In this analysis, participants were derived from the ongoing Kailuan Study, which was a population-based prospective cohort study conducted in the Kailuan community of Tangshan, China. The study design has been described in previous publications [22,23,24]. In brief, the study enrolled 101,510 adult participants (81,110 men and 20,400 women) during 2006 to 2007. Furthermore, between 2008 and 2009, an additional 24,540 adult participants (19,356 men and 5,184 women) were enrolled. Participants completed questionnaires, and underwent physical measurements, clinical evaluations, and lab tests were performed in a single measurement at baseline. Subsequently, same assessments were conducted biennially. The study obtained approval from the ethics committees of Kailuan General Hospital, and all participants provided written informed consent.

This study investigated the relationship between physical activity levels and risk of T2D in the prediabetes population. Physical activity and other covariates of participants were collected from the 2014 survey (as baseline of this study) of the Kailuan Study. Participants included in present analyses were followed until December 31 2018, the average follow up time was 3.6 years. The fasting blood glucose (FBG) values were collected twice, at 2014 survey and end of the follow up period. In 2014, out of 101,588 participants, 1,718 lacked fasting plasma glucose data, 31,567 had no physical activity data, 52,626 did not meet the prediabetes criteria for FBG levels, 2,649 were lost to follow-up, and 604 deaths were excluded (Supplementary Fig. 1). Finally, a total of 12,424 participants were eligible for analysis.

Physical activity information of participants was collected once during 2006–2009 by using the International Physical Activity Questionnaire-Short Form [25]. Participants were asked how many days per week they did exercise; what kind of physical activity they performed; and how long each exercise session lasted (< 30 min; 30–60 min; ≥60 min). We categorize MET intensity into three levels: high (8 METs), moderate (4 METs), and low (3.3 METs). The formula is: weekly exercise days × daily exercise duration (minutes) × MET intensity = weekly total MET minutes. Add up all weekly MET minutes to get the total MET value. After collecting frequency and intensity of the weekly physical activity, participants were categorized into three groups based on their weekly MET value: low (< 600 MET-minutes per week), moderate (600–3000 MET-minutes per week), and high (≥ 3000 MET-minutes per week). These categories represented different levels of physical activity engagement, as detailed elsewhere [26, 27].

The primary outcome of our study is the incidence of T2D among prediabetic participants during the follow-up period. Prediabetes is defined as FBG level range between 5.6 and 6.9 mmol/L according to the guidelines of American Diabetes Association [28]. T2D is defined as meeting any of the following criteria: FBG ≥ 7.0 mmol/L, a self-reported physician diagnosis, or self-reported use of anti-diabetic medication. The secondary outcome is the change in FBG levels between the baseline (2014 resurvey) and final FBG tests (2016 or 2018 resurvey) among participants from different physical activity groups. Metabolic syndrome was diagnosed by three or more of the following criteria: TG ≥ 150 mg/dl (1.7 mmol/L), HDL-C < 40 mg/dl (1.03 mmol/L) in men and < 50 mg/dl (1.29 mmol/L) in women, fasting glucose ≥ 100 mg/dl (5.6 mmol/L) or previously diagnosed with type 2 diabetes, blood pressure ≥ 130/85 mmHg or on antihypertensive medication, and central obesity (defined as waist circumference ≥ 90 cm in men and ≥ 80 cm in women, according to the ethnic criteria for Asians) [29, 30].

Demographic information (e.g., age, sex), socioeconomic status (e.g., income), family history of diabetes, and lifestyle factors (e.g., smoking status, alcohol consumption, sleep duration, physical activity, and dietary intake) were collected through a structured questionnaire. Education level was categorized as middle school and below, high school, or college and above. Average monthly income of each family member was classified as < 1,000 Yuan, 1,000–3,000 Yuan and ≥ 3,000 Yuan. Smoking status was defined as either a current smoker or never having smoked. Alcohol consumption was defined as either a current drinker or never having drunk.

Height and weight measurements were taken by trained nurses under standardized conditions, with participants wearing light clothes and being barefoot. Heights were measured to the nearest 0.1 cm, and weights were measured to the nearest 0.1 kg. Body mass index (BMI) was calculated as weight in kilograms divided by height in meters squared. Waist circumference was also measured at standing position. Blood pressure was measured in the seated position and the average of three readings was recorded as systolic blood pressure (SBP) and diastolic blood pressure (DBP). Hypertension was defined as SBP ≥ 140 mm Hg or DBP ≥ 90 mm Hg, the use of antihypertensive medication, or a self-reported history of hypertension. Blood samples were collected after an 8 to 12-hours fasting period. FBG levels were measured using the Hexokinase/Glucose-6-phosphate dehydrogenase method (Mind Bioengineering Co., Ltd., Shanghai, China), with an upper detection limit of 30.07 mmol/L. Triglyceride (TG), total cholesterol (TC), low-density lipoprotein-cholesterol (LDL-C), and high-density lipoprotein-cholesterol (HDL-C) levels were measured by automatic analyzer (Hitachi 747, Hitachi, Tokyo, Japan).

Continuous, normally distributed data were summarized as mean and standard deviations; while categorical data were summarized as the numbers and percentages. For 12,424 prediabetes participants included in the analysis, their demographic information (sex, age, education, income, marital status), behavioral status (smoking status, sleep duration, TV hours), and biochemical characteristics (FBG, DBP, and SBP) were described according to three levels of physical activity. For each individual, the person-time of follow-up was calculated from the 2014 survey date until the occurrence of T2D, being lost to follow-up, death or end of follow-up (December 31, 2018), whichever happened first. Supplementary Table 1 shows baseline characteristics for all participants and participants with prediabetes in the 2014 resurvey.

Cox proportional hazard models were utilized to determine hazard ratios (HR) and 95% confidence intervals (CIs) for the incidence of T2D in the three physical activity groups (low, moderate, and high), using the low physical activity group as reference. According to Schoenfeld residuals, our models satisfied the proportional assumption criteria. In Model 1, age and sex (men, women) were adjusted as covariates. Model 2 was further adjusted for smoking status (current or never), alcohol consumption (current drinker or never drink), sleep hours (< 5, 5–7, or ≥ 7 h/day), TV time (< 2, 2–3, or ≥ 4 h/day), central obesity (yes or no), family history of diabetes (yes or no), education level (middle school and below, high school, or college and above), average monthly income of each family member (< 1000, 1000–3000, or ≥ 3000 Yuan), marital status (married or single), dietary approaches to stop hypertension (DASH) diet score (≤ 25, 26–30, or ≥ 31 points), and BMI categories (< 24, 24 ≤ BMI < 28, or ≥ 28 kg/m2). In model 3, systolic blood pressure, LDL-C, and TG were further adjusted for as the full model. Supplementary Table 2 indicated how DASH score was classified. Sensitivity analyses were conducted after excluding participants with conditions such as cancer, cerebral hemorrhage, cerebral infraction, subarachnoid rainbow, ankylosing spondylitis, rheumatoid joint disease, gout, and Parkinson’s disease to test the consistency and stability of the results. Stratified analyses were performed based on sex, age, BMI, smoking status, alcohol consumption, nighttime sleeping hours, metabolic syndrome, hypertension, and DASH diet score. A likelihood ratio test was applied to examine the significance of interactions.

For FBG changes, mixed-effects liner models were used to analyze the effect of group (physical activity levels), time, and group by time interaction with adjusting for age, sex, and baseline FBG. In the analysis, individual participant was treated as random effects, while the group, times, and their interaction were considered as fixed effects. Statistical analysis was performed using R software (version 4.2.2) and SAS version 9.4 (SAS Institute, Cary, NC, USA). The level of significance was taken as two-sided P < 0.05.