Genetic architecture of multiple pain phenotypes in UK biobank and cross-site AI-ECG classification

Abstract

This thesis brings together two strands that address challenges in human health: the genetic architecture of common musculoskeletal pain phenotypes and the reliability of artificial intelligence for clinical electrocardiography (ECG). The first strand asks what inherited factors contribute to hip, knee, neck or shoulder, and widespread pain, and how these factors relate to other traits and phenotypes. The second strand asks how to build ECG classifiers that remain accurate when deployed across hospitals that differ in devices, populations, and workflows.
Musculoskeletal pain imposes a major health and economic burden, with a high prevalence and disability in older adults. For common phenotypes, including hip, knee, neck or shoulder and widespread pain, the genetic underpinnings remain only partially characterized. Electrocardiography is central to the di agnosis of cardiac disease. Although AI-based ECG classification has advanced rapidly, performance often degrades across hospitals because of variability in devices, populations and workflows, leading to systematic error.
For pain genetics research, phenotypes were derived from harmonized UK Biobank questionnaires using consistent inclusion and exclusion rules. We introduced a refined case–control definition to capture a broader spectrum of the targeted pain phenotypes. A genome-wide association study (GWAS) was then performed for each pain phenotype, with genotypes undergoing standard sample and variant quality control and associations estimated using mixed models that account for relatedness and population structure. Significant sig nals were assessed for reproducibility through replication in multiple exter nal cohorts. Post association analyses included gene and tissue enrichment with Functional Mapping and Annotation (FUMA) and Multi-marker Analy sis of GenoMic Annotation (MAGMA), estimation of single nucleotide poly morphism (SNP) heritability and cross trait genetic correlation using linkage disequilibrium score regression (LDSC), phenome-wide association study (Phe WAS)oflead variants, transcriptome-wide association study (TWAS) integrat ing reference expression resources, and two-sample Mendelian randomization with sensitivity checks for instrument validity and horizontal pleiotropy. For AI-ECG research, a conditional adversarial multi-task framework was imple mented that simultaneously optimizes ECG classification and removes con founding variables, with the adversarial site-ID serving as a proxy variable to ensure marginalization over the learned conditional distribution of site-specific variability. Evaluation comprised stratified cross-validation, external testing on independent cohorts, repeated training runs to assess stability, and formal comparisons against ResNet baselines.
In pain genetics, we identified multiple novel loci across hip, knee, neck or shoulder, and widespread pain. We present the first sex-stratified GWAS of these phenotypes and demonstrate genetic differences between females and males. Integrative post-GWAS analyses, including tissue expression analysis, LDSC, TWAS, PheWAS, and Mendelian randomization, provided biological context and revealed links to other clinical and behavioral traits. In AI-ECG, we developed a practical and generalizable classification framework that re mains robust under distribution shift across test sites and improves accuracy and efficiency over convolutional baselines.

Date of Award	15 Apr 2026
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Weihua Meng (Supervisor), Tania Dottorini (Supervisor) & Mainul Haque (Supervisor)