Time Series Analysis of Cognitive Scores for Alzheimer's Prediction: An LSTM Deep Learning Approach
Year: 2024
Type: article
Abstract: Timely identification of persons predisposed to Alzheimer’s disease (AD) dementia is crucial for the development of disease-modifying therapeutics. This research aims to predict the clinical diagnosis, cognitive function, and ventricular volume of a person at each subsequent month indefinitely, based on multimodal Alzheimer's disease indicators and clinical diagnoses from one or more timepoints. We introduced a recurrent neural network (RNN) model and used it on data from The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) challenge, which includes longitudinal data from 1,677 people (Marinescu et al. 2018) sourced from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). We evaluated the efficacy of the RNN model against three baseline algorithms over a forecast period of six years. Most prior research on forecasting Alzheimer's disease development neglects the problem of missing data, a common challenge in longitudinal studies. We examined three distinct ways for addressing missing data. Two of the solutions addressed the missing data as a "preprocessing" concern by imputing the absent data via the prior timepoint ("forward filling") or linear interpolation ("linear filling"). The third technique used the RNN model to complete the missing data during both training and testing, referred to as "model filling." Our findings indicate that the RNN using "model filling" outperformed baseline techniques, such as support vector machines/regression and linear state space (LSS) models. Nonetheless, there was no statistically significant difference between the RNN and LSS in predicting cognition and ventricular volume. Significantly, while using longitudinal data in the training process, our analysis revealed that the trained RNN model had comparable performance whether utilizing either one or four input timepoints, indicating that our methodology may be effective with just cross-sectional data. An previous iteration of our methodology achieved a 5th place ranking among 53 submissions in the TADPOLE competition in 2019. The present methodology is positioned 2nd among 56 submissions as of August 12, 2019.
Author: Amuloju Bala Bhavani
Language: English
Subfield: Psychiatry and Mental health
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7217741464
