Cardiovascular events and mortality in hip fracture patients : an epidemiological and machine learning study using electronic health records

Abstract

Hip fractures represent a critical outcome of osteoporosis, significantly affecting morbidity and mortality among older adults. Emerging evidence has shown an association between hip fractures and cardiovascular diseases (CVD), with CVD emerging as a leading cause of death among hip fracture patients. The complexity introduced by the comorbidities and varied healthcare needs of older adults with hip fractures complicates research and care in this domain. Nonetheless, advancements in electronic health record (EHR) databases and machine learning (ML) technologies present new opportunities to improve outcomes through enhanced research and clinical strategies.

This thesis aims to explore the association between hip fractures and cardiovascular diseases, and to leverage machine learning for patient stratification and mortality prediction in the context of hip fracture management. All the studies in this thesis were conducted using the Clinical Data Analysis and Reporting System (CDARS), a population-wide EHR database in Hong Kong.

The initial study was an epidemiological investigation evaluating the immediate risk of major adverse cardiovascular events (MACE) following hip fractures. The study showed a significant association between hip fractures and increased 1-year risk of MACE, with the risk being the highest near the time of hip fracture. This study illustrated the need for prompt cardiovascular evaluation in older adults diagnosed with hip fractures.

Building upon the epidemiological findings, the focus shifts to the utilisation of ML on EHR data. The second study leveraged unsupervised learning to uncover clinical subphenotypes of hip fracture patients, and examined their associations with MACE, mortality, and healthcare utilisation in Hong Kong and the United Kingdom. Through latent class analysis, this study identified hip fracture patient clusters with differential risk profiles, and consistently pinpointed heart failure as a key baseline characteristic associated with poor prognosis. This study showed that personalised care of hip fracture patients, considering their specific subphenotypes, is required to prevent MACE.

The final study adopted supervised learning to develop 1- and 5-year mortality prediction models for hip fracture patients, using variables including age, sex, medical diagnoses, and medication histories from the CDARS database. Advanced ML models, particularly the Gradient Boosting Machine (GBM) and eXtreme Gradient Boosting (xgbTree), showed promising predictive and calibration performance. These models could potentially support clinicians in assessing mortality risk, thereby aiding clinical decision-making.

Collectively, the findings from this thesis enrich our understanding of the associations between hip fractures, CVD, and mortality. This thesis also demonstrated the potential of machine learning to refine risk assessment and support personalised care for hip fracture patients. Future research could validate these findings across diverse clinical and population settings. This thesis contributes to the broader effort of harnessing big data and machine learning technologies to improve healthcare outcomes and optimise resource allocation.