Retinopathy prediction in type 2 diabetes: Time-varying Cox proportional hazards and machine learning models
Issued Date
2023-01-01
Resource Type
ISSN
23529148
Scopus ID
2-s2.0-85161337257
Journal Title
Informatics in Medicine Unlocked
Volume
40
Rights Holder(s)
SCOPUS
Bibliographic Citation
Informatics in Medicine Unlocked Vol.40 (2023)
Suggested Citation
Looareesuwan P., Boonmanunt S., Siriyotha S., Lukkunaprasit T., Thammasudjarit R., Pattanaprateep O., Nimitphong H., Reutrakul S., Attia J., McKay G., Thakkinstian A. Retinopathy prediction in type 2 diabetes: Time-varying Cox proportional hazards and machine learning models. Informatics in Medicine Unlocked Vol.40 (2023). doi:10.1016/j.imu.2023.101285 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/85143
Title
Retinopathy prediction in type 2 diabetes: Time-varying Cox proportional hazards and machine learning models
Other Contributor(s)
Abstract
Background: Diabetic retinopathy (DR) is one of the most common complications in type 2 diabetes (T2D) with an estimated prevalence of 22%. Predictive modelling has largely been dependent on Cox proportional hazards (CPH) with assumptions of linearity and constant hazards. Machine learning (ML) approaches may prove advantageous in more adequately capturing non-linear effects. Objective: To construct and compare DR prediction models using CPH and ML models with time-varying covariates. Design: Real-world, retrospective cohort study. Setting: A tertiary care hospital in Thailand. Participants: Data on 48,622 T2D patients from electronic health records between 1st January 2010 and 31st December 2019. Methods: Time-to-event time-varying models that included 13 variables were trained in diabetic retinopathy prediction. The CPH and ML models were compared using left-truncated right censoring relative risk forest (LTRC-RRF) and left-truncated right censoring conditional inference forest (LTRC-CIF) algorithms. Results: The CPH model outperformed both ML approaches with a Harrell's C-index (c-index) of 0.70 compared to c-indices of 0.51–0.57 for the ML models in the test dataset. Both CPH and ML models showed insulin use and the presence of chronic kidney disease increased DR risk. Sodium glucose transporter 2 inhibitors and dyslipidemia were associated with reduced DR risk. Conclusion: CPH provided better predictive power for DR risk than ML modelling using real world data. The presence of comorbidities and the use of antidiabetic medications were associated with the greatest drivers of DR risk.