Premium
P2‐315: MODELS FOR PREDICTING RISK OF DEMENTIA: A SYSTEMATIC REVIEW
Author(s) -
Hou Xiao-He,
Yu Jin-Tai
Publication year - 2018
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1016/j.jalz.2018.06.1005
Subject(s) - dementia , receiver operating characteristic , medicine , population , neuropsychology , risk assessment , disease , gerontology , clinical psychology , cognition , psychiatry , environmental health , computer science , computer security
(ADNI)-1, Go&2, including ADAS-Cog13, RAVLT immediate, RAVLT learning, Functional Assessment Questionnaire, and MMSE. Baseline characteristics of the subjects are summarized in table 1. An autoencoder of cognitive measures was built on longitudinal data of ADNI-1 subjects using deep long short-termmemory (LSTM) recurrent neural networks, in order to characterize each individual subject’s temporal dynamics of cognitive measures using the autoencoder’s latent features. The autoencoder was then applied to all the MCI patients from ADNI-1, Go&2 to extract their latent features for building a Cox regression model to predict their conversion to AD. The Cox regressionmodel was built on data from ADNI-1, and its prognostic performance was evaluated based on the data from ADNI-Go&2. Age, gender, education years, and APOEε4 at baseline were used as covariates in the Cox modeling. We built prediction models on longitudinal data of the first 2 and 3 time points separately, and compared them with those built on cognitive measures at each single visit. The prediction performance was evaluated using concordance measure (C-index).Results:When the prediction models build upon ADNI-1 longitudinal data of 3 and 2 time points applied to ADNI-GO&2 subjects, they achieved C-index values of 0.896 and 0.873 respectively, significantly better than the prediction models built upon single-visit data at 6 and 12 months. All these prediction models had better prediction performance than the model built upon the baseline measures (C-index1⁄40.848) (Fig. 1). Conclusions: We developed a deep learning based method to characterize longitudinal dynamics of cognitive measures and build prognostic models to predict individual MCI subjects’ conversion to AD. The evaluation results have demonstrated that the proposed model can achieve promising performance for predicting pathological progressions using data within 1-year follow-up. The prediction models could potentially be further improved by including imaging measures.