The proposal for this study was reviewed by Chinese Academy of Medical Science & School of Basic Medicine (033-2018). The insurance claims data were obtained from the MSB of Weifang. In order to protect the privacy of patients, all personally identifiable information in the database is obfuscated.

The database used in this study included all in- or outpatient claim records of patients in Weifang for 2016 to 2020 provided by the local MSB. The social health insurance system covered approximately all residents in Weifang (close to 9 million individuals) during the period from 2016 to 2020. Each claim record contains (1) basic information about the patient, including personal identification number, gender, date of birth, and type of medical insurance; (2) medical information, including hospital name and code, admission and discharge time, and diagnosis of disease; and (3) details of medical service expenses, including total expenses, out-of-pocket expenses, and reimbursement expenses.

Firstly, due to the changes in the Weifang disease coding system and the special disease codes in outpatient chronic-disease records, the initial disease codes in the database were partly based on the International Classification of Diseases, 10th Revision (ICD-10). Therefore, the diagnoses were harmonized with the ICD-10 according to the code-disease mapping table provided by the MSB of Weifang. Later, records with critical missing information, such as gender (418/21,855,530, <0.001%) or disease diagnosis (1,068,910/21,855,530, 4.89%), were deleted. The data processing is shown in Figure 1.

First, we defined diabetes (ICD-10 codes E10-E14) cases as those with at least 1 inpatient claim record or 3 outpatient claim records with a diagnosis of diabetes, considering that inpatient records are more reliable than outpatient records [16]. In this database, we found that all patients defined as having diabetes had more than 3 outpatient records, reflecting the regularity of medical contact.

Second, we defined incident diabetes cases. The procedure to identify incident cases was to exclude cases with any in- or outpatient records of diabetes in the given LP, before the index year. Different LPs may classify patients as having different statuses. Two scenarios for assumed different LPs are shown in Figure 2. Scenario 1 (LP1) shows that patient B will be identified as an incident case but patient A will not. Scenario 2 (LP2) shows that with the extension of the LP, more records of patient B might be caught, but patient A and patient B will both be identified as prevalent cases. In order to use as long an LP as possible in this study, we used 2020 as the index year and a 4-year LP (2016-2019) as the longest LP.

Two methods were developed to identify the optimal LP. First, consistency examination indexes (CEIs) were used to determine the level of agreement between different LPs from a quarter of a year to 4 years. Second, a retrograde survival function method was used. Finally, the optimal LP was selected after considering the results from the above approaches.

The CEIs include the κ value, positive predictive values (PPVs), and overestimation value. The relative formulas have been reported in other papers [4,15]. The κ value is categorized as follows: values between –1 and 0 represent “no agreement”; between 0 and 0.20 represent “poor agreement”; between 0.21 and 0.40 represent “slight agreement”; between 0.41 and 0.60 represent “fair agreement”; between 0.61 and 0.80 represent “good agreement”; between 0.81 and 0.90 represent “very good agreement”; and between 0.91 and 1.0 represent “excellent agreement” [17]. The PPVs are used to estimate the probability of identifying a true incident case among those identified as being new cases [18], calculated using the 4-year LP as the reference time. The overestimation of incident cases decreased with extension of the LP, which reflects the impact of the length of the LP impacts on the number of incident cases. All the CEIs mentioned above were calculated at quarter-year intervals during the first year of the LP and 1-year intervals after the first year, since the number of identified incident cases varied greatly within the first year of the LP in previous studies [7,12,16].

With the retrograde survival function method, survival (ie, being an incident case) means not having a preceding record at the given time. In detail, for each record in this study, let a represent the date of the last record of diabetes of a patient in 2020, b the date of the most recent previous record in the LP, and c the first date of the LP. The retrograde survival time is (a–b) for patients with preceding records. Patients without preceding records during the LP, indicating survival, are defined as censored cases, with the censored time (a – c). Based on the survival time and censored time, we established the survival function. Survival probability, S(t), the probability of surviving at a specific time, is represented by the equation S(t)=P (T>t), where T is the year of the LP, calculated using the Kaplan-Meier estimator. Then, we used the following 2 approaches to determine the optimal LP based on the survival function. For the first approach, we constructed a hazard function. The hazard function represents the instantaneous probability of having the end point event during (t,∆t) conditional on survival up to t or later. In actuarial terms, it is calculated by , where d_∆t is the number of patients with the previous record during (t,∆t) and n_t is the patient at risk during the interval. Since the retrograde hazard of diabetes will decrease and converge to zero and the survival function will converge to the “true” probability of being incident, we assumed that a time point (tf) exists where . At this point (tf), all cases with preceding records have been found, and the remaining cases are risk-free; this is exactly the optimal LP. To estimate (tf), for practical purposes, we defined (tf) as the first time point when h_t ≤0.0005. For the second approach, we counted the frequency of S(t) corresponding to each day calculated by the retrograde survival function. We assumed that there was a stable S(t), where the frequency was high and increased significantly, representing that S(t) stays at S(t_f) relatively constantly over a long time and that the change in S(t) approaches zero. The t_f will be considered as the point when S(t) first reaches S(t_f).

Statistical analyses were performed using SAS (version 9.4; SAS institute).

In our study, a 2-year LP had good CEIs with excellent κ agreement (0.950>0.9), a high PPV (92.2%) and a lower than 10% overestimation rate. As for the retrograde survival function, it showed that a 2-year LP can reliably distinguish new cases in a prevalent pool. According to the results of the methods above, we confirmed that the optimal LP for identifying incident cases of diabetes using the Chinese insurance database provided by the MSB of Weifang is around 2 years, and all the methods are effective and stable.

In our study, there were 109,467 diabetes cases in 2020 before LP adjustment. After using the 2-year LP, the number of incident cases decreased to 21,472, which means this adjustment is important. In addition, our study showed that 84.74% (18,192/21,472) of incident cases were between the ages of 45 to 79 years, indicating that the middle-aged and older population has a high incidence of diabetes. We should make efforts to focus on this population to prevent diabetes.

Our study is the first to identify 2 years as the optimal LP for diabetes in the Chinese population when using insurance data, which supports efforts to estimate the incidence of diabetes among populations using insurance data. More importantly, we used 2 different methods to identify the optimal LP, a method that could be referred to when analyzing LPs for other diseases in administrative data.

Our study has some limitations. In our database, without more detailed records for disease classification, it was not possible to differentiate between type 1 and type 2 diabetes. However, there are far fewer patients with type 1 diabetes than with type 2 diabetes in the overall population. Thus, we considered the study population to be generally homogeneous and that these results may be representative and stable. Furthermore, we only applied up to a 4-year LP. Although the results of the retrograde survival function analysis showed stabilization of risk of being a preceding case after a 2-year LP, we expect to use longer LPs in future studies with more accumulated data.

All CEIs (the κ coefficient, PPV, and overestimation rate) and the retrograde survival function method can effectively identify the optimal LP for diabetes incident cases, with stable and constant results. We found that a 2-year LP was optimal to remove preceding diabetes cases and identify the true incidence of diabetes using medical insurance data in the Chinese population.