The protocols of MIND-China and MRI sub-study were reviewed and approved by the Ethics Committee at Shandong Provincial Hospital in Jinan, Shandong, China. Written informed consent was obtained from all participants or proxies.

We used data derived from the baseline assessments of the Multimodal INtervention to delay Dementia and disability in the rural China (MIND-China) that is a participating project in the World-Wide FINGERS Network, as described previously (11, 12). Briefly, the MIND-China study targeted people aged 60 years and older and living in 52 rural communities of Yanlou Town, Yanggu County, western Shandong province, China. From March to September 2018, 5,765 participants were examined for MIND-China. Of them, 1,304 participants from 26 villages randomly selected from all the 52 villages accomplished the structural brain MRI scans.

All participants underwent the brain MRI scans either on the Philips Ingenia 3.0 T MR System (Philips Healthcare, Best, The Netherlands) in Southwestern Lu Hospital or on the Philips Achieva 3.0 T MR System (Philips Healthcare, Best, The Netherlands) in Liaocheng People’s Hospital. The MRI sequences included the sagittal 3D sT1-weighted, axial T2-weighted, sagittal 3D Fluid-attenuated inversion recovery (FLAIR) images, and axial susceptibility weighted imaging (SWI). The detailed parameters of the core MRI sequences have been reported previously (13).

We assessed the following four MRI markers for CSVD. Cerebral microbleeds (CMBs) were focal, rounded hypodense lesions measuring less than 5 mm in diameter on SWI. The CMBs was quantitatively acquired by AccuBrain® (BrainNow Medical Technology Ltd., Shenzhen, Guangdong, China) as described previously (14, 15). Briefly, CMBs were detected on SWI images via a fully connected network that was trained by deep learning technique. For a given SWI image, the network showed CMB location by exporting a probability map.

The visual assessment of lacunes, WMH, and PVS was performed by two well-trained raters (M.Z. for EPVS and J.W. for lacunes and WMH) according to the standards for reporting vascular changes on neuroimaging 1 (STRIVE-1) (2). The raters were blinded to participants’ clinical data and under the supervision of an experienced clinical neurologist (L.S) and an experienced neuroradiologist (T.G.). Lacunes were rounded or ovoid lesions, 3–15 mm in diameter, generally in the intensity of cerebrospinal fluid (CSF) signal on T2 and FLAIR sequences. The trained rater (J.W.) counted lacunes in each hemisphere on FLAIR sequence, and then added up the numbers of lacunes in bilateral hemispheres. WMH was defined as symmetrical hyperintense on T2 images in the brain white matter, and was evaluated in periventricular and deep white matter region according to the Fazekas scale (16). We defined periventricular WMH (PWMH) and deep WMH (DWMH) following the “continuity to ventricle” rule (17). The rater (J.W.) evaluated WMH on the slice with the most severe white matter lesions. We defined EPVS as small (<3 mm) punctate (if perpendicular) or linear (if longitudinal to the plane of scan) hyperintensities on T2 images. We rated EPVS using the validated semiquantitative scale (18). Briefly, the rater (M.Z.) visually counted bilateral EPVS in basal ganglia (BG) and centrum semiovale (CSO) on all slice, and categorized EPVS in BG and CSO according to the highest counts on the slice and hemisphere with the most EPVS.

Six months after the initial assessment, the rater re-assessed MR images of 200 randomly selected subjects for lacunes and WMH, which yielded the weighted ĸ statistic of 0.84 for lacunes, 0.86 for DWMH, and 0.89 for PWMH. Similarly, three months after the initial assessment, the rater re-evaluated EPVS in MRI images of 30 randomly selected subjects, which yielded a weighted ĸ statistic of 0.75 for BG-EPVS and 0.74 for CSO-EPVS.

Total CSVD score was assessed as previously reported (19). One point was assigned for the presence of: (a) lacunes; (b) CMBs; (c) moderate-to-severe EPVS (>10) in BG; and (d) DWMH (Fazekas score 2–3) or PWMH (Fazekas score 3). We dichotomized the total CSVD score into the absence of CSVD (total CSVD score = 0) and the presence of CSVD (total CSVD score ≥ 1) (7).

In March–September 2018, extensive data were collected by trained staff through face-to-face interviews, clinical examinations or laboratory tests. According to the literature (5–8, 10, 20), we selected the following candidate risk factors for CSVD: age, sex, education, body mass index (BMI), smoking, alcohol drinking, high blood pressure, pulse pressure, history of cerebral infarction, fasting blood glucose (FBG), low-density lipoprotein (LDL), WBC count, NC, NLR, and SII. Education was dichotomized into low education (primary school and below) and high education (middle school and above). Alcohol consumption and smoking status were categorized as current, former, and never drinking or smoking, respectively. Arterial pressure was measured as described previously (21), and high blood pressure was defined as systolic pressure ≥ 140 mmHg or diastolic pressure ≥ 90 mmHg. FBG and LDL was measured using an automatic biochemical analyzer (DIRUICS-600B; DIRUI Corporation, Changchun, China) (22). The automated blood cell analyzer (BC1800, Mindray Corporation, Shenzhen, China) was used for routine blood tests (12). NLR (neutrophil count/lymphocyte count) and SII (platelet count × neutrophil count/lymphocyte count) were calculated based on the absolute NC (×10⁹/L), lymphocyte count (×10⁹/L), and platelet count (×10⁹/L).

We defined clinical risk factors (i.e., high blood pressure, high pulse pressure, obesity, and high FBG) according to current clinical criteria, as described above. For laboratory measurements (e.g., LDL, WBC, NC, NLR, and SII), abnormal values were defined according to the optimal cutoff points determined using the receiver operating characteristic (ROC) analyses and Youden index (Supplementary Table S1).

We performed the analyses among participants with complete data. We presented frequencies (%) for categorical variables and the median (interquartile range, IQR) for continuous variables. The Chi-square test or Mann–Whitney test was used for the comparisons of categorical or continuous variables, respectively. Univariate logistic regression analyses were performed to screen potential risk factors at the level of p < 0.10, and multivariable logistic regression analyses using backward stepwise approach was used to select independent risk factors associated with CSVD.

We assessed the model performance using discrimination and calibration. The discrimination refers to as the model’s ability to distinguish between high- and low-risk participants and was assessed by calculating the area under the ROC (AUC) that ranged from 0.5 (no better than chance) to 1.0 (perfect discrimination) (23). The calibration was defined as the agreement between the predicted and observed probability. It was determined by Hosmer-Lemeshow test and calibration curve (23), where people were evenly divided into ten groups based on decile of predicted risk, and the predicted probability was plotted against the observed probability. A diagonal line with intercept of 0 and slope of 1 represented ideal calibration.

We assessed the clinical usefulness of the model using decision curve analysis (DCA), which compared the net benefit of using the diagnostic model vs. the strategy of assuming that all people were at high or low risk of CSVD (24). Graphically, the line parallel to the x-axis was drawn to show no net benefit when assuming all people with low CSVD risk and no intervention given, whereas the solid black curve represented all people with high CSVD risk and received intervention. The DCA curve (black dotted curve) was drawn for the established diagnostic model, and the curve with the highest net benefit corresponded to a higher clinical value.

All analyses were performed using IBM SPSS Statistics for Windows, Version 26.0 (IBM Corp., Armonk, NY, USA) and Stata Statistical Software: Release 15.0 (Stata Corp LLC., College Station, TX, USA) for Windows. Two-sided p < 0.05 was considered statistically significant.

In 1304 participants accomplished the structural brain MRI scans, we excluded 112 participants due to suboptimal image quality (n = 70) or missing data on clinical features and laboratory measures (n = 42), leaving 1,192 participants for the current analysis. These participants were randomly divided into the training set (70%, n = 847) and the validation set (30%, n = 345). Figure 1 shows the flowchart of the study participants.

The median age of the 1,192 participants was 69 years (IQR: 66–72 years), 58.56% were female, and 80.37% had limited education (i.e., primary school or no school education) (Table 1). Out of these, 934 (78.36%) were presented with CSVD (i.e., total CSVD score ≥ 1), including 659 (77.80%) in the training set and 275 (79.71%) in the validation set. There was no statistical significance between the two datasets with respect to all the examined variables (Table 1).

Univariate logistic regression analysis suggested nine of the 15 candidate risk factors were associated with CSVD at the p < 0.10 (Table 2). These nine risk factors were then entered into the multivariable logistic regression model, and five risk factors, i.e., older age, high blood pressure, history of cerebral infarction, increased WBC count, and increased NLR, were independently associated with CSVD and selected for the final diagnostic model (Table 2).

A nomogram was built based on the multivariable logistic regression (Figure 2). Each variable was assigned a weighted score based on odds ratio. A total score was generated by adding each weighted score of the risk factors, then the probability of having CSVD was determined by projecting the total score to the total point scale. For example, a person aged 62 years with a history of cerebral infarction, high blood pressure, WBC count of 7.0 × 10⁹/L, and NLR of 2.2 had a total point 19.4, representing approximately 92% probability of having CSVD (Supplementary Figure S1).

The diagnostic model yielded AUC of 0.71 (95% CI, 0.67–0.76) in the training set (Figure 3A) and 0.69 (95% CI, 0.63–0.76) in the validation set (Figure 3B). Ten-fold cross-validation was performed in the training set, with the AUC being 0.69 (Figure 3A). The calibration plot showed a good agreement between the predicted and the observed probabilities of having CSVD in both the training (Figure 3C) and the validation sets (Figure 3D), which was verified by Hosmer-Lemeshow test, with χ² being 10.50 (p = 0.40) and 10.36 (p = 0.41) for the training and the validation set, respectively, indicating no significant difference between the predicted and observed probabilities.

We assessed the clinical utility of the nomogram using DCA. In the training set (Figure 4A), the curve of applying the model crossed the curves of treating all participants as having low and high risk approximately at threshold probability of 90 and 50%, respectively, suggesting that the diagnostic model had higher net benefits for risk thresholds between 50 and 90%. In the validation set (Figure 4B), the model had higher net benefits than the strategy assuming all participants either at high risk or low risk of CSVD for risk thresholds between 65 and 98%.