Exploring Prediabetes Pathways: Using Machine Learning and Counterfactual Explanations for Type 2 Diabetes Prediction and Prevention

Console D., Lenatti M., Simeone D., Keshavjee K., Guergachi A., Mongelli M., Paglialonga A., “Exploring Prediabetes Pathways Using Explainable AI on Data from Electronic Medical Records,” Proceedings of the 34th Medical Informatics Europe Conference (EFMI MIE 2024), Aug 25-29, 2024, Athens, Greece. Studies in Health Technology and Informatics, 2024. In press.

Introduction

Type 2 Diabetes Mellitus (T2DM) is a chronic metabolic disorder characterized by hyperglycemia due to impaired insulin secretion. Early detection and personalized intervention are crucial to reducing the risk of T2DM and associated healthcare costs. Prediabetes (PD) is a reversible state that precedes T2DM, making it an important focus for early intervention.

This work leverages Machine Learning (ML) and counterfactual explanations to predict and prevent transitions between normoglycemia (NG), PD, and T2DM using Electronic Medical Record (EMR) data.

Materials & Methods

Dataset Extraction

Data was extracted from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN) using SQL queries. Features included were blood exams, glycemic biomarkers, general health indicators and presence of comorbidities.

Dataset Characterization

Univariate and bivariate analyses were conducted, including the Shapiro-Wilk test and Spearman’s correlation matrix. The separability of the target variable was assessed using boxplots and statistical tests.

Model Training and Evaluation

Four models were tested: Decision Tree, Random Forest, XGBoost, and Bagging of Logistic Regressions. Models were trained using stratified 5-fold cross-validation on scaled data. Total population, subgroup of patients with NG and subgroup of patients with PD were separately analysed. XAI techniques, such as feature importance and partial dependence plots (PDPs), were used to understand model predictions.

Counterfactual Explainability

Counterfactual explanations were generated using DiCE (Diverse Counterfactual Explanations) to determine the minimal changes needed for a patient with PD to regress to NG and prevent T2DM. Random and genetic search methods were tested.

Results

Dataset Characterization

Non-Gaussian distributions were observed for all features. Spearman’s correlation matrix showed low inter-correlation of features, except for total cholesterol and LDL. Discriminable distributions were found between the target classes, particularly for transitions ending in T2DM.

The similarity in feature ranges between PD and NG patients indicates the complexity of predicting PD. High correlation was found between some features, like FBS and HbA1c. LDL showed counter-intuitive decreases, likely due to medication.

Model Performance

The XGB model performed best on the total population and CurrentState=PD subgroup. For the CurrentState = NG subgroup, no model achieved satisfactory performance due to the imbalance of the dataset.

Population	Model	F1Macro	Sensitivity	Specificity
Total population	XGB	83%	86%	90%
CurrentState = PD	XGB	81%	76%	86%
CurrentState = NG	DT	58%	13%	99%

Glycemic biomarkers, BMI, and LDL were identified as the most important features, as can be seen from the feature importance and PDP graphs.

From left to right: Feature importance and PDP computed on total population

Counterfactual Explainability

Counterfactuals for transitions from PD to T2DM highlighted the importance of improving glycemic biomarkers and BMI to reduce T2DM risk. The random method generated fewer counterfactuals per record compared to the genetic method, which changed more features but produced fewer outliers.

Metric	Random Method	Genetic Method
Availability [%]	100	100
Mean number of CFs	10/10	10/10
Features changed per CF	1.63 (0.49)	6.80 (0.78)
Outliers on all features	49%	10%

In the example below some examples of counterfactual explanations are reported for a 40-year-old female subject without comorbidities.

Spiderplots showing counterfactual explanations

Conclusions

This work contributes to understanding transitions between glycemic states using ML on primary care data, focusing on prediabetes for early intervention. Counterfactual explanations provide actionable insights for personalized prevention plans.

Instructions

This work is not reproducible due to fact that the CPCSSN is not a public database and therefore our dataset cannot be legally shared.