The aim of the project is to develop explainable and trustworthy machine-learning models for the early detection of cognitive decline and dementia. By integrating multimodal data (cognitive assessments, neuroimaging, genetics, lifestyle factors and comorbidities) from diversified information sources like partner cohorts (e.g., PROTECT and evergreen life), and ADNI. we will develop AI tools that predict individuals’ risk trajectories with high accuracy, transparency and fairness, so clinicians…
can identify high-risk patients earlier and have confidence in recommendations.
We formulate the task as a longitudinal prediction modelling problem to forecast cognitive decline and dementia progression. For early detection of cognitive decline and dementia, there exists a significant semantic gap among multi-source heterogeneous modalities, including neuroimaging (e.g. MRI or PET scans), and discrete textual data (e.g. cognitive assessment scores and clinical notes), polygenic risk scores and blood biomarkers, and the rich concept entities in biomedical Knowledge Graphs (KGs, as in Chen et al., 2023). To address this challenge, we propose an unsupervised, neural-based and latent knowledge-enhanced approach. The approach learns latent representations (e.g., neural entity embeddings) to elicit and align with entities across multimodal data and knowledge to ensure high explainability.
We will leverage an explainable agentic framework for early diagnosis with the developed modelling tools, where an example of an agentic framework is in the work (Li et al., 2025). Vision language models in the agentic framework will be leveraged to form natural language explanations from the visual-concept rationales (Zhao et al., 2024). The explanations will be evaluated by clinicians, with criteria including trustworthiness and bias. Bias will be measured through counterfactual data, e.g. perturbed patients’ demographical data, with clinical expert validation, from diverse populations. Bias will also be revealed through visual-concept rationales. The visual-concept rationales will be evaluated through two fairness metrics, comprehensiveness and sufficiency in NLP (DeYoung et al., 2020), with extension to the vision-language domain.
Uncertainty quantification in the agentic framework will be achieved through a novel, post-hoc calibration model based on the sub latent graph, the extracted rationales with quantified scores (e.g. attention or mutual information scores), and probabilistic prediction outcomes. The proposed approach will be validated through an integrated longitudinal multimodal prediction and explanation system. The outcomes and explanations will be presented to clinicians and patients during PPIE.
References:
- Chen, H. Dong et al., ‘Knowledge Graphs for the Life Sciences: Recent Developments, Challenges and Opportunities’, 33 pages, 2023, doi: 10.4230/TGDK.1.1.5.
- Li, X. Wang, D. Berlowitz, J. Mez, H. Lin, and H. Yu, ‘CARE-AD: a multi-agent large language model framework for Alzheimer’s disease prediction using longitudinal clinical notes’, npj Digit. Med., vol. 8, no. 1, p. 541, Aug. 2025, doi: 10.1038/s41746-025-01940-4.
- Zhao et al., ‘Explainability for Large Language Models: A Survey’, ACM Trans. Intell. Syst. Technol., vol. 15, no. 2, p. 20:1-20:38, Feb. 2024, doi: 10.1145/3639372.
- DeYoung et al., ‘ERASER: A Benchmark to Evaluate Rationalized NLP Models’, in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault, Eds, Online: Association for Computational Linguistics, Jul. 2020, pp. 4443–4458. doi: 10.18653/v1/2020.acl-main.408.
Location
University of Exeter, Streatham Campus
£21,805. The studentship will provide funding of fees and a stipend per annum for 2026/7