Development of an artificial intelligence approach that employs genomic and brain imaging features to improve the diagnosis of Alzheimer’s Disease (AD)

Background In the past 20 years, genome wide association studies (GWAS) and brain imaging analyses have identified various genetic risk variants and structural brain differences between AD and cognitively normal (CN) subjects 1–3 . However, due to the relatively large and colinear feature space, it remains a challenge to use these variants and features to classify AD status at an individual level for personalized diagnosis, treatment, and prevention 4 . In this study, we applied a novel artificial intelligence‐based technique to transform both genomic and brain imaging features into 2D artificial images, and further utilized a deep convolutional neural network (CNN) to classify these artificial images 5,6 . Together, we show that the approach significantly improves the classification accuracy of AD subjects. Methods Participants were selected from the AD Neuroimaging Initiative. Table 1 details the samples and features for each classification analysis. Briefly, 808 subjects with 2451 single‐nucleotide‐variants (SNV) with p‐values<1e‐4 in previous AD‐GWAS 1 were included in classification of AD using SNV features. For classification of AD using brain imaging features, 1625 structural magnetic resonance imaging (sMRI) images from 941 subjects were included as samples. Each T1‐weighted sMRI image was entered into the FreeSurfer pipeline 7 to compute features of regional thicknesses, volumes, and surface areas. Next, every feature was normalized to a scale of 0‐255, and the entire feature set from every subject was transformed into a 2D artificial image 5 . A CNN classifier was then trained on 80% of the samples with internal validation and tested on the remaining 20% of the samples. Results Since our samples were significantly biased towards non‐AD subjects, both the area under the ROC curve (AUC) and the accuracy were reported in Table 2 to assess the performance of the classifier. As listed in Table 2, the artificial intelligence approach outperforms 10 other regular/popular classifiers in discriminating AD from non‐AD samples, with an accuracy and AUC of 70.99% and 0.64 for SNV features, 86.50% and 0.74 for sMRI features, and 91.60% and 0.85 for combined features. Conclusion Taken together, these results indicate that our approach can efficiently deal with large and colinear feature spaces to classify AD subjects more accurately at an individual level.