An interpretable machine learning framework with data-informed imaging biomarkers for diagnosis and prediction of Alzheimer’s disease

Comprehensive Summary

Kang et al. developed an interpretable machine learning model called Glo and Loc EBM to improve the diagnosis and risk prediction of Alzheimer’s disease using MRI scans. This model examines both the whole brain and regional features of the MRI images and offers explanations for the involvement of different brain regions in the final prediction. This model has shown promising results using the data provided by the ADNI study, obtaining an AUC of 0.969 for distinguishing Alzheimer’s disease from healthy controls and 0.750 for predicting the progression of Alzheimer’s disease among patients with mild cognitive impairment. In a separate data set, the model has also shown an AUC of 0.871 for distinguishing Alzheimer’s disease from patients suffering from subjective cognitive decline. Importantly, the model maintained high accuracy while also allowing clinicians to understand how the decision was made.

Outcomes and Implications

This study is important because many artificial intelligence models for Alzheimer’s disease act as a black box, which means they can make a prediction but cannot explain it. In contrast, the current approach not only performs well but also provides a high level of interpretability using brain imaging features. This is particularly important in the clinic because understanding why a model is making a certain prediction is key to building trust among physicians. If validated further in broader clinical populations, this tool could support earlier diagnosis and better identification of patients at risk for progression from mild cognitive impairment to Alzheimer’s disease.