Methods
Hospital (target facility)
Kitasato University Kitasato Institute Hospital (Hospital A).
Kitasato University Hospital (Hospital B).
Patients
Selection criteria
The subjects were patients with type 2 diabetes who visited Hospital A or Hospital B from January 2013 to December 2013 and continued treatment for the following 5 years until 2018.
Exclusion criteria
Patients who refused to participate in the study or had a history of any of the following diseases were excluded: angina, myocardial infarction, stroke, peripheral arterial disease, familial hypercholesterolemia, familial type III hyperlipidemia, nephrotic syndrome, renal diseases other than diabetic nephropathy, microhematuria, pre-proliferative and proliferative retinopathy, or major ocular diseases (e.g., glaucoma, dense cataract, or a history of cataract surgery).
This study was conducted in accordance with the Ethical Guidelines for Medical and Health Research Involving Human Subjects. The Kitasato University Kitasato Institute Hospital, Research Ethics Committee, approved the study (Control Number: 20051 and 20051-2) and provided permission to review patient records and use the corresponding data. The option to opt-out of the study was provided to the patients at the start of the study (2021).
Statistical analysis
We developed a prediction equation based on the Cox proportional hazard model using patient data from Hospital A [15]. The backward stepwise method was used for the selection of variables [16].
Discrimination
It is an index that evaluates how accurately the presence or absence of an event can be predicted by a prediction model. The C-statistic, which is calculated based on the receiver operating characteristic (ROC) curve, is used as a criterion for measuring the predictive accuracy [17, 18].
Calibration
It is an index to measure the degree of agreement between the prediction by the model and the actual outcome. The significance probability calculated using the Hosmer–Lemeshow test is used as the criterion for predictability. The significance level was set at 0.05 (p < 0.05) [18, 19].
External validation
We developed prediction equations using Hospital A data and then performed external validation using Hospital B data.
Outlier testing using box plots
Outlier tests with box plots were performed to reduce the impact of outliers of each risk factor on the prediction accuracy.
A box-and-whisker diagram of data for age, total cholesterol, HDL cholesterol, body mass index (BMI), urinary albumin, upper quartile (Q3), and lower quartile (Q1) was generated, and values that deviated from the range of Q1–(Q3–Q1) to Q3 + (Q3–Q1) were considered outliers [20].
Although blood pressure was measured at the time of medical examination, it was excluded from the risk factors in this study because the time of measurement varied among subjects [21].
R version 2.5.1 (http://www.r-project.org, library Design, Hmisc, ROCR) was used to determine discrimination and calibration, whereas the ROC curve, Hosmer–Lemeshow test, and box plot were used for calculation [18,19,20].