Premium
P1‐126: LEARNING BASED GENETIC RISK PREDICTION OF CONVERSION: COMBINATION WITH HIPPOCAMPAL VOLUME, APOE AND AGE IN MCI WITH 7‐YEAR FOLLOW‐UP
Author(s) -
Mikhno Arthur,
Redei Janos
Publication year - 2019
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.2019.06.681
Subject(s) - context (archaeology) , voxel , population , artificial intelligence , artificial neural network , hippocampal formation , medicine , computer science , psychology , machine learning , biology , paleontology , environmental health
Background: There is a need for non-invasive approaches to detect early AD, particularly in the context of population screening and clinical trial recruitment. Evaluation of existing genetic risk prediction approaches and their combination with other measures in MCI patients with only short-term follow-up (Kauppi et al., Frontiers in Neuroscience, 2018) has provided initial insights as to the clinical utility. The aim of the present study was to assess the utility of a novel learning based genetic risk prediction approach alone and in combination with other non-invasive measures, for detecting conversion to AD in MCI with long follow-up. Methods: Data were obtained from the ADNI-GO/ADNI2 cohort: 334 subjects for training (128 healthy controls, 159 MCI, and 47 AD); 54 independent MCI subjects for testing, divided into 33 converters to AD (PET 1 to 4 years prior to conversion) and 21 non-converters (5-7 years of MCI follow-up). Corresponding [F]florbetapir PET, T1MRI, Illumina Omni2.5M genotyping data, and APOE genotypewere obtained. Avoxel-based-SUVRwas then calculated for each PET/MRI scan pair (Mikhno et al.,AAIC 2018). Hippocampal volume (HCV) was quantified with GPU-accelerated software for hippocampus segmentation of the MRI that was validated against EADC/ADNI harmonized protocol (Mikhno et al., HAI 2016), calculated for left (L) and right (R) hippocampi, and the averagevolume (avgLR). A deep neural network (DNN) using TensorFlowwas trained to predict voxel-based-SUVR from SNPs. Test dataset performance for identifying MCI conversion was assessed for DNN alone, and for logistic regressionmodels that includedDNN in combinations with HCV, APOE, and Age in leave-one-out analyses. Performance was also benchmarked for APOE alone. Results: ROC AUC for identifying MCI conversion was 0.749 (DNN), 0.729 (DNN+HCV[avgLR]+APOE), 0.729 (DNN+HCV[avgLR]+Age), and 0.762 (DNN+HCV[R]+APOE). Sensitivity/specificity/accuracy for conversion was 91/62/80 (DNN), 76/71/74 (DNN+HCV [avgLR]+APOE), 79/71/76 (DNN+HCV[avgLR]+Age), 79/76/78 (DNN+HCV[R]+APOE), and 76/67/72 (APOE alone).Conclusions: Preliminary results of learning based genetic risk prediction demonstrate high AUC and improved sensitivity and accuracy for detecting MCI conversion in subjects with long follow-up, as compared to APOE alone. Combining the SNP model with HCV/ APOE yields sensitivity and specificity above 70%, and for the highest AUC model approaches 80% accuracy.