Prediction of Dementia Types Using Machine Learning Methods

• Activity Number: 13 - Advances in Longitudinal Methods in Research on Aging and Dementia from the MEthods for LOngitudinal Studies of DEMentia (MELODEM) Initiative
• Type: Invited
• Date/Time: Sunday, August 8, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Biometrics Section
• Author(s): Chung-Chou H. Chang and Yueting Wang* and Yichen Jia and Mary Ganguli
• Companies: University of Pittsburgh and University of Pittsburgh and University of Pittsburgh and University of Pittsburgh
• Keywords: dementia; etiology; HyDaP clustering; maching learning; risk prediction; variable importance

Abstract

We used machine learning (ML) algorithms for survival data with competing risks to predict type of dementia (neurodegenerative or not) once a person has been identified as a case of incident dementia. Data come from the first 12 annual cycles of the Monongahela-Youghiogheny Healthy Aging Team (MYHAT), which is an on-going population-based cohort study. We include 120 individuals who reached the Clinical Dementia Rating (CDR) stage of 1 or higher during the follow-up period. Baseline and yearly assessed feature characteristics, including more than 300 sociodemographic, behavior, functional, medication, medical, and neuropsychological test variables for each eligible individual, were used in the analysis up to the cycle when CDR >=1 was observed. These variables had been posted on a secure database where expert clinicians reviewed the information on each case and rendered an etiological diagnosis based on available data on each incident case. We assessed model performance, stability of the ML algorithms, the identify the highest estimated variable importance and features that are the most predictive for distinguishing between neurodegenerative and nonneurodegenerative dementia cases.