Survival Analysis in Cognitively Normal Subjects and in Patients with Mild Cognitive Impairment Using a Proportional Hazards Model with Extreme Gradient Boosting Regression

• Published in: Journal of Alzheimer's disease Vol. 85; no. 2; p. 837
• Main Authors: Khajehpiri, Boshra, Moghaddam, Hamid Abrishami, Forouzanfar, Mohamad, Lashgari, Reza, Ramos-Cejudo, Jaime, Osorio, Ricardo S, Ardekani, Babak A
• Format: Journal Article
• Language: English
• Published: Netherlands 01.01.2022
• Subjects: Aged; Aged, 80 and over; Alzheimer Disease - diagnosis; Alzheimer Disease - epidemiology; Alzheimer Disease - genetics; Cognitive Dysfunction - diagnosis; Cognitive Dysfunction - epidemiology; Cognitive Dysfunction - genetics; Disease Progression; Female; Genetic Testing - methods; Humans; Magnetic Resonance Imaging; Male; Neuropsychological Tests; Prognosis; Proportional Hazards Models; Risk Assessment - methods; Survival Analysis; magnetic resonance imaging; Alzheimer’s disease; Xgboost; survival analysis; proportional hazards model; mild cognitive impairment; hippocampal atrophy; brain; machine learning
• DOI: 10.3233/JAD-215266

Evaluating the risk of Alzheimer's disease (AD) in cognitively normal (CN) and patients with mild cognitive impairment (MCI) is extremely important. While MCI-to-AD progression risk has been studied extensively, few studies estimate CN-to-MCI conversion risk. The Cox proportional hazards (PH), a widely used survival analysis model, assumes a linear predictor-risk relationship. Generalizing the PH model to more complex predictor-risk relationships may increase risk estimation accuracy. The aim of this study was to develop a PH model using an Xgboost regressor, based on demographic, genetic, neuropsychiatric, and neuroimaging predictors to estimate risk of AD in patients with MCI, and the risk of MCI in CN subjects. We replaced the Cox PH linear model with an Xgboost regressor to capture complex interactions between predictors, and non-linear predictor-risk associations. We endeavored to limit model inputs to noninvasive and more widely available predictors in order to facilitate future applicability in a wider setting. In MCI-to-AD (n = 882), the Xgboost model achieved a concordance index (C-index) of 84.5%. When the model was used for MCI risk prediction in CN (n = 100) individuals, the C-index was 73.3%. In both applications, the C-index was statistically significantly higher in the Xgboost in comparison to the Cox PH model. Using non-linear regressors such as Xgboost improves AD dementia risk assessment in CN and MCI. It is possible to achieve reasonable risk stratification using predictors that are relatively low-cost in terms of time, invasiveness, and availability. Future strategies for improving AD dementia risk estimation are discussed.