Data used in this study were obtained from the China Health and Retirement Longitudinal Study (CHARLS), a nationally representative longitudinal cohort initiated by Peking University in 2011. The study targets the middle-aged and older adults population in China and employs a multi-stage, stratified cluster sampling strategy. Survey content includes psychological health, chronic disease status, and socioeconomic indicators of participants (14–16). To date, five Waves of data have been collected: 2011 (Wave 1), 2013 (Wave 2), 2015 (Wave 3), 2018 (Wave 4), and 2020 (Wave 5), with response rates consistently exceeding 80%. Ethical approval for data collection was granted by the Biomedical Ethics Review Committee of Peking University (IRB00001052-1015), and all participants provided written informed consent. The CHARLS dataset is publicly available at http://charls.pku.edu.cn/en.

For this study, we used data from Wave 3 (2015) and Wave 4 (2018). A total of 6,187 participants were included after applying inclusion and exclusion criteria. The detailed screening process is presented in Figure 1. Only participants aged 45 years and above were included, consistent with the CHARLS study design.

The FI used in this study was constructed based on the deficit accumulation model proposed by Rockwood and Mitnitski (5), and further refined using the practical framework introduced by Theou et al. (17). This version of the FI is widely used in geriatric research due to its flexibility, strong predictive validity, and applicability across different datasets and populations. This scale mainly includes 37 variables, including 5 aspects (Activities of daily living Instrumental activities of daily living, Physical functional limitations, Chronic disease, Mental health), These variables can reflect the “deficits” of participants' physical functions from different aspects. According to Theou et al. (17), a frailty index with 30 or more variables across multiple domains ensures robustness and validity. Therefore, 37 variables were chosen based on data availability and domain coverage (17). In addition, the selection of variables was guided by previous FI studies using CHARLS and other large aging cohorts, ensuring comparability and methodological consistency.

We constructed a 37-item Frailty Index (FI) following the deficit accumulation approach, as shown in Supplementary Table S1. Each item was recoded such that “0” indicated no deficit and “1” indicated the presence of a deficit. For variables with ordinal responses (e.g., mental health items), intermediate values of 0.33 and 0.67 were used to reflect partial deficits based on the degree of symptom severity or frequency. Positively framed items such as “feel happy” were reverse-coded to maintain consistency across all items.

The FI score for each participant was calculated as the sum of non-missing item scores divided by the total number of items (n = 37), resulting in a continuous variable ranging from 0 to 1. FI values were computed separately for 2015 and 2018. The difference between FI2018 and FI2015 was then used to quantify frailty progression. Participants with an FI increase >0.1 were classified as having rapid frailty progression (Frailty Progress Rapid, FPR), while others were classified as having non-rapid progression (Non-Frailty Rapid Progression, NFRP). These groups are hereafter referred to as the Frailty Progress Rapid Group (FPRG) and the Non-Frailty Rapid Progression Group (NFPRG).

We defined rapid frailty progression (FPR) as an increase in FI >0.1 between 2015 and 2018. This threshold was chosen to reflect clinically meaningful changes in frailty status over a 3-year period. While some prior studies have modeled frailty progression as a continuous outcome [e.g., (18, 19)], or have considered any increase in frailty scale scores as indicative of progression [e.g., (20)]. Our use of a fixed threshold enabled binary classification and clearer group comparisons.

To ensure the robustness of this threshold, sensitivity analyses were conducted using alternative cutoffs (ΔFI > 0.05 and ΔFI > 0.15), and the associations between temperature and frailty progression remained consistent across these definitions, as shown in Supplementary Table S2.

Meteorological data were obtained from the China Meteorological Administration's Land Data Assimilation System (CLDAS v2.0), integrating multi-source observations with high-resolution spatial (0.0625° × 0.0625°) and temporal (hourly) precision. We acquired 4-year daily and annual average temperature records (2015–2018) for participants' residential cities through standardized measurements at municipal meteorological stations. Three temperature metrics were derived: (a) the lowest daily average temperature (TLDAT) across the study period, (b) the highest daily average temperature (THDAT), and (c) the multi-year average annual temperature (AAT). Data completeness exceeded 95%, with limited gaps (<5%) addressed via spatial interpolation from adjacent stations. Geospatial distribution patterns of these thermal metrics across China are illustrated in Supplementary Figures S1–S3, reflecting substantial regional climatic heterogeneity.

The covariates included in the regression models were derived from participants' 2015 data and covered three domains: (1) Demographic and socioeconomic factors: age, sex, body mass index (BMI; overweight vs. non-overweight), education level (literacy vs. illiteracy), marital status (married vs. others), place of residence (urban vs. rural), smoking status (yes/no), and alcohol consumption (yes/no); (2) Routine blood parameters: white blood cell count (WBC)and platelet count (PLT); (3) Lipid metabolism indicators: total cholesterol (TC), low-density lipoprotein cholesterol (LDL-C), and triglycerides (TG).

In this study, participants were classified as illiterate or literate depending on whether they had completed formal schooling. Marital status was divided into married and others based on whether the respondent was currently married and cohabiting. Residence was categorized as urban or rural based on the administrative classification recorded in the CHARLS dataset. BMI was grouped into overweight (>24) and non-overweight (≤ 24) categories according to Chinese criteria.

Participant characteristics were summarized using appropriate descriptive statistics. Continuous variables were assessed for normal distribution using Shapiro-Wilk tests, with results presented as mean ± standard deviation for parametric data or median (interquartile range) for non-parametric distributions. Categorical variables were reported as frequencies and percentages. Temperature metrics (TLDAT, THDAT, AAT) were initially analyzed as continuous variables, with subsequent quartile categorization (Q1–Q4) applied to variables demonstrating univariate significance (P < 0.05). Group comparisons between Frailty Progression Rapid Group (FPRG) and Non-Frailty Rapid Progression Group (NFPRG) employed Student's t-tests for normally distributed continuous variables, Wilcoxon rank-sum tests for skewed data, and χ² tests for categorical variables, as detailed in Table 1.

Logistic regression models evaluated associations between temperature metrics and frailty progression risk (as shown in Table 2), calculating adjusted odds ratios (ORs) with 95% confidence intervals (CIs). Three sequential models were developed: (1) unadjusted model; (2) demographic-adjusted model (age, sex, BMI, residence); (3) fully adjusted model incorporating demographic, clinical, and laboratory covariates. Restricted cubic spline regression with three knots (25th, 50th, and 75th percentiles) was employed to explore nonlinear relationships between continuous temperature exposures and outcomes, with the resulting curves illustrated in Figure 2.

Stratified analyses examined effect modification by sex, BMI status, drink, smoke, education level, marital status, and urban-rural residence. Interaction terms were incorporated into adjusted models, with likelihood ratio tests evaluating subgroup heterogeneity. Sensitivity analyses tested the robustness of the 0.1 FI change threshold using alternative cutoffs (ΔFI >0.05 and >0.15), confirming consistent temperature effects across definitions.

All statistical analyses were performed using R version 4.1.2 (R Core Team, 2021) and SPSS version 26.0(IBM Corp., Armonk, NY, USA). A P-value < 0.05 was considered statistically significant, including adjustments for multiple comparisons where applicable (21, 22).

Our study included 25,504 participants who underwent the baseline (2015) survey. Considering the study objectives, we excluded the following participants: (1) those lacking age information and those under 45 years of age; (2) those lacking information related to FI and other informations at baseline; (3) those who were not followed up;4) those lacking information related to FI and other informations during follow-up surveys. Detailed inclusion and exclusion processes are shown in Figure 1.

To assess potential selection bias, we compared key baseline characteristics between included (n = 6,187) and excluded (n = 19,317) participants, as shown in Supplementary Table S3. Although some variables showed statistically significant differences (P < 0.05)—likely due to the large sample size—their standardized mean differences (SMDs) all remained below 0.1, indicating that the magnitude of these differences was minimal. Consequently, any residual selection bias is expected to be small, and our included sample broadly represents the target population.

A comparative analysis of baseline characteristics (Table 1) revealed that participants in the Frailty Progress Rapid Group (FPRG) were generally older (median age 62 vs. 59 years, P < 0.001, SMD = 0.254) and had a higher proportion of females (59.34 vs. 49.88%, P < 0.001, SMD = 0.191). Although the median BMI was only slightly higher in FPRG (P < 0.001, SMD = 0.019), total cholesterol (P = 0.010, SMD = 0.094) and triglycerides (P = 0.022, SMD = 0.083) were also elevated compared to NFPRG. Additionally, FPRG showed a greater percentage of illiterate individuals (71.41 vs. 60.28%, P < 0.001, SMD = 0.236) and had slight but significant differences in drinking and smoking status (P = 0.03 and P = 0.04, respectively). In contrast, no statistically significant differences were observed for white blood cell count, platelet count, or LDL levels.

From an environmental standpoint, FPRG participants tended to live in regions with significantly lower average annual temperatures (AAT, P < 0.001, SMD = 0.139), highest daily average temperatures (THDAT, P = 0.007, SMD = 0.094), and lowest daily average temperatures (TLDAT, P < 0.001, SMD = 0.129), suggesting that cooler climatic conditions may contribute to accelerated frailty progression.

These findings collectively indicate that older age, female gender, higher lipid profiles, and residence in cooler regions are associated with rapid frailty progression, supporting the notion that both biological and environmental factors can significantly influence frailty trajectories in older adults.

Our multivariate analysis revealed significant associations between some thermal exposure metrics and frailty progression (Table 2). After adjusting for all covariates, each 1 °C increase in annual average temperature (AAT) was associated with a 12% reduction in the odds of rapid frailty progression (OR = 0.88, 95%CI: 0.83–0.94), and higher minimum temperatures (TLDAT) likewise conferred a protective effect (OR = 0.91 per 1 °C, 95%CI: 0.86–0.96). Quartile-based analyses for AAT and TLDAT also showed statistically significant, nearly linear trends (P-trend < 0.01 for both).

In contrast, for THDAT, the quartile-based trend test was not statistically significant (P-trend > 0.05), suggesting no clear linear dose-response across THDAT quartiles. However, restricted cubic spline (RCS) models (Figure 2) revealed a significant non-linear relationship (P-nonlinear = 0.008), characterized by a J-shaped curve and a turning point at ~31.8 °C. Below this threshold, each 1 °C increase was associated with a 9% lower risk of frailty progression (OR = 0.91, 95%CI: 0.87–0.96), but above 31.8 °C, the frailty risk rose by 13% per 1°C increase (OR = 1.13, 95%CI: 1.05–1.21). These findings indicate that older adults experience a protective effect of moderate warmth, yet extreme heat may sharply exacerbate frailty progression.

We conducted subgroup analyses to examine whether the association between temperature and frailty progression was consistent across different populations (Supplementary Figures S4–S6). Stratified analyses were performed based on gender, BMI, marital status, residence, drink, smoke, and education level.

The results indicated that the negative associations between AAT and TLDAT and frailty progression remained consistent across most subgroups, with all interaction terms showing P > 0.05. This suggests that the effects of these temperature variables on frailty progression were stable across different groups. However, the interaction terms for education level and place of residence with THDAT were significant (P < 0.05). Individuals with lower education levels or those living in rural areas were more vulnerable to extreme temperatures, likely due to limited access to climate-responsive resources, while urban populations were less affected.

These findings highlight that the impact of temperature on frailty progression is modulated by factors such as BMI, education level, and residence. Public health interventions should consider these factors to better mitigate the adverse effects of temperature extremes.