U-Shaped Link of Health Checkup Data and Need for Care Using a Time-Dependent Cox Regression Model with a Restricted Cubic Spline

Study design

We employed a community-based retrospective cohort study to explore the risk indicators for certified long-term care needs and mortality among older adults. We integrated and analyzed two sets of data received from Kitanagoya City, Aichi Prefecture, Japan. The first dataset comprises long-term care information on certification, mortality, and transfers out of Kitanagoya City from April 2011 to September 2020. The second dataset is health checkup data comprising specific health checkup information and late-stage health checkup data from 2011 to 2019. Figure 1 shows the selection of the study participants. Using the second dataset, we first confirmed that 6527 residents of Kitanagoya City in Aichi Prefecture participated in at least one health checkup between April 1, 2011 and March 31, 2012. Then, we applied the following exclusion criteria to the 6527 participants: (1) age and gender unknown; (2) lost insured status owing to death or moving out of Kitanagoya City by April 1, 2012; (3) certified as needing assistance or care by April 1, 2012; and (4) under age 65. The final sample included 3,718 persons aged 65 years or above as of April 1, 2012, who had undergone health examinations between April 1, 2011 and March 31, 2012. Participants were eligible for follow-up from April 1, 2012 to September 30, 2020. Those newly certified as needing long-term care and those who had died were identified from the first dataset.

The study protocol adhered to the guidelines of the Declaration of Helsinki. This study was approved by the Ethics Review Committee of Nagoya University Graduate School of Medicine (No. 2019-0100) and was conducted under contract with Kitanagoya City, Aichi Prefecture. As the data used in this study were anonymized and non-personalized by Kitanagoya City, Aichi Prefecture, which is the data management authority, the participants’ informed consent is not necessary based on the “Ethical Guidelines for Medical and Health Research Involving Human Subjects” in Japan²⁵.

Outcome variables

In this study, the outcome was whether the older adult was certified as requiring Level 1 or higher nursing care, or had died. Follow-up was terminated if the participant had died or moved out of the city, or was certified as needing long-term care. The outcome variables were obtained as time-to-event data.

To use the long-term care insurance system, one must apply for a long-term care (support) certification and undergo a two-stage assessment to determine the level of care required. The first is a computerized examination of 74 survey items (primary assessment), followed by an examination (secondary assessment) by the Long-Term Care Certification Examination Board established by the municipality based on the primary assessment results, the attending physician's written opinion, and other materials⁵. We focused only on the certification of the need for long-term care (Care Need Levels 1 to 5). If the participant was certified as requiring Level 1 or higher care in the secondary assessment, the date of this secondary assessment was used as the date of certification.

As the death data dates obtained from Kitanagoya City were monthly, the last day of the month was used as the date of death.

Health checkup data

The health checkup data consist of specific medical checkup and late-stage health checkup information from 2011 to 2019. Of the health checkup items, those items consistently obtained for all participants in all years were included in the analysis. The health checkup data include basic attributes, test items of the specific health checkups and health checkups for specific older adults, and the participant’s questionnaire responses. The basic attributes include age and gender. Anthropometric information includes height, weight, and BMI. Clinical laboratory information includes blood pressure (systolic blood pressure or SBP and diastolic blood pressure or DBP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), γ-glutamyl transpeptidase (γ-GTP), HbA1c content, and serum lipid profile. Tests for glucose and protein in urine were also performed. HbA1c was measured by the Japanese Diabetes Society until 2012 and by the National Glycohemoglobin Standardization Program from 2013. The National Glycohemoglobin Standardization Program values were converted to Japanese Diabetes Society values and used for further analysis. Information on smoking status, drinking status, and disease history was obtained through a questionnaire.

Diabetes mellitus, hypertension, and dyslipidemia were defined, respectively, as (1) HbA1c of ≥ 6.1%, or treatment with blood glucose-lowering drugs; (2) SBP of ≥ 140 mmHg, DBP of ≥ 90 mmHg, or treatment with antihypertensive drugs; and (3) high-density lipoprotein cholesterol (HDL-cho) concentration of < 40 mg/dL, low-density lipoprotein cholesterol (LDL-cho) concentration of ≥ 140 mg/dL, triglyceride concentration of ≥ 150 mg/dL, or treatment with antidyslipidemic drugs.

Statistical analyses

The distributions of continuous variables at the baseline were compared between men and women participants using the Student’s t-test or Wilcoxon rank sum test. The distributions of categorical variables at the baseline were compared between men and women using Fisher’s exact test.

Crude rates of certified care needs and death are shown as the number of certified care need cases and deaths per 1,000 person-years, respectively. The crude rates were compared between men and women using the Wald test.

For continuous clinical variables, a time-dependent Cox regression analysis²⁶ was performed to assess the association of each clinical parameter with certified care need and death, using data measured at the baseline and over several years of follow-up. For the Cox regression, we used the time to the first certification of need for care or the time to death as time-to-event data. The level of care required was not considered in the analysis of nursing care needs. Gender and age were included in the model as potential confounders. Furthermore, the same analysis was performed by adding treatment with blood glucose-lowering drugs (yes = 1, no = 0), treatment with antidyslipidemic drugs (yes = 1, no = 0), treatment with antihypertensive drugs (yes = 1, no = 0), smoking status (habitual smoker: yes = 1 vs. no = 0), and drinking status (habitual or chance drinker = 1 vs. non-drinker = 0) to the model as potential confounders, in addition to age and gender. Given that the distribution of triglyceride, ALT, AST, γ-GTP, and HbA1c levels was skewed, the values were log₂-transformed (Supplementary Fig. S1). Cox regression analysis was compared by building both linear and nonlinear models. In the case of the nonlinear model, a restricted cubic spline was applied. Details of the analysis conditions are described in Supplementary Methods. Statistical significance in the nonlinear model was tested by comparing the spline and null models, while that for nonlinearity was tested by comparing the spline and linear models. A p-value < 0.05 for the test of nonlinearity and a p-value < 0.05 for the spline model depict a statistically significant nonlinear relationship between the clinical parameter and the event, and a spline model is adopted. Conversely, a linear model is adopted for a p-value ≥ 0.05 for the test of nonlinearity and a p-value < 0.05 for the linear model. Values of the Akaike information criterion (AIC) were compared between linear and nonlinear models to assess the quality of each model.

For categorical clinical variables, we performed a time-dependent Cox regression model to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association of each clinical variable with a certified need for long-term care and death data, measured at the baseline and over several years of follow-up. One of the categories—for each categorical parameter—was used as the reference category and the HR for the remaining categories was estimated. Gender and age were included in the model as potential confounders.

All statistical analyses were performed using R version 3.6 (http://www.r-project.org/). A p-value under 0.05 was considered significant. We used the survival package²⁷ and the rms package²⁸.