Multi-Modality Image Data Fusion and Machine Learning Approaches for Personalized Diagnostics and Prognostics of MCI due to AD

Alzheimer’s Disease (AD) is a devastating neurodegenerative disease. The greatest treatment potential lies in early stages before irreversible brain damage occurs. Early treatment requires early detection of the disease. Imaging holds great promise for capturing early signs of AD. This capability can be substantially strengthened by integrating neuroimages of different modalities that characterize brain structure and function from complementary aspects. However, although various machine learning (ML) algorithms have been developed to integrate multi-modality images for diagnosis and prognosis of AD, there is a lack of novel, robust, effective algorithms to address patient-wise missing modalities in the integration. In real clinical data, it is inevitable that some image modalities are unavailable to some patients due to high cost, insurance coverage, and safety constraints. Thus, the existing algorithms may only work for a small portion of patients who have complete modalities. This significantly reduces the access to advanced imaging-based diagnostic systems from the general patient population and in broad clinical settings. Because of the limited clinical utility, it is difficult to commercialize the existing ML algorithms into clinical systems/products, whereas the current imaging-based products on the market focus on single image modalities or image measurement, processing, visualization, and statistical analysis (without advanced ML capabilities). To fill the unmet market niche, this STTR Phase II project will develop the first-ever broadly-applicable clinical decision support system, Multi-neuroimaging for Detecting AD (Mind-AD), which can accommodate varying availability of image modalities across different patients to build classifiers and provide accurate diagnosis and prognosis of AD for each individual at the early MCI stage. Our Phase I has successfully demonstrated the feasibility of the Mind-AD system. At Phase II, we propose functional optimization and validation of Mind-AD in three aims. Aim 1 will optimize the accuracy and robustness of the diagnostic/prognostic models by integrating our Phase I IMTL model with efficient PSO feature selection. The integrated IMTL-PSO is very efficient in selecting optimal feature subsets to yield accurate, robust diagnostic/prognostic models especially on independent validation datasets. Aim 2 will develop a novel IMTL- DL (deep learning) model to integrate incomplete multi-modality volumetric images. While IMTL-PSO is based on features defined using anatomical knowledge of the brain, IMTL-DL extracts features in a data-driven manner. Aim 3 will integrate IMTL-PSO and IMTL-DL through decision fusion to best leverage their complementary, joint strength, and validate the resulting Mind-AD system using two independent datasets. Our project is significant because Mind-AD is the first early diagnostic/prognostic system for AD using advanced ML algorithms to integrate incomplete multi-modality image datasets. Mind-AD will facilitate early detection, early intervention, patient selection in drug trials targeting the early stage, and will help achieve these goals in in broad clinical settings due to the capability of accommodating varying availability of image modalities from different patients.The public health relevance of this project is to provide a clinical decision support system, Mind-AD, to facilitate accurate diagnosis and prognosis of AD at the early stage, through novel machine learning integration of multi- modality image datasets with patient-specific missing modalities. Early detection and early intervention bear the greatest potential for combating this devastating disease. Therefore, our system helps address the serious public health concern regarding AD which currently affects over 5.5 million people and incurs over $200 billion health care cost per year in the US.