Towards Data-Driven Cognitive Disease Classification using Machine Learning and the Digital Symbol Digit Test

Abstract

There is no cure for Alzheimer’s and other cognitive diseases; however, there are treatments that help slow disease progression. Early detection of neurological dysfunction is often caught through screening tests like the Symbol Digit Test. Researchers at MIT and Lahey have been administering the Symbol Digit Learning Test using a digitizing pen that records data and pen strokes made by patients. This new data allows additional insights to be uncovered that a clinician may miss in a physical examination. We are the first group to perform analysis on this data for the digital Symbol Digit Learning Test – with this comes a number of challenges to build a working model that can aid clinicians in diagnosing this class of diseases. One challenge is creating an accurate multi-digit classifier that generalizes well when given messy digits drawn by patients with cognitive diseases. Additionally, more classifiers will need to be created to analyze the digital pen time-series data. Our research provides a computational approach to detecting cognitive diseases and brings to light novel insights such as diagnostic signals when subjects switch to a new row during the test and the impaired subject’s inability to match the healthy controls’ performance across the delayed recall task. Lastly, we developed a logistic regression classifier that flags dementia, Parkinson’s disease, and healthy subjects with an area under the curve of 0.93, 0.97, and 0.89 respectively. These new findings can be used in the clinical setting when administering these tests and can be fruitful for future machine learning model development.