Multimodal, multiclass prediction of disease status in Alzheimer’s

Project Summary As the number of Americans living with Alzheimer’s disease (AD) is projected to reach 13 million by 2050, we must prioritize efforts for early disease detection. The bottleneck of early AD detection has greatly hindered clinical treatment and development of successful therapeutics. With greater availability of multimodal AD

biomarkers in clinical practice, we have a unique opportunity to leverage statistical machine learning for earlier detection of AD. Past work has demonstrated good classification accuracy of clinical diagnosis of AD using binary classifications (i.e., AD-dementia vs healthy cognition, HC; mild cognitive impairment, MCI vs HC; AD vs

MCI). These classifiers, however, are often reliant on unimodal biomarker inputs, and no formal comparison of multimodal biomarker integration (“fusion”) methods exist for predicting either AD clinical diagnosis or biomarker status as defined by the A/T(N) framework. Lack of optimal multimodal fusion strategies and holistic

diagnosis prediction beyond binary classification reduce the translational value of statistical machine learning classifiers in clinical practice. This proposal fills these gaps by evaluating several competing strategies for multimodal fusion and multiclass classification (e.g., AD vs MCI vs HC) using data from the National

Alzheimer’s Coordinating Center and Alzheimer’s Disease Neuroimaging Initiative. The strength of using large, multimodal datasets for disease prediction is accompanied by the challenge of handling missing data, a barrier for building a reliable classifier. This proposal will address these challenges with two specific aims: (1) compare

techniques for optimal data imputation and multimodal fusion, and (2) develop a multiclass model to accurately predict AD status (AD/MCI/HC and A+T+/A+T-/A-T-) using multimodal inputs. Preliminary analyses of multimodal data fusion in binary classification using random forest and sparse group lasso classifiers motivate

Aim 1. Preliminary analysis of two strategies of multiclass classification demonstrates feasibility of developing a multimodal, multiclass classifier for Aim 2. The proposed work will be enhanced by the excellent training and research environment at the University of California, Irvine (UCI), including direct access to 1 of 33 NIA-funded

Alzheimer’s Disease Research Centers (ADRCs). The ADRC offers a third, independent dataset to serve as a validation set to improve the rigor of results from the proposed experiments. The applicant will be supported by the joint mentorship of Dr. Craig Stark, the ADRC Biomarker Core Leader, and Dr. Babak Shahbaba, Director

of the UCI Data Science Initiative, and will receive advanced training in both aging and AD research and statistics and machine learning techniques. Fellowship training will be further strengthened by the additional mentorship of Dr. Peter Chang for machine learning, Dr. Michele Guindani for multimodal data fusion, and Dr.

S. Ahmad Sajjadi for clinical expertise in AD. The proposed training and research plans will result in the development of a reliable, multimodal, and multiclass classifier for AD status prediction to enable earlier disease detection and stratification of patients for more effective clinical trials.