Background
Schizophrenia causes substantial personal and societal burden, with about one-third of patients developing treatment-resistant schizophrenia (TRS), defined as non-response to two adequate antipsychotic trials. Clozapine is the only proven TRS treatment, yet its initiation is often delayed and unequally distributed—ethnic minorities, women, and younger patients face longer waits or reduced access. These challenges highlight two key needs: early, reliable TRS identification and equitable, timely clozapine use.
Recent advances in Generative AI (GenAI) and large language models (LLMs) offer new possibilities. LLMs can extract rich clinical data from electronic health records (EHRs) to predict TRS onset and treatment outcomes, while fairness-aware AI methods can identify and mitigate biases. LLM-based synthetic data generation further enables balanced, reproducible model development.
Novelty and Importance
This project is the first to jointly address early TRS detection and equitable clozapine prescribing using AI. It will develop LLMs to extract treatment adequacy, adherence, and side-effect patterns; create transformer and survival models to forecast TRS before two failed trials; and apply fairness-aware approaches—such as invariant representation learning and adversarial debiasing—to reduce disparities. Synthetic data will be used to generate counterfactual cohorts for fairness testing and robustness.
Earlier TRS detection could reduce ineffective polypharmacy, hospitalisations, and long-term disability. Fairer clozapine use addresses enduring inequalities in psychosis care and supports equitable AI in mental health.
Aims and Objectives
Aim:
- Develop and evaluate AI methods for early TRS identification and fair, timely clozapine use.
Objectives:
- Extract treatment data from EHRs using LLMs.
- Predict TRS onset with transformer-based models.
- Apply fairness-aware AI to detect and mitigate bias.
- Generate synthetic data to test fairness and robustness.
References:
- Howes, O. D. et al. Treatment-Resistant Schizophrenia: Treatment Response and Resistance in Psychosis (TRRIP) Working Group Consensus Guidelines on Diagnosis and Terminology. Am. J. Psychiatry 174, 216–229 (2017).
- De Freitas, D. F. et al. Ethnic inequalities in clozapine use among people with treatment-resistant schizophrenia: a retrospective cohort study using data from electronic clinical records. Soc. Psychiatry Psychiatr. Epidemiol. 57, 1341–1355 (2022).
- Chen, Z. Z. et al. A Survey on Large Language Models for Critical Societal Domains: Finance, Healthcare, and Law. Preprint at https://doi.org/10.48550/ARXIV.2405.01769 (2024).
- Da Silva, I. L. et al. Weak Reward Model Transforms Generative Models into Robust Causal Event Extraction Systems. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14707–14719, (2024)
- Pezoulas, V. C. et al. Synthetic data generation methods in healthcare: A review on open-source tools and methods. Comput. Struct. Biotechnol. J. 23, 2892–2910 (2024).
- Yan, H. et al. Position: LLMs Need a Bayesian Meta-Reasoning Framework for More Robust and Generalizable Reasoning. International Conference on Machine Learning (ICML2025)