This is a project which is currently making use of HPC facilities at Newcastle University. It is active.
For further information about this project, please contact:
This PhD research develops a condition-agnostic adaptive treatment planning framework for intensive care unit (ICU) patients. The system integrates four heterogeneous data modalities (time-series vital signs, tabular laboratory results, clinical notes, and chest X-ray imaging) into fixed-dimensional embeddings to enable consistent patient trajectory modelling despite asynchronous data arrival patterns. The framework employs reinforcement learning with multi-level uncertainty quantification to generate treatment recommendations with explicit confidence estimates, addressing critical gaps in current ICU decision support systems that provide static predictions rather than adaptive recommendations. Validation uses MIMIC-IV, a comprehensive database of over 40,000 ICU patients.
The project uses Python with PyTorch for deep learning model development. The research involves training neural networks for processing time-series data, text analysis, medical image analysis, and reinforcement learning for treatment policy optimization. Monte Carlo Dropout techniques are used for uncertainty quantification across multiple levels. The work requires large-scale training on the MIMIC-IV dataset with extensive hyperparameter optimization experiments across different model architectures, embedding sizes, and fusion strategies. Computationally intensive tasks include model pretraining, ablation studies, reinforcement learning agent training with parallel simulations, and uncertainty calibration experiments.