Improving the diagnostics and treatment of patients with endometriosis : an algorithmic approach

Endometriosis is a chronic condition in which tissue resembling the uterine lining grows outside the uterus, often causing debilitating symptoms such as severe pelvic pain, dysmenorrhoea, dyspareunia, and infertility. Its aetiology remains uncertain, and in the absence of a cure, management focuses on alleviating symptoms. Definitive diagnosis still relies on laparoscopic surgery and biopsy, an invasive and resource-intensive procedure that contributes to average diagnostic delays of 8-11 years. Such delays can exacerbate disease progression, compromise fertility, and impose significant physical, psychological, and economic burdens. These challenges highlight the urgent need for non-invasive, data-driven approaches to accelerate diagnosis and improve patient outcomes. Using endometriosis as a motivating example, this thesis first applies a six-phase thematic analysis (with an added preparatory step for translation and cleaning) to free-text survey responses from UK and Brazilian cohorts. It then develops a proof-of-concept diagnostic pipeline that integrates insights from these data with causal inference methods, producing probability estimates and symptom-level explanations that are transparent to clinicians. The long-term aim is to contribute to earlier, safer, and more interpretable decision support for endometriosis and related conditions. A modular Bayesian framework was designed to represent risk factors, conditions, and symptoms within a causal structure. The framework combines marginal probability propagation with counterfactual sufficiency and disablement scoring, adapted from Richens et al., to provide clinically meaningful explanations. It is parameterised with prevalence values derived from survey data and evaluated through clinician-designed vignette testing and parameter perturbation experiments, assessing plausibility, stability, and robustness under uncertainty. This work delivers a prototype pipeline linking patient-reported survey data to a causal diagnostic engine. It introduces population-specific profiles to make prevalence assumptions explicit, employs dual conditional probability tables and background terms to handle uncertainty, and offers transparent explanation scores aligned with clinical reasoning. While focused on endometriosis, the design principles are general, offering a transferable blueprint for causal, auditable decision support in medicine.