Background
Climate change is driving unprecedented health emergencies through extreme weather events, floods, droughts, and disease outbreaks. Policy makers and humanitarian organisations responding to these crises need rapid, evidence-based guidance, but traditional systematic reviews take over a year to complete. While Large Language Models (LLMs) like ChatGPT show promise for automating evidence synthesis, they lack critical scientific rigour: accepting claims uncritically, failing to assess study quality, and potentially perpetuating biases. Research synthesis requires systematic critical appraisal to ensure high-quality evidence is appropriately weighted and findings consider health equity and local context.
Novelty & Importance
This PhD pioneers a novel modular approach to automated evidence synthesis, where specialised LLM components perform distinct systematic review sub-tasks: critical quality assessment, structured data extraction including equity considerations, and scientifically robust synthesis. Unlike existing black-box approaches, this transparent methodology ensures outputs meet rigorous scientific standards. The research directly addresses global health inequities by focusing on climate-health emergencies in low- and middle-income countries where impacts are most severe but evidence is often neglected. This work will be embedded within SOLACE-AI, a £4 million Wellcome Trust-funded initiative, ensuring real-world impact through deployment with humanitarian partners in Ethiopia, South Africa, and India, and globally with organisations like Médecins Sans Frontières.
Aims & Objectives
This PhD aims to develop and rigorously evaluate novel language modeling approaches for producing scientifically robust evidence syntheses for climate-health emergencies. Specific objectives include: (1) developing evaluation datasets from climate-health systematic reviews; (2) creating and testing modular LLM architectures for critical appraisal using established frameworks; (3) comparing zero-shot, few-shot, and fine-tuning approaches across diverse models; (4) implementing reliability mechanisms including methods for models to decline answers when confidence is low; and (5) integrating validated methods into SOLACE-AI for field evaluation with humanitarian organisations responding to real climate-health crises.