Laboratory of AI for Digital Health | Engineering Science Department | University of Oxford

Reinforcement learning (RL) shows promise for optimising dynamic treatment regimes (DTRs) in personalised medicine, but lacks standardised benchmarking. We introduce DTR-Bench, a platform with four healthcare simulations (chemotherapy, diabetes, etc.), evaluating RL algorithms under real-world challenges like noise and variability. Results reveal performance gaps, emphasising the need for more robust, adaptive RL methods in DTRs.

Deep learning models excel in various applications but typically require extensive, high-quality data, posing challenges in biomedical settings with incremental, limited data. Few-Shot Class-Incremental Learning (FSCIL) methods offer a promising approach while also depending on strong base models that face the same aforementioned limitations. To address limitations, we propose AnchorInv, which uses synthetic samples generated from anchor points in feature space for buffer replay. Tested on three physiological datasets, AnchorInv efficiently reduces forgetting and adapts effectively to new categories, outperforming existing state-of-the-art methods while enhancing data privacy.

Multimorbidity, the co-occurrence of multiple chronic conditions in an individual, has become a global health challenge affecting populations in high-income and low- to middle-income countries. Despite its increasing prevalence, critical gaps remain in understanding its progression, burden, and determinants to better guide prevention and treatment. Here, by leveraging linked primary care, hospitalisation, and mortality records from 3.3 million individuals with multimorbidity in England, we conducted a longitudinal cohort study to characterise multimorbidity across multiple dimensions, including condition profiling, progression trajectories, healthcare burden, and social and biological factors.

We propose CAMELOT++, a novel deep clustering method for multi-task learning using temporal electronic health records (EHRs), enhancing ICU outcome prediction and management. First, CAMELOT+ improves clustering and accuracy through mixed-data modelling that separately encodes static and dynamic features. Further extending to multi-task learning, CAMELOT++ integrates survival analysis, enabling interpretable cluster-specific survival curves, and significantly improving prediction and interpretability for clinical applications.

Longitudinal Electronic Health Records (EHRs), though valuable for personalised medicine, often have informative missing data. To address this, we propose IGNITE, a deep-learning model using a conditional dual-variational autoencoder with dual-stage attention and an Individualised Missingness Mask (IMM) to impute personalised realistic values. Validated on three large datasets, IGNITE outperforms state-of-the-art models for missing data reconstruction and prediction tasks, also enabling generation of synthetic patient data.