This project aims to combine causal analysis, large language models (LLMs), and multi-agent reasoning to accelerate scientific discovery in mental health support. Modern psychiatry and clinical psychology generate vast volumes of heterogeneous evidence, from electronic health records and therapy transcripts to randomised controlled trials and the rapidly growing biomedical literature, yet the pace at which clinicians can synthesise this evidence, identify causal risk factors, and translate findings into practice remains limited. LLMs offer an unprecedented ability to read, summarise, and reason over scientific text at scale, but they are known to hallucinate associations, conflate correlation with causation, and inherit biases from their training data, making them unreliable as standalone scientific assistants. Deep learning models trained directly on clinical data face a complementary problem: they exploit spurious correlations and produce opaque predictions that clinicians cannot verify.
This project will develop a neuro-symbolic scientific discovery framework that integrates causal discovery and inference with LLM-based literature reasoning and multi-agent verification. Rather than treating clinical data and the published evidence base as separate sources, the framework will jointly mine causal hypotheses from both, use specialised agents to retrieve, critique, and reconcile claims against statistical evidence, and produce verifiable causal explanations that psychologists and clinicians can interrogate, extend, and challenge. By grounding LLM-generated hypotheses in identifiable causal structure rather than surface associations, the framework will support clinicians in formulating and verifying their own hypotheses more efficiently, predicting risks more reliably, and uncovering previously unknown risk factors from heterogeneous evidence at a speed and scale unattainable by manual review, leading to better health care for patients.
The project aims to:
- Design and develop a system that integrates mental health data in different formats from clinical, behavioural, and biomedical literature sources, including LLM-based extraction of structured causal claims from unstructured clinical notes and published text.
- Develop cutting-edge deep causal discovery and causal inference models that combine observational mental health data with LLM-encoded prior knowledge to identify and verify causal relationships between risk factors, treatments, and outcomes, while remaining robust to spurious correlations and confounding.
- Devise multi-agent reasoning models, comprising hypothesis-generation, literature-retrieval, statistical-analysis, and critic agents, that propose, debate, and refine causal hypotheses with calibrated uncertainty over their conclusions.
- Produce explanations that are easily understandable by health professionals and that are auditable against both the underlying data and the cited literature, supporting verifiable scientific discovery rather than opaque prediction.
