Combining Machine Learning Methods and Scalable In Vivo Biomarkers for the Early and Accurate Detection of Alzheimer's Disease in Veterans

Va Boston Health Care System, Boston MA

Investigators

Abstract

Veterans are aging more rapidly than the US population, leading to greater numbers of Veterans with Alzheimer’s disease (AD) and related disorders (ADRD). Improved detection of AD pathology will provide Veterans with earlier access to services and initiation of treatment. Early and accurate detection of AD is critical for cost savings and effective treatment. However, current in vivo biomarkers for AD pathology require procedures not generally available such as amyloid positron emission tomography (aPET) or lumbar puncture. This project has the potential to provide Veterans across the country with a convenient and cost-effective tool to detect AD pathology, ideally before the disease has resulted in significant degeneration. This proposal aims to enhance the detection of ADRD with machine learning applied to results from a routine clinical neuropsychological battery and digital phenotyping of the clock drawing test (CDT). Machine learning and digital phenotyping show promise as scalable means to detect ADRD, but few studies have used methods with non-proprietary technologies and none have used Veteran samples. Traditional machine learning models will be computed to detect aPET positivity using archival data from a single VA medical center. [Record review will be completed on approximately 550 Veterans with aPET imaging followed by the VA Boston Memory Disorders Clinic between January 2015 and April 2023. Next, the project will examine high-dimensional but scalable digital phenotyping of a video CDT. A sample of approximately 450 Veterans with aPET imaging will be collected from the VA Boston Memory Disorders Clinic.] This project will assess deep learning classifiers derived from video recording of the clock drawing process. These classifiers will be assessed as standalone predictors of aPET positivity and in ensemble models with other clinical neuropsychological predictors. In addition, a sample of approximately 1000 Veterans with available clinical data will be collected from the VA Boston Memory Disorders Clinic. The project will examine whether clock drawing classifiers can be fine-tuned to detect AD as a suspected etiology, based on clinical diagnostic impression from the comprehensive neurological evaluation. This trained model will be externally validated in a prospective sample of 400 Veterans recruited from the National TeleMental Health Center (NTMHC) Memory Disorders service, which provides care for Veterans from over 25 VA healthcare systems. Finally, clock drawing classifiers will be fine-tuned to detect the presence of cognitive impairment in the same training and external validation cohorts. It is expected that machine learning algorithms will classify aPET positivity with modest accuracy using clinically available neuropsychological tests consisting of several hundred data points. However, it is expected that deep learning algorithms applied to thousands of images from the clock drawing process will produce excellent classification accuracy for all clinical outcomes across samples. This project could result in an accurate and scalable tool to detect AD pathology at an early stage and inform the need for confirmatory aPET imaging or lumbar puncture. The research and career development activities proposed will prepare Dr. Frank for his long-term goal of a research career in phenotyping of neurodegenerative conditions in Veterans. He proposes focused coursework, directed study, and mentoring from a group of well-established investigators who are leaders in their respective fields, as well as practical research experience completing the project proposed. The proposed training plan focuses on the following objectives: (1) training in traditional machine learning model development and validation, (2) training in cognitive neuroscience and the neuropathology of Alzheimer’s disease, (3) training in software pipeline creation to process and manage high-dimensional data, (4) training in deep learning methodologies applied to image analysis, and (5) training in VA culture related to ethical research and translational clinical science. The knowledge, techniques, and experience gained through this proposal will allow Dr. Frank to successfully compete for a Merit Review Award and become an independent investigator.