AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Machine Learning-Derived Histologic Phenotypes for Genetic Mapping of Hippocampal Sclerosis of Aging
By Bahramy A, Tahmasebidehkordi H, Julian D et al.
Background: Hippocampal sclerosis of aging (HS) is a common, under-recognized contributor to cognitive decline that often co-occurs with Alzheimer's disease, LATE neuropathologic change (LATE-NC), and cerebrovascular small-vessel disease. Current assessments often rely on subjective binary neuropathologic ratings, limiting reproducibility and the power to detect genetic and vascular contributors to HS severity. Hypothesis: Objective, machine learning (ML)-derived quantitative HS phenotypes capture regional neuronal loss and increase power to detect its genetic and cerebrovascular determinants. Methods/Approach: H&E-stained hippocampal whole-slide images (WSIs; n=460) from the University of Pittsburgh Alzheimer’s Disease Research Center (ADRC) autopsy cohort were analyzed. For model development, 30 WSIs representing HS, evolving HS, and non-HS were sampled across CA1, CA1/subiculum junction, and subiculum. Manual annotations generated 1,440 tiles and 14,994 labeled neuronal and non-neuronal cells. A StarDist/U-Net model was trained/validated for cell detection and classification. CA1 and subiculum were subdivided into five regions each to quantify neuronal density and neuronal fraction across 10 hippocampal subregions. Continuous traits were generated for CA1, subiculum, and overall hippocampus, and Gaussian mixture models were used to derive binary phenotypes. Genome-wide association analyses (GWAS) were performed in 381 genotyped cases, adjusting for age at death, sex, and three principal components. A validated deep-learning vessel-segmentation pipeline from our group is being applied to quantify arteriolosclerosis, enabling future integration of vascular pathology with image-based genetic analyses. Results/Data: At an intersection-over-union threshold of 0.5, the model showed strong object-level performance (precision=0.88, sensitivity/recall=0.83, F1=0.86, matching accuracy=0.75). ML-derived phenotypes captured regional heterogeneity of neuronal loss, with CA1 more vulnerable than subiculum. Using these phenotypes for image-based quantitative trait locus mapping (image-QTL), GWAS models showed no genomic inflation (λ=0.99-1.02). In this discovery sample, a genome-wide significant locus (P<5х10 ⁻ ⁸ ) for CA1 neuronal loss mapped to rs7732684 near DRD1 (dopamine receptor D1). Suggestive loci included TCF7L1, TCIM, MIR99AHG/miR-99a, and AGT. Established HS loci (TMEM106B, ABCC9) showed concordant nominal signals. Conclusions: ML-derived subregional HS phenotypes enable objective, high-resolution image-QTL mapping and reveal candidate loci (DRD1, TCF7L1) for hippocampal neuronal vulnerability. Ongoing integration of TDP-43 pathology and arteriolosclerosis measures will allow evaluation of neurodegenerative and small-vessel disease contributions to hippocampal neuronal loss.