Explainable AI (XAI) Models
This method uses artificial intelligence models that can explain their predictions to forecast serious health events in adults with type 1 diabetes or multiple sclerosis. It helps understand which factors most influence the risk of severe hypoglycemia, diabetic ketoacidosis, or relapses, and considers differences between men and women.
At a glance
Use when
Predicting clinical events where interpretability is critical, such as risk stratification in chronic diseases; when sex-specific insights are needed; when integrating AI into clinical decision support systems.
Avoid when
When data is highly sparse or unstructured without preprocessing; when real-time prediction is required without computational resources; when patient-level explanations are not needed.
Inputs
Clinical, socioeconomic, physical, and mental health data from adult patients with type 1 diabetes or relapsing-remitting multiple sclerosis, including sex-stratified datasets.
Outputs
Predictions of severe hypoglycemia episodes, diabetic ketoacidosis events, and MS relapses, along with interpretable explanations (via SHAP values) of feature importance and model decisions.
How it works
The study develops and evaluates multiple machine learning models (XGBoost, LightGBM, CatBoost, Adaboost, random forest, linear regression) enhanced with SHAP (SHapley Additive exPlanations) for interpretability. A three-step process includes: baseline model development, performance improvement via ReliefF feature selection and sex-stratified modeling, and model explanation using SHAP. Models were applied to predict severe hypoglycemia (SH), diabetic ketoacidosis (DKA) in type 1 diabetes, and relapses in relapsing-remitting multiple sclerosis (RRMS). Feature selection and sex stratification improved F1 scores, with best results reaching 84.95% (SH in females), 78.67% (DKA), and 84.55% (RRMS relapse in males).
- Project
- HTx
- Funding
- Horizon 2020
- Project status
- Completed 2024
- HTA domains
- Clinical Effectiveness
- Categories
- ML/AIPredictive Modelling
- Technology
- Medicines
- Assumptions
- Relevant predictive features are available and measurable; sex-specific patterns exist in disease progression; historical data reflects future outcomes; feature selection improves model generalizability.
- Strengths
- Incorporates explainability via SHAP to increase transparency; uses feature selection to enhance performance; accounts for sex differences; evaluates multiple strong ML algorithms; validated on real clinical outcomes with high F1 scores.
- Limitations
- Limited to adult populations with T1D or RRMS; model performance may vary in different healthcare settings or populations; reliance on quality and completeness of input data; not tested prospectively in clinical workflows.
- Geographic & clinical scope
- T1D, RRMS
- Also known as
- XAI models, Explainable AI for clinical prediction, SHAP-based prediction models
Questions this answers
- › Can AI predict dangerous health events in type 1 diabetes and multiple sclerosis patients?
- › How can we understand which factors are most important in AI predictions for patient outcomes?
- › Do prediction models work differently for men and women?
- › What machine learning methods work best for predicting hypoglycemia, ketoacidosis, or MS relapses?
- › How can we make AI models in healthcare more transparent and trustworthy?
- › Can feature selection improve the accuracy of clinical prediction models?
References & sources
Related methods
Similar by meaning
Beta record. Generated from the primary source via AI extraction and independent audit, pending final human review.