Explainable and transparency machine learning approach to predict diabetes develop

Abstract

Purpose

This study aims to address the problem of type 1 diabetes by utilizing machine learning techniques and developing a decision support system based on Explainable Artificial Intelligence (XAI). The main research question is to predict the risk of developing type 1 diabetes in a population using different machine learning algorithms, while ensuring interpretability and transparency of the decision support system. The study builds upon a case-control study conducted by previous researchers, who approached the problem from a statistical-parametric perspective.

Method

In this work, various machine learning algorithms, including Decision Trees (DT), Deep Neural Networks (DNN), XGBoost (XGB), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Support Vector Classifier (SVC), are employed. The algorithms are evaluated based on their ability to predict the disease risk accurately and consistently on both the training and validation datasets. Additionally, Explainable AI techniques such as LIME (Local interpretable model-agnostic explanations) are employed to contextualize and interpret each prediction and assess the importance of various characteristics influencing the probability of developing the disease.

Results

The results obtained from the application of machine learning algorithms show promising outcomes on both the training and validation datasets. However, the best-performing algorithms are not necessarily those with the highest accuracy, as they may suffer from overfitting. Instead, algorithms such as DNNs (97%) or KNNs (93%) exhibit similar behavior on both training and test datasets, making them more reliable, LR and SVC both around (98.3%). The adoption of Explainable AI techniques enables the measurement of each characteristic’s importance and the analysis of factors influencing the disease’s development probability. This allows the development of a clinical decision support system (CDSS) that is immediately understandable, transparent, and interpretable. By leveraging machine learning techniques and Explainable AI, this study addresses the challenge of type 1 diabetes prediction and decision support.

Conclusion

The results indicate that algorithms like DNNs and KNNs offer reliable performance in predicting the risk of developing type 1 diabetes. The integration of Explainable AI techniques, specifically LIME, enhances the interpretability of predictions and provides insights into the factors influencing the disease. The developed CDSS based on XAI can potentially assist healthcare professionals in making informed clinical decisions, thereby improving patient care and management of type 1 diabetes.