Multimodal Machine Learning for Early Alzheimer's Disease Prediction: Integrating Biomarkers, Artificial Intelligence, and Precision Neurology

Early and accurate prediction of the onset and progression of Alzheimer's disease represents a major medical challenge, since years of pathological changes precede the first clinical sign of cognitive decline. So far, clinical prediction has relied on the analysis of individual biomarkers or, at best, data from a limited number of modalities. However, the complex and largely unknown biological mechanisms underlying the heterogeneity of the disease make it unlikely that a single biomarker or a single data modality will allow robust prediction of disease in individual subjects or of the rate of progression in already affected subjects. There is therefore a strong need for machine learning algorithms that integrate, within a unifying framework, the different classes of data generated in the context of precision medicine. Here, a framework of multimodal machine learning is presented and its potential for early prediction is discussed, with special emphasis on individualised risk assessment, explainable artificial intelligence and precision neurology. A series of limitations that may hinder implementation of the proposed framework in the clinical setting is also presented. These limitations are largely due to the heterogeneity of the different classes of data, the lack of large multimodal patient cohorts, the lack of external validation of performance, and the lack of model interpretability. By using models that are both interpretable by humans and validated against clinical endpoints, it may be possible to move towards more individualised methods of early prediction of Alzheimer's disease.