📄 medicine

Enhancing Explainability in AI Models for Blood Glucose Prediction in Diabetes Patients

This study demonstrates that machine learning models, particularly Support Vector Regression and tree-based ensembles, can accurately predict short-term blood glucose levels for both Type 1 and Type 2 diabetes patients while leveraging SHAP and LIME techniques to provide transparent, interpretable insights that support clinical decision-making.

Original authors: Lam Nhi Nguyen, Kwong-Cheong Wong

Published 2026-07-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Lam Nhi Nguyen, Kwong-Cheong Wong

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather for the next 30 minutes. You have a super-smart computer that looks at the current temperature, wind speed, and cloud cover to guess if it will rain. But here's the problem: the computer is a "black box." It gives you an answer, but it won't tell you why it thinks it's going to rain. If you are a pilot or a farmer, you might not trust that answer because you don't understand the reasoning.

This paper is about building a similar "weather forecast" for people with diabetes. Instead of rain, the computer predicts blood sugar levels 30 minutes into the future. The goal is to make these predictions not just accurate, but also explainable—so doctors and patients can trust them.

Here is a simple breakdown of what the researchers did and found:

1. The Goal: A Trustworthy Crystal Ball

The researchers wanted to build a machine learning model (a type of computer program) that could look at a diabetic patient's recent blood sugar history and guess what their sugar level will be in 30 minutes.

The Problem: Many smart computer models exist, but they are like a magician pulling a rabbit out of a hat. You see the rabbit (the prediction), but you don't know how the trick was done. Doctors are hesitant to use tools they can't understand.
The Solution: They added a special "translator" layer called Explainable AI (XAI). This layer acts like a tour guide, explaining exactly which clues the computer used to make its guess.

2. The Ingredients: Two Types of Patients

The study looked at two different groups of people, like two different types of gardens:

Type 1 Diabetes: A smaller group (12 people). Their blood sugar is like a garden with a very sensitive sprinkler system—it can swing wildly up and down very quickly.
Type 2 Diabetes: A larger group (100 people). Their blood sugar is more like a steady river, but it has more rocks and debris (like age, weight, and kidney function) affecting the flow.

They used data from Continuous Glucose Monitors (CGM), which are like tiny sensors that take a "temperature reading" of blood sugar every 15 minutes, 24 hours a day.

3. The Race: Five Different Predictors

The researchers built five different "forecasters" to see which one was best:

Linear Regression: A simple, straight-line guesser.
Random Forest: A committee of decision trees that vote on the answer.
XGBoost: A super-charged, fast-learning tree committee.
Support Vector Regression (SVR): A clever model that finds the best curve to fit the data.
LSTM: A deep learning model designed to remember long sequences (like a human trying to memorize a long story).

The Result: The "simple" and "committee" models (SVR, Random Forest, and XGBoost) won the race. They were the most accurate. The "deep learning" model (LSTM) actually did the worst.

Why? Think of the LSTM like a student trying to learn a complex language by reading a whole library, but they only have a single page of text to study. Because the data for Type 1 patients was so small, the complex model got confused and overthought things. The simpler models just looked at the most recent clues and did a great job.

4. The Translator: SHAP vs. LIME

Once they picked the best forecasters, they needed to explain how they worked. They used two different "translators":

SHAP (The Detailed Detective): This method is like a forensic accountant. It looks at every single clue, calculates exactly how much each one contributed to the final answer, and gives a mathematically perfect report. It is very stable and accurate, even when the blood sugar is swinging wildly.
- Analogy: It's like a high-definition map that shows every street, alley, and building. It's precise but takes a bit of time to read.
LIME (The Quick Sketch Artist): This method zooms in on just one specific prediction and draws a simple sketch to explain it. It's very easy to understand at a glance but can sometimes get the details wrong if the situation is messy.
- Analogy: It's like a quick napkin sketch. It gets the main idea across instantly, but if you look too closely, the lines might be wobbly or exaggerated.

The Comparison:

SHAP was the winner for accuracy and stability. It told the truth about what the computer was thinking, even in chaotic situations.
LIME was better for a quick, simple explanation, but it sometimes made the "clues" look bigger or more important than they really were, especially when the blood sugar was changing fast.

5. What Did They Learn?

The Most Important Clue: For both types of diabetes, the single most important thing the computer looked at was what the blood sugar was doing just a few minutes ago.
- If your sugar was rising 15 minutes ago, the computer predicts it will keep rising.
- Other factors like age, weight, or medication mattered, but they were like background noise compared to the immediate trend.
The Difference: In Type 1 patients, the computer relied almost only on the recent sugar readings because their levels change so fast. In Type 2 patients, the computer also paid a little attention to other health factors (like kidney function), but the recent sugar trend was still the boss.

The Bottom Line

The paper concludes that we can build computer models that predict blood sugar very accurately without needing to be "black boxes." By using tools like SHAP, we can see exactly why the computer made a guess.

For Doctors: This means they can trust the AI because they can see the logic behind it.
For the Future: The researchers suggest that to make these models even better, they need more data (especially for Type 1 patients) and should try to include more "real-life" clues like what the patient ate or how much they exercised, which weren't fully included in this specific study.

In short: Simple models work best, recent history is king, and SHAP is the best translator to help us understand the machine.

Technical Summary: Enhancing Explainability in AI Models for Blood Glucose Prediction in Diabetes Patients

Problem Statement
Accurate short-term blood glucose prediction is critical for managing diabetes and preventing adverse glycemic events. While machine learning (ML) models have demonstrated promising performance in forecasting glucose levels, their clinical adoption is hindered by a lack of transparency. The "black-box" nature of complex models creates barriers to trust and safety, particularly in life-critical scenarios. Furthermore, existing research often focuses predominantly on Type 1 diabetes (T1DM), leaving Type 2 diabetes (T2DM) underexplored despite its higher prevalence. There is a need to evaluate not only the predictive accuracy of ML models but also their interpretability across different diabetes populations to ensure clinical viability.

Methodology
The study employs a two-phase framework: Predictive Modelling and Model Explainability, utilizing data from the ShanghaiT1DM (12 subjects) and ShanghaiT2DM (100 subjects) datasets. These datasets include continuous glucose monitoring (CGM) time-series data (15-minute intervals) and static clinical features (demographics, biochemical measurements, medications).

Predictive Modelling:
- Data Preprocessing: Missing values were interpolated (variables with >30% missingness were excluded), and categorical data were binarized.
- Feature Selection: From an initial 93 variables, a two-stage selection process (variance threshold and mutual information-based SelectKBest) reduced the input to the 20 most relevant predictors.
- Model Development: Five distinct ML models were trained to predict glucose levels 30 minutes ahead (two 15-minute intervals):
  - Linear Regression (baseline)
  - Random Forest (ensemble)
  - XGBoost (gradient boosting)
  - Support Vector Regression (SVR, kernel-based)
  - Long Short-Term Memory (LSTM, deep learning sequence model)
- Evaluation: Performance was assessed using Root Mean Squared Error (RMSE) and the coefficient of determination ( $R^2$ ) on a 60/20/20 train-validation-test split.
Model Explainability (XAI):
- Techniques: The study applied SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations).
- Global Analysis: SHAP was used to aggregate feature importance across the entire dataset to identify dominant predictors.
- Local Analysis: SHAP and LIME were applied to specific test instances representing low-error (stable) and high-error (volatile) prediction scenarios for both T1DM and T2DM.
- Comparison Criteria: The explainability methods were evaluated based on five dimensions: fidelity, interpretability, completeness, stability, and clinical usability.

Key Contributions

Comparative Performance Analysis: The study provides a direct comparison of linear, ensemble, kernel-based, and deep learning models for 30-minute glucose forecasting in both T1DM and T2DM cohorts.
Dual-Population Insight: It addresses the gap in T2DM research by analyzing model behavior and feature importance specifically for Type 2 patients, contrasting them with Type 1 dynamics.
Systematic XAI Evaluation: Rather than simply applying XAI tools, the paper systematically compares SHAP and LIME across global and local contexts, highlighting their respective strengths and weaknesses in the specific domain of glucose prediction.
Physiological Validation: The study validates whether ML models rely on clinically plausible features (e.g., recent CGM trends vs. static clinical variables) and identifies population-specific differences in model reasoning.

Results

Predictive Performance:
- T1DM: SVR and Linear Regression achieved the best performance (RMSE ~20.7–20.8 mg/dL, $R^2$ ~0.90). The LSTM model performed significantly worse (RMSE 26.57 mg/dL), likely due to the small cohort size limiting deep learning training.
- T2DM: Random Forest performed best (RMSE 22.92 mg/dL, $R^2$ 0.73), followed closely by XGBoost and SVR. LSTM again underperformed (RMSE 33.25 mg/dL).
- General Trend: Simpler models (SVR, Random Forest) outperformed deep learning approaches, suggesting that for short-term horizons and limited data, complex temporal architectures may not be necessary or effective.
Interpretability Findings:
- Feature Importance: In both T1DM and T2DM, recent CGM measurements (specifically $t-1$ to $t-4$ ) were the dominant predictors, accounting for the vast majority of predictive influence. Static clinical variables contributed minimally in T1DM but showed slightly more structured, secondary contributions in T2DM.
- Model Behavior: T1DM predictions were characterized by high volatility and reliance on immediate glucose history. T2DM predictions showed smoother trajectories with slightly more influence from clinical variables (e.g., HbA1c, creatinine).
- SHAP vs. LIME:
  - Fidelity & Stability: SHAP demonstrated higher fidelity and stability, providing consistent explanations even in high-error, volatile scenarios. LIME exhibited numerical volatility, often exaggerating feature magnitudes or flipping signs in nonlinear regions.
  - Completeness: SHAP provided a comprehensive view of feature contributions, whereas LIME produced sparse explanations that sometimes omitted relevant but weaker features.
  - Interpretability: LIME offered simpler, more accessible visualizations (bar charts) suitable for rapid case review, while SHAP required more familiarity but offered deeper insight into nonlinear interactions.

Significance and Claims
The paper claims that accurate short-term glucose forecasting can be achieved alongside meaningful transparency through the integration of ML and XAI. The findings support the development of trustworthy AI systems for diabetes management by demonstrating that:

Model Selection: SVR and tree-based ensembles are robust choices for 30-minute glucose prediction, particularly when dataset sizes are constrained.
Clinical Plausibility: ML models primarily rely on recent glucose history, a clinically sound mechanism, though T2DM models incorporate additional clinical context.
Complementary XAI: SHAP and LIME serve complementary roles; SHAP is superior for validating model trustworthiness and global behavior, while LIME is effective for rapid, instance-level reasoning.
Future Direction: The study emphasizes that explainability should be a core component of AI deployment in healthcare, not an optional add-on, to facilitate clinician confidence and safe integration.

The authors note limitations, including the small T1DM cohort size, the exclusion of time-dependent contextual variables (meals, insulin, activity), and the inherent constraints of post-hoc explainability methods. They suggest future work should focus on larger, more diverse cohorts and the integration of dynamic contextual data.