Machine Learning to Predict Clinical Outcomes and Uncover Multimorbidity Patterns in Older Adults

: (Alternative Format Thesis)

Abstract

Background: Older adults aged ≥65 years account for a disproportionate number of hospital admissions, many of which are avoidable and driven by complex multimorbidity and polypharmacy patterns. As populations age, the ability to predict and prevent hospitalisations and identify patterns of multimorbidity becomes increasingly critical for healthcare systems. Older adults are also at heightened risk of hospital-associated complications, including hospital-acquired infections, delirium, and functional and cognitive decline. These adverse outcomes further reinforce the importance of prospective risk stratification of older adults. Traditional statistical approaches are based on a priori assumptions and are limited in their application to high-dimensional data and non-linear interactions common in real-world clinical settings. This thesis aims to investigate how machine learning (ML) can facilitate risk prediction and identify underlying patterns of multimorbidity and other clinically relevant factors in older adults, ultimately enhancing clinical decision-making and healthcare planning.

Methods: This thesis comprises four studies. First, a systematic review and meta-analysis investigated the performance of ML models in predicting mortality among older adults, highlighting key methodological gaps. Second, unsupervised learning approaches (hierarchical clustering, association rules) were used to identify multimorbidity clusters among 95,994 older adults from the UK Biobank who experienced emergency hospitalisations. Third, supervised ML models (random forest, XGBoost, logistic regression) were developed using the Inter-Resident Assessment Instrument data from 14,198 community-dwelling older adults with complex care needs to predict 30-day hospitalisation and identify modifiable risk factors. Finally, the models were validated on the UK Biobank cohort of 86,870 older adults to assess generalisability and consistency of the identified important risk factors across heterogeneous datasets. Interpretability techniques, such as variable importance plots and calibration curves, were utilised throughout.

Results: The systematic review and meta-analysis found that ML models demonstrate good discriminatory power for short-term and long-term mortality (AUC 0.80-0.81) but suffer from poor external validation and inconsistent outcome definitions. Clustering analysis revealed four distinct multimorbidity phenotypes associated with emergency hospitalisation. The four clusters were i) age-associated conditions, ii) immediately modifiable lifestyle factors and inappropriate polypharmacy, iii) cardiovascular symptoms, and iv) autoimmune disease and cancer. The supervised models achieved strong performance (AUC 0.79-0.97), good calibration, and identified the Drug Burden Index, alcohol use, and mobility issues as important predictors of acute hospitalisation. These findings were successfully validated in an external cohort, reinforcing the utility of modifiable variables for predicting acute hospitalisation.

Discussion: Across all studies, ML models outperformed traditional statistical approaches in handling complex interactions and large datasets. The integration of interpretable ML enabled clinically relevant insights, while external validation enhanced clinical credibility and generalisability. Notably, the research demonstrated how unsupervised and supervised ML can be used synergistically to both identify distinct multimorbidity patterns and predict clinical outcomes in older populations.

Conclusions: This thesis demonstrates the feasibility and clinical value of ML for predicting hospitalisation and identifying actionable healthcare patterns in older adults. Future work should prioritise longitudinal and multimodal data, integration into clinical workflows, and development with key stakeholders to ensure fairness and impact in real-world care.

Date of Award	25 Mar 2026
Original language	English
Awarding Institution	University of Bath
Supervisor	Prasad Nishtala (Supervisor) & Sandipan Roy (Supervisor)