This project aims to develop multimodal and longitudinal generative AI models to optimize diagnostic decision-making and patient care pathways in outpatient medicine. Outpatient care involves complex sequences of clinical actions-consultations, prescriptions, lab tests, and referrals-whose coordination is critical to improving diagnosis timeliness, resource use, and patient outcomes.
The proposed research will design and evaluate generative models capable of recommending optimal sequences of diagnostic actions based on both structured (e.g., lab results, prescriptions) and unstructured (e.g., clinical notes, reports) health data over time. The objective is to generate interpretable, guideline-compliant diagnostic plans that minimize diagnostic delay, unnecessary procedures, and cost, while maintaining clinical safety and fairness.
Methodologically, the work will focus on: (1) building and benchmarking datasets and baseline models for outpatient diagnostic prediction; (2) developing temporally-aware multimodal architectures that integrate structured and textual information; (3) incorporating probabilistic modeling to quantify uncertainty and improve robustness; and (4) leveraging reinforcement learning to train models capable of planning diagnostic actions aligned with clinical guidelines.
The UK Biobank dataset will be used to validate the generalizability of these approaches across large-scale, longitudinal, multimodal health data. Complementary evaluations will be performed on both open (e.g., MIMIC-IV) and proprietary datasets (e.g., curated outpatient vignettes from Doctolib). The ultimate goal is to develop clinically grounded, explainable AI tools that support practitioners in decision-making and care coordination in real-world outpatient practice.
