The PERMIT Project: Personalised Renal Function Monitoring via Information Technology

Al-Naher, Ahmed
(2020) The PERMIT Project: Personalised Renal Function Monitoring via Information Technology. PhD thesis, University of Liverpool.

[img] Text
201171661_Aug2020.pdf - Unspecified
Access to this file is embargoed until 1 January 2022.

Download (7MB)


Patients with heart failure are typically elderly and are among those most at risk of renal failure due to both their condition and their medication. Regular monitoring of renal function may allow early detection of renal decline and appropriate intervention to prevent renal failure. However, clinical guidance on renal function monitoring in heart failure is sparse and based on anecdotal evidence. To reduce unnecessary admissions caused by renal impairment in heart failure due to inadequate monitoring, standardised practice for renal monitoring would be of benefit. Given that each patient has individual co-morbidities and rates of renal decline, general guidelines may have minimal impact and there may be a need for renal monitoring that is personalised case-by-case. The aim of the PERMIT project (Personalised Renal Function Monitoring via Information Technology) was to develop the framework for creating such personalised guidance by using machine-learning on large clinical datasets. The goal was to create a prediction model that could highlight which patients with heart failure were most at risk of renal decline, in order to intervene before they required hospital admission. In light of developing a future predictive algorithm for use in clinical care, patient and clinician engagement with heart failure-related remote healthcare technologies was investigated. The aim of this was to improve the knowledge base so that future technologies, such as remote renal monitoring, can improve upon their accessibility and acceptability in this patient cohort. Studies examining remote care in heart failure were thematically synthesised in a qualitative systematic review. This generated 5 core themes of engagement: Clinical Care, Convenience, Communication, Ease of use, and Education, with different perspectives from patients and healthcare staff. The themes which were generated were assessed prospectively via a discrete-choice questionnaire survey given to heart failure patients (n=93). Binary logit analysis showed that ‘Clinical care’ was most valued by patients with heart failure and was almost twice as important as ‘Communication’, the lowest ranked theme. The study provided important insights into the lived experiences of patients with heart failure that will allow the development of future interventions with greater acceptability and engagement rates. To create the predictive model for renal decline, retrospective primary care data was obtained from SIR (Salford Integrated Records). This data was processed into a longitudinal dataset which included 3800 adult patients with newly diagnosed heart failure, over an 8.5 year study window. The clinical parameters of each patient were mapped longitudinally with creatinine over time. A model-based clustering algorithm known as ‘flexmix’ was applied to the data. In order to select appropriate clinical variables to input into the clustering predictive model, pairwise mixed-model linear regression was used to determine correlation between each clinical parameter and log(creatinine). The most correlative covariates were serum urea and serum potassium, with urea showing the highest R-squared value for explaining variance in creatinine over time. The final clustering model therefore used the inputs of: age at heart failure diagnosis; time since heart failure diagnosis; gender; IMD decile; and serum urea. This process produced seven discrete clusters of renal change over time which were ranked by severity. Evaluation of the algorithm was made using the assigned cluster models to predict creatinine over time in patients with heart failure. The MAPE (mean absolute percentage error) of the creatinine prediction was between 17-33% depending on the cluster assigned. The work outlined in this thesis represents an important step towards developing personalised renal monitoring guidance. Important clinical correlates of renal function decline, identified in the process, can be used for prognostic research in future studies. The error of the prediction values was variable and will thus require further optimisation using additional datasets and clinical studies in the future.

Item Type: Thesis (PhD)
Uncontrolled Keywords: renal failure, heart failure, machine learning, remote care technology, personalised medicine
Divisions: Faculty of Health and Life Sciences > Institute of Life Courses and Medical Sciences > School of Medicine
Depositing User: Symplectic Admin
Date Deposited: 15 Jan 2021 15:57
Last Modified: 20 Aug 2021 11:40
DOI: 10.17638/03104527