PT - JOURNAL ARTICLE AU - Yang, Fan AU - Song, TzuAn AU - Dutta, Joyita TI - Alzheimer’s Disease Classification by Combining Tau PET Imaging and Genomics DP - 2021 May 01 TA - Journal of Nuclear Medicine PG - 1060--1060 VI - 62 IP - supplement 1 4099 - http://jnm.snmjournals.org/content/62/supplement_1/1060.short 4100 - http://jnm.snmjournals.org/content/62/supplement_1/1060.full SO - J Nucl Med2021 May 01; 62 AB - 1060Objectives: Alzheimer’s disease (AD) is characterized by the aggregation of two types of pathological proteins, tau and amyloid-β. Our objective here is to evaluate whether and to what extent the joint use of tau measures from positron emission tomography (PET) and genomics data can improve AD classification accuracy relative to tau PET alone or genomics alone. While genetics is known to play an important role in AD, no single genetic factor is able to accurately predict late-onset AD. We, therefore, incorporate genomic information either in the form of a polygenic risk score (PRS) or as raw single nucleotide polymorphism (SNP) data corresponding to genes considered critical in AD. Methods: For this study, we used data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. Our classifier training and validation data were based on the ADNI3 protocol (N = 103; 71.44 ± 7.76 years; 55 females; 65 CN, 38 cognitively impaired - CI). We used tau PET standardized uptake value ratio (SUVR) measures based on the 18F-Flortaucipir radiotracer. Tau PET SUVR means were computed for 85 anatomical regions-of-interest (ROIs) obtained via FreeSurfer parcellation. For the genomics data, quality control and imputation steps were performed using PLINK followed by PRS calculation using PRSice. A PRS is an estimate of an individual's genome-wide genetic risk for a trait or disease and is calculated from genome-wide association studies (GWAS) data. For PRS calculations, ADNI1 GWAS data (N = 756) were used as the base and the cognition status (known for the base data) was used as the phenotype, while ADNI3 data were set as the target. The following genomic data formats were compared: 1) PRS, 2) Raw genomics data consisting of 575 SNPs corresponding to 41 genes considered critical for AD, 3) Raw genomics data consisting of 17 of these SNPs highlighted in subsequent research. We trained and validated two classifiers to differentiate between CN and CI cases: 1) a random forest classifier based on bagged trees with parameters set to MATLAB defaults and 2) a neural network with 6 fully connected layers implemented using PyTorch. The tau PET inputs were concatenated with the genomics raw data vectors or the scalar PRS scores to create the classifier input vectors. Independent subsets of the data were used for classifier training (N = 60; 44 CN and 26 CI) and validation (N = 33; 21 CN and 12 CI). Results: Tau PET alone led to a CN vs. CI classification accuracy of 66.67% with the random forest classifier and 78.79% with the neural networks classifier. In comparison, with genomics data alone, the classification accuracies were in the 60.61-63.64% range with random forests and in the 63.63-66.66% range with neural networks. The random forest classifier achieved its best accuracy of 72.73% when the tau PET data was supplemented by scalar PRS inputs. The neural network classifier, on the other hand, achieved its best accuracy of 81.82% when the tau PET inputs were directly concatenated with the raw 17-SNP genomic data. Conclusions: Late-onset AD is influenced by multiple genetic variants. Our results suggest that the computation of a simple scalar feature derived from multiple genetic variants is particularly beneficial for boosting the performance of a simpler random forest classifier. A neural network with higher representational capacity, on the other hand, can extract the requisite information directly from the raw genomic data. Overall, our results show that the joint use of tau PET and genomic data could lead to improved accuracy in the prediction of cognitive impairment over tau PET alone or genomic data alone.