefficient of determination, or R², which measures how much of the variability in future glucose values the model actually explains. Crucially, they did not stop at statistical scores. They also applied Clarke Error Grid analysis, a clinical assessment tool developed in 1987 that judges predictions the way clinicians would, by asking whether a forecast error would lead to a harmless decision or a dangerous one. A model can look numerically decent while still making the kind of mistake that would trigger a wrong insulin dose, and the Clarke Error Grid is designed to catch exactly that.

Feature selection was explored through three combinations drawn from the literature. The simplest used blood glucose readings alone, testing how far a model can get on the glucose time series itself. The second added insulin-on-board, an estimate of how much previously injected insulin is still actively working in the body, together with meal information, since carbohydrates are the dominant driver of glucose excursions. The third configuration added physical activity data on top of everything else, acknowledging that exercise can send glucose plummeting or, in some cases, spiking, in ways that neither insulin nor food history fully explains. Each configuration was tested with a sliding input window of five historical time steps sampled at five-minute intervals, meaning the models looked back twenty-five minutes into each patient’s recent metabolic past.

Two prediction horizons were evaluated: thirty minutes and sixty minutes ahead. This choice reflects the practical realities of diabetes management. A thirty-minute forecast gives enough lead time for a pre-emptive snack or a delayed bolus, while a sixty-minute forecast is considerably harder because the effects of meals, insulin and activity compound over time, but it offers a much larger safety margin. The gap in difficulty between the two horizons is one of the central challenges in the field, and any framework that maintains accuracy at the longer horizon is doing genuinely useful work.

The centerpiece of the study is the Tri-Ensemble model, constructed by combining the predictions of the three strongest benchmarked models using simple averaging. Ensembling is a well-established strategy in machine learning: when individually strong models make somewhat different errors, averaging their outputs can cancel out idiosyncratic mistakes and produce a more robust overall prediction. The authors describe this aggregation choice as robustness-oriented, and the results supported that framing. The Tri-Ensemble demonstrated competitive predictive performance across multiple feature configurations and both prediction horizons, with statistically significant improvements observed in several of the experimental settings. In other words, the gains were not just artifacts of chance but held up under statistical scrutiny.

Beyond raw accuracy, the study contributes an interpretability layer. The researchers conducted an exploratory explainable AI analysis using SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each input feature a contribution value for every individual prediction. Rather than treating the model as a black box that spits out a glucose number, SHAP lets researchers open the box and see which variables pushed each forecast up or down, and by how much. The team applied this analysis over the complete available feature space, going beyond the three curated feature combinations used in the benchmarking experiments.

The SHAP analysis was performed patient by patient, and this patient-wise perspective is one of the study’s most interesting aspects. Type 1 Diabetes is famously heterogeneous: two people with the same diagnosis can have very different insulin sensitivities, eating patterns, activity levels and glucose dynamics. A feature that dominates predictions for one patient, such as insulin-on-board, may play a secondary role for another whose glucose is driven more by activity or meal timing. By examining feature importance patterns for each individual, the authors aimed to identify potentially informative variables that could guide feature engineering in future blood glucose prediction studies, effectively using the explainability tool as a discovery instrument rather than merely a compliance checkbox.

The significance of this work lies partly in its restraint. Rather than proposing an ever-larger deep learning architecture trained on heavily curated data, the study asks a more deployable question: how well can we predict glucose with modest models, light preprocessing and honest clinical evaluation? The answer, supported by the OhioT1DM benchmark and Clarke Error Grid analysis, is encouraging. A simple averaged ensemble of the strongest learners, given just twenty-five minutes of history and a handful of clinically meaningful signals, can deliver forecasts that are both statistically and clinically competitive. That combination of simplicity and rigor matters for anyone hoping to move glucose prediction out of the lab and into the pumps, pens and smartphone apps that patients actually use.

The road ahead, as the authors frame it, involves using the SHAP-derived feature relevance patterns to design smarter, personalized feature sets, potentially letting each patient’s model emphasize the signals that matter most for their own physiology. As continuous glucose monitoring spreads globally and the International Diabetes Federation projects continued growth in diabetes prevalence, tools that forecast blood glucose accurately, transparently and with minimal data fuss could become a quiet but transformative part of daily care. This study is a reminder that in medical machine learning, sometimes the most valuable innovation is not a bigger model, but a clearer view of what the model is actually seeing.

Subject of Research: Machine learning prediction of blood glucose in Type 1 Diabetes using the OhioT1DM dataset with SHAP-based explainability

Article Title: Blood glucose prediction in Type 1 Diabetes under minimal preprocessing using machine learning and exploratory SHAP analysis

Article References: Abdelaziz, D. M., Refaat, A. M., & Muhammad, S. T. (2026). Blood glucose prediction in Type 1 Diabetes under minimal preprocessing using machine learning and exploratory SHAP analysis. Neural Computing and Applications, 38(17), Article 706. https://doi.org/10.1007/s00521-026-12330-6

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12330-6

Keywords: Type 1 Diabetes, blood glucose prediction, machine learning, OhioT1DM dataset, Tri-Ensemble, SHAP, explainable AI, Clarke Error Grid, insulin-on-board, continuous glucose monitoring, patient-specific modeling, Neural Computing and Applications