Purchase this article with an account.
Simon P Harding, Antonio Eleuteri, Deborah M Broadbent, Marta van der Hoek, Christopher P Cheyne, Amu Wang, Irene M Stratton, Anthony C Fisher; Individualised risk-based screening for sight threatening diabetic retinopathy - the Liverpool Risk Calculation Engine. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):2017.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a risk calculation engine (RCE) to estimate risk of progression to sight threatening diabetic retinopathy (STDR) and assign personalised screening intervals. To develop a generaliseable individualised approach to screening and assess improvements in applicability of screening interval.
Data from established digital photographic screening and primary care systems were combined in a purpose built data repository. The local ethics committee approved an opt-out approach to consent. A development dataset of candidate covariates likely to predict progression was created using a mixed qualitative/quantitative approach. <br /> After data processing a series of Markov mathematical models were fitted. Disease state (background retinopathy /mild NPDR in neither, one or both eyes) informed the baseline risk. Multiple imputation dealt with missing data. Covariates were selected using 2 steps: ranking (Wald statistic), selection (corrected Akaike Information Criterion). Alternative risk thresholds were reviewed by a patient engagement group.
Data were from 11,806 people with diabetes (46525 episodes) who were screen -ve at the first of at least 2 episodes between 20 Feb 2009 and 4 Feb 2014. 388 screen +ve events occurred. Covariates selected as of having sufficient predictive value were: duration of known disease, HbA1c, age, systolic BP, total cholesterol.<br /> Corrected C-index for the model was 0.687 and corrected AUCs (95% confidence intervals) were 0.88 (0.83-0.93) at 6 months, 0.90 (0.87-0.93) at 12 months and 0.91 (0.87- 0.94) at 24 months. A 4-way random data split gave sensitivities and specificities for a risk threshold of 2.5% at 6, 12 and 24 months respectively: 6 months 0.64-0.72, 0.93-0.94; 12 months 0.63-0.72, 0.90; 24 months 0.79-0.84, 0.81-0.82. An alternative extended Cox model was fitted to the data and gave similar results to the Liverpool RCE.<br /> Implementing personalised RCE based intervals would reduce the % of people in the dataset who become screen +ve before the allocated screening date by >50% and the overall number of screening episodes by 30%.
The Liverpool RCE has sufficient performance to allow its introduction into clinical practice within a robustly monitored environment. This approach offers the potential for an important enhancement in screening with significant improvements in performance and cost-effectiveness.
This PDF is available to Subscribers Only