Using Natural Language Processing to identify psychiatric symptoms related to Alzheimer’s disease: a novel tool to enhance molecular brain research in Alzheimer’s disease

Abstract

AbstractBackgroundNeuropsychiatric symptoms (NPS), such as anxiety, depression, or agitation, are common across different stages of Alzheimer’s disease (AD). However, their molecular course is still poorly understood. While post‐mortem brain research provides a unique source to study brain‐specific molecular changes, clinical characterization of post‐mortem brain donors is challenging and primarily subject to case‐control designs. Here, we present a novel approach to post‐mortem brain donor characterization using natural language processing (NLP).MethodWe analyzed electronic health records (HER) of 92 brain donors (46 males, 46 females, mean age 81.9 years, B&B 0 ‐ VI) using NLP‐based NIMH Research Domain Criteria (RDoC) quantification. Individual RDoC scores were regressed to their neuropathological findings or RNAseq data from two independent brain regions, BA32 and insula.ResultNeuritic plaques (NP) in all four cortices show a high association with RDoC “cognition”, controlled for age and sex (frontal: ß = 0.38, p < 0.001; parietal: ß = 0.35, p < 0.001; temporal: ß = 0.37, p < 0.001; occipital: ß = 0.37, p < 0.001). These findings align with prior clinicopathological correlation studies where cognitive examinations were performed within one year before death (Nelson et al., J Neuropathol Exp Neurol, 2012). At pFDR<0.05, we identified 492 differentially expressed genes (DEGs) in the insula and 333 DEGs in the BA32. We have further built a model regressing individual RDoC levels against their DEGs in the two brain regions. In each domain, we observed a cluster of genes that is up‐ or downregulated in relation to their RDoC score (higher/lower domain burden). In addition, we saw that the majority of DEGs are RDoC‐specific. 223 DEGs were uniquely related to one domain (170 DEGs in the insula and 53 DEGs in BA32), while a total of 211 DEGs were associated with at least two domains (151 DEGs in the insula and 60 DEGs in BA32).ConclusionNLP‐based analysis of RDoC can be applied to post‐mortem case collections to phenotype brain donors and show promising associations with neuropathology or RNAseq data that can be used to discover novel molecular pathways related to psychiatric symptoms.