University of Edinburgh
Harnessing routine data and healthcare contacts to predict risk and treatment requirements in older people
Abstract
dc:description.abstractRecent demographic shifts towards an ageing population have placed unprecedented demands on health and social care services. These pressures often contribute to inactivity and deconditioning among older people with complex conditions in hospital. Frail older adults are increasingly likely to require extended hospital care and rehabilitation following an acute admission, with a high probability of independence loss. Current guidelines place person-centred care at the heart of improving clinical pathways in older people. Data-driven algorithms incorporating Electronic Health Records (EHRs) could enable such approaches through early predictions of resource use and health outcomes to support targeted interventions in those at high risk of functional decline. The main aim of this thesis was to predict risk and resource requirements in older people using routine healthcare data. The research objectives included the use of observational health data and care provider contacts to integrate Machine Learning (ML) approaches for the measurement and prediction of risk and care needs in older adults. Secondly, I used Process Mining (PM) techniques to compare and explain care patterns in clinical pathways over this population. To understand the rapid shifts in clinical practices during the COVID-19 pandemic, I first conducted a literature and data overview of the COCHRANE REH-COVER systematic review series showcasing evidence of rehabilitation interventions. There was an evident research gap, as many studies did not clearly describe the content and length of rehabilitation programmes, reducing the ability to share best practices. I then conducted an observational data study using EHRs to report the quantity and intensity of rehabilitation services in consecutive hospitalised patients across three acute sites in the Lothian region of Scotland. I used longitudinal measures of rehabilitation delivery (accumulated rehabilitation sessions and minutes, time from admission to first contact), defined across 651 patients with and 10,940 without acute COVID-19 infection. A total of 128,646 rehabilitation-specific provider contacts were captured during hospitalisation across four rehabilitation disciplines. Patients with COVID-19 received more than double the rehabilitation time compared to those without the infection (365 [165, 772] vs 170 [95, 350] mins, p<0.001). A follow-up Process Mining (PM) data study used these health contacts to compare care provider patterns across the first two pandemic waves. I generated standardised event logs for 1,153 unique patients with COVID-19 (302 in Wave 1 and 851 in Wave 2), where cross-log conformance checking (log fitness, precision and generalisation) and graph distance measures were adopted to estimate differences in care patterns across subpopulations. For individuals older than 75, more dependencies between care providers were discovered in Wave 2, suggesting higher degree of complexity than in Wave 1. Observed contact patterns showed higher complexity in patients with extended hospital stay but were similar in those who required out-of-hours contacts and intensive therapy. The next data study demonstrated the clinical utility of routine data in pre-symptomatic risk identification of dementia in older people. In 144,113 community-dwelling older adults with linked primary and secondary care records, I aimed to predict 5, 10 and 13-year dementia diagnosis outcomes confirmed via routine health records. Over 13 years of follow-up, 11,143 (8%) of individuals were diagnosed with dementia. I developed a Gradient-boosted Ensemble ML algorithm (XGBoost) to generate risk predictions adopting both a data-driven (171 routine health variables) and clinically-supervised (22 curated variables) approach. I observed comparable diagnostic quality in both approaches, with the best performance measured using the clinically-supervised features (specificity 0.88 [95% CI 0.87-0.88] and sensitivity 0.55 [95% CI 0.53-0.57]) in 13-year outcomes. At the top decile of predicted risk, 1 in 3 individuals were diagnosed with dementia within 13 years, compared to a baseline population risk of 1 in 13 people. The final data study adopted an XGBoost regression model to predict derived estimates of nursing and rehabilitation care needs in older hospitalised patients who were admitted via an emergency department (ED) pathway. This study included 98,242 patients excluding those under immediate surgical services. A total 1,122,006 nursing and 575,629 rehabilitation care provider contacts were considered for prediction, where 42% of the study population received any rehabilitation. The regression model achieved weak to reasonable forecasting accuracy for received healthcare contacts with a Conditional Mean Absolute Percentage Error (cMAPE) between 49% [95% CI 48%-49%] at point of ED attendance and 34% [95% CI 33%-35%] at 72 hours post-admission. Increased pressure ulcer risk (Waterlow score) and delirium (4A’s Test) were uniquely important indicators of reduced nursing and rehabilitation contact frequency. Blood tests linked to muscle function and inflammation markers were associated with requiring more contacts. The models showed future potential for decision-support, targeting novel approaches for triaging and discharge planning in hospital front-door services for older people. Taken together, these EHR-driven algorithms have substantial potential for supporting healthcare providers in providing efficient treatment protocols in older individuals focused on improving their journey through the healthcare system.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Georgiev, Konstantin
- Advisors dc:contributor.advisor
-
- Anand, Atul
- Shenkin, Susan
- Fleuriot, Jacques
- McPeake, Joanne
Subjects
dc:subject × 6Rights
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- https://doi.org/10.7488/era/7035