Back to search

The University of Edinburgh

Personalised screening intervals in diabetes care: evaluating risk-based models for diabetic retinopathy and diabetic foot disease

Abstract

dc:description.abstract

BACKGROUND AND AIMS: Diabetes is a major public health concern, affecting 4.4 million people in the UK, with complications such as diabetic retinopathy and diabetic foot leading to significant morbidity. Routine screening programs aim to detect these complications early, but current guidelines rely on fixed screening intervals based on prior test results. This approach overlooks variations in individual risk profiles, potentially leading to unnecessary screening for low-risk individuals and delayed detection for high-risk patients. This research investigates whether personalised, risk-based screening schedules can improve upon current guidelines. THIS STUDY AIMS TO: Develop and validate survival models for diabetic retinopathy and diabetic foot complications. Compare risk-based screening schedules with existing guidelines to evaluate their ability to stratify patients according to risk, and their impact on healthcare workload. Collect feedback from healthcare professionals on the utility of those models and gain insight into the barriers and requirements for implementation. materials and methods: This retrospective cohort study analysed data from 102,638 individuals with type 1 or type 2 diabetes, attending Greater Glasgow and Clyde NHS clinics between 2006 and 2021. Interval-censored parametric survival models, and a GLM-cloglog were developed using forward selection from over 200 features, including laboratory results, comorbidities and prescriptions. Model performance was assessed using precision-recall (PR) AUC, receiver-operating characteristic (ROC) AUC, and Uno’s concordance (C) index, while calibration was evaluated using calibration plots. Separate models were developed for diabetic foot and retinopathy, each predicting time to severe or any (mild or severe) adverse outcomes. Personalised time-to-screen recommendations were obtained by sampling the survival function, conditional on individual covariates, at risk thresholds matching the prevalence of the corresponding outcome. The proposed screening strategy was compared with current guidelines using confusion matrix metrics and workload change analyses. Additionally, the calculated survival estimates were reviewed by diabetes care professionals and their views were gathered on: 1) whether the model’s predictions align with clinicians’ expectations, and 2) their opinions on extending the screening interval length limits. RESULTS: Among 401,550 retinopathy screening appointments, 4.8% resulted in severe outcomes, while 27.6% resulted in mild or severe outcomes. Similarly, in 368,042 foot screenings, 9.4% led to severe and 28.1% to mild or severe outcomes. The retinopathy model (log-normal accelerated failure times) for severe outcomes contained 13 covariates and on the withheld test set achieved ROC AUC 0.92, PR AUC 0.34 and C 0.70, while the model for any outcomes contained 12 features and achieved ROC AUC 0.89, PR AUC 0.81 and C 0.74. The diabetic foot model (log-normal proportional hazards) for severe outcomes contained 19 covariates and achieved ROC AUC 0.89, PR AUC 0.81 and C 0.76, while the model for any outcomes contained 19 features and achieved ROC AUC 0.73, PR AUC 0.34 and C 0.70. When applying the best-performing models at a decision threshold equal to outcome prevalence, the proposed approach outperformed current policy in terms of sensitivity, negative predictive value, specificity and precision. Due to the generally low risk across the cohort, models recommend longer re-screening intervals for the majority of appointments in the dataset compared to current guidelines. Personalised screening intervals, with interval length limits identical to that of current guidelines, achieve 11% reduction in workload for retinopathy and 3% for diabetic foot screening. Lowering risk thresholds by 40% for retinopathy and 20% for foot screening, therefore facilitating earlier detection, allowed for a 2% and 0% reduction in workload compared to current policy. An expert panel of seven diabetologists and two nurses found that the model’s risk predictions aligned with their clinical intuition. They supported extending screening intervals but expressed concerns about dynamic risk changes, patient adherence, and potential anxiety over missed healthcare opportunities. CONCLUSION: This thesis successfully developed an analytical framework for generating models for personalised screening of diabetic complications. The resulting risk trajectories were favourably assessed by healthcare professionals, who recognised their potential for improving risk stratification. Overall, personalised screening schedules, coupled with a reduction of the level of risk at the time of screening, could enhance healthcare resource efficiency while improving patient outcomes.

Degree

thesis:*
Grantor dc:publisher
The University of Edinburgh
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Banas, Wojciech
Advisors dc:contributor.advisor
  • Mercer, Stewart
  • Manataki, Areti
  • Wake, Debbie

Subjects

dc:subject × 8

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
University of Edinburgh
Base URL
era.ed.ac.uk/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Banas, Wojciech. Personalised screening intervals in diabetes care: evaluating risk-based models for diabetic retinopathy and diabetic foot disease. The University of Edinburgh, 2025. https://hdl.handle.net/1842/43995