Machine learning on longitudinal multi-modal data enabling the understanding and prognosis of Alzheimer’s disease progression

Abstract

Background Alzheimer’s disease (AD) is a complex pathophysiological disease. Its progression is heterogenous and associated with clinical symptoms as well as prognosis. Hence, understanding the progression of AD is important in guiding patient management, evaluating treatment response and designing therapeutic trials. We developed and validated a novel machine learning model for revealing the underlying disease progression patterns from longitudinal multi-modal data of AD. Methods The prospective ADNI study data (1192 patients with cognitive impairment and 388 healthy controls) was used to train, validate and test a novel personalized hidden Markov model to identify disease-related states. The transition diagram of disease-related states exhibited two distinct disease progression patterns. We used the index of disease-related states as a predictor, and evaluated the prediction performance on time to conversion to AD dementia. External validation was performed on the AIBL cohort (87 patients with cognitive impairment and 179 healthy controls). Findings Our model identified ten biologically and clinically meaningful disease-related states from longitudinal multi-modal data, which constitute three non-overlapping stages (i.e., neocortical atrophy [NCA], medial temporal atrophy [MTA], and whole brain atrophy [WBA]) and two distinct disease progression patterns (i.e., NCA→WBA and MTA→WBA) hidden in the trajectories of AD patients. Disclosed patterns showed different atrophy trends of brain regions, and presented distinct genetic, clinical and biomarker characteristics, especially in the early course of the disease. The index of the identified disease-related states provided a remarkable performance on predicting the time to conversion to AD dementia (C-Index: 0.923±0.007), in comparison with the other commonly recognized AD risk factors, including APOE (0.607±0.012), MMSE score (0.808±0.006), Abeta (0.7445±0.013), pTau (0.689±0.006), and hippocampal volume (0.774±0.010). Interpretation A machine learning algorithm was proposed to capture the full spectrum of AD progression by identifying meaningful AD-related states as well as distinct disease progression patterns from longitudinal multi-modal data, showing its potential for promoting the understanding of heterogeneous disease progression, and early predicting the conversion time to AD dementia.