Abstract
dc:description.abstractPatient trajectories within hospital are inherently dynamic and require repeated assessments to prevent clinical deterioration. Current early warning systems rely on single time-point assessments, ignoring temporal dynamics and missing gradual trends that unfold over time. Artificial intelligence, specifically Machine Learning (ML), offers an alternative approach that can leverage temporal patterns encoded within physiological time-series data. This thesis develops temporally aware ML models for clinical deterioration prediction, addressing three research questions: Can temporal deep learning architectures leverage sequential patterns to improve prediction compared to snapshot-based approaches? How do strategies for handling irregular temporal sampling affect model performance? How does pooling training data across multiple hospital sites affect model performance and stability? Using recurrent neural network architectures, models were trained on three datasets: MIMIC-IV, a curated NHS surgical inpatient dataset, and eICU; using death and critical care admission as proxies for clinical deterioration. Models were evaluated through time-series and time-point classification tasks, with comparisons against established early warning scores including NEWS2, eCART, and MEWS. Temporal models consistently outperformed traditional early warning systems. In NHS data, models detected 70% more observations within 24 hours of death or critical care admission than the best-performing early warning system (NEWS2), with a typical 6-hour earlier detection. Explicitly modelling irregular time intervals through temporal decay mechanisms further improved performance by capturing richer temporal dynamics. Multi-site data pooling improved mortality detection but revealed a mortality-intervention detection trade-off, reducing sensitivity to other deterioration events. APACHE-based hospital clustering achieved comparable performance to fully pooled models while using 60-73% less data. This thesis demonstrates the feasibility of temporal deep learning for developing clinical deterioration models across diverse care settings. It provides a curated NHS dataset for future research and identifies critical trade-offs in multi-site model development, offering practical strategies for collaborative model training without extensive data sharing.
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Leeds
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Pang, Allan Yiu Lun
- Advisors dc:contributor.advisor
-
- Johnson, Owen
- Kotze, Alwyn
- de Kamps, Marc
- Hall, Geoff
- ORCID person:identifier.orcid
Rights
- Language dc:language.dcterms_iso639_2
- eng