The TLS is a community-based cohort initiated in Taizhou, China since 2007 [13]. Among TLS participants considered for this analysis, 16,358 individuals enrolled from September 2011 to January 2014 and 26,308 individuals enrolled from July 2018 to November 2021 were initially considered for the discovery and temporal validation cohorts, respectively. After restricting to participants with available valid serum or plasma samples and applying further exclusions for baseline major NCD history, missing key data, or incident major NCDs within 180 days post-baseline, 14,478 and 25,018 participants were finally included, respectively. The study flowchart is presented in Supplementary Figure S1.

The study was approved by the Ethics Committee of Fudan University Taizhou Institute of Health Sciences (Ethics Approval No. B016) and was conducted in accordance with the World Medical Association Declaration of Helsinki. Written informed consent was obtained from all participants prior to enrollment in the TLS.

Baseline data were collected via face-to-face interviews, physical examinations, and blood sampling. Peripheral blood samples (8–10 mL) were collected into K2 EDTA vacuum tubes, stored at 4 °C until same-day processing. After centrifugation to separate plasma, samples were aliquoted into barcoded cryovials and stored at −80 °C or below. Biochemical tests for both cohorts were centrally conducted at Shanghai Fourth People’s Hospital from June to September 2024 [14, 15], using consistent laboratory platforms and quality-control procedures. The study ultimately collected 54 circulating blood biomarkers and 26 questionnaire- and physical examination-derived indicators [16]. The latter covered demographic and socioeconomic factors, lifestyle factors, anthropometric and body composition factors, blood pressure-related factors, and dietary factors. Details of blood biomarkers are provided in Supplementary Table S1, candidate variable categories and their relevance to major NCD risk are summarized in Supplementary Table S2, and definitions, missing rates, and screening status of questionnaire- and physical examination-derived indicators are provided in Supplementary Table S3.

Variables with >20% missingness in either cohort were excluded; missingness of blood biomarkers and questionnaire/physical indicators is summarized in Supplementary Tables S1, S3. For missing values, categorical variables were imputed with the most frequent category, continuous variables with the median. Except for Helicobacter pylori antibody, circulating blood biomarkers were standardized within each cohort and entered association models as continuous variables, with effect estimates expressed per 1-standard-deviation increase. For highly correlated biomarkers (Pearson’s correlation >0.8), the one with lower missing rate or higher variance was retained. After filtering by missingness, overlap, correlation, and Cox pre-screening, 25 blood biomarkers and 11 traditional factors were retained for multi-method selection.

The primary outcome was incident major NCDs (cancer, CVD, and CRD). All-cause mortality was treated as a secondary outcome to examine whether the derived risk score and risk groups also differentiated all-cause mortality risk. Disease outcomes were determined using International Classification of Diseases, 10th Revision (ICD-10) codes from the Taizhou Center for Disease Control and Prevention and local medical security systems; all-cause mortality was defined as death from any cause during follow-up and was ascertained through the same record linkages. Diagnoses were validated via pathological reports, imaging, or clinical assessments. Cancer outcomes included five common types: lung cancer (C33, C34), esophageal cancer (C15), gastric cancer (C16), liver cancer (C22), colorectal cancer (C18-C20). CVD included coronary heart disease (I20-I25) and stroke (I60-I64, I69); CRD was represented by chronic obstructive pulmonary disease (J41-J44).

For individuals with multiple major NCD diagnoses, only the earliest incident event is considered for the composite outcome; for disease-specific analyses, only the first recorded diagnosis within each disease category is retained. To reduce the influence of undiagnosed disease at baseline and ensure valid incident cases, only events occurring ≥6 months after baseline were included. Follow-up started 6 months after baseline and ended at the first target event, withdrawal, death, or the administrative censoring date of December 31, 2023, whichever came first.

In the discovery cohort, multivariable Cox proportional hazards regression was first used to screen candidate predictors. Least absolute shrinkage and selection operator (LASSO), stepwise regression, and Boruta random forest were then applied in parallel, and variables consistently selected across these three methods were retained for construction of the final NIMO score (Supplementary Figure S2). LASSO used cross-validation to select the penalty parameter; stepwise regression used bidirectional selection based on the Akaike information criterion; Boruta random forest was run with 500 trees and a maximum of 50 runs.

Variables were categorized into subgroups. Clinical reference limits or commonly used clinical thresholds were prioritized where appropriate; for variables lacking uniform clinical thresholds or with limited clinical discrimination, cut-offs were determined using the maxstat.test function in the discovery cohort. Cox model coefficients were then used to derive relative risk points for each subgroup, with the reference subgroup assigned 0 points. For interpretability, coefficients were multiplied by 10 and converted into point values. Total risk scores were summed across variables for each participant. Participants were classified into low-risk (<25th percentile), intermediate-risk (25th–75th percentile), and high-risk (>75th percentile) groups.

All cut-offs were fixed when applied to the temporal validation cohort. The scoring rule derived from the discovery cohort was fixed and directly applied to participants recruited during a later period, allowing assessment of temporal reproducibility. Stratification performance was evaluated in both cohorts using cumulative incidence curves and hazard ratios for incident major NCDs; all-cause mortality was examined as a secondary outcome to assess whether the risk groups also differentiated mortality risk.

Robustness analyses included multiple imputation and complete-case analysis [17], alternative risk-group thresholds, alternative predictor-selection strategies, continuous z-standardized biomarker modeling, sex- and age-stratified analyses, non-CVD and subtype-specific outcomes to examine whether the composite-outcome findings primarily reflected CVD events, an expanded outcome definition including diabetes (E10–E14), breast cancer (C50), and asthma (J45–J46), glucose-excluded scoring, and a fixed 3-year follow-up window.

Internal robustness was assessed using 50 repeated 7:3 splits within the discovery cohort, stratified by incident major NCD status. In each split, the final predictors were retained, Cox coefficients were re-estimated in the training subset, and performance was evaluated in both subsets. Variable-selection stability was assessed by repeating the full predictor-selection procedure in 50 bootstrap resamples [18, 19].

Cancer-, CVD-, and CRD-specific scores were developed in the discovery cohort using the same candidate predictor pool, selection procedure, categorization strategy, and coefficient-based scoring framework as the NIMO score, with each first incident disease subtype as the target outcome. Within the discovery cohort, we compared high-risk overlap, stratification performance, and discrimination between the NIMO score and the disease-specific scores across composite and subtype outcomes.

To evaluate information-source contributions, two additional scores were developed using the same framework: a traditional-factor score based on questionnaire- and physical examination-derived variables, and a blood-based score based on circulating blood biomarkers plus age and sex. Both were developed in the discovery cohort, applied to the temporal validation cohort, and compared with the NIMO score using risk gradients, hazard ratios, and areas under the receiver operating characteristic curves (AUCs).

Continuous variables are presented as mean ± standard deviation for approximately normally distributed data or median (interquartile range) otherwise. Continuous variables were compared using Welch’s t test or the Wilcoxon rank-sum test, as appropriate. Categorical variables were compared using the χ² test or Fisher’s exact test. Risk stratification was evaluated using hazard ratios (HRs) and 95% confidence intervals (CIs). For biomarker-outcome association analyses, multivariable Cox proportional hazards models were adjusted for age, sex, body mass index, smoking status, alcohol drinking, hypertension, and daily vegetable intake. Discrimination was assessed using AUC, with the concordance index (C-index) reported where applicable. Differences in AUCs between scores were compared using the DeLong test.

A two-sided P value <0.05 was considered statistically significant. All analyses were performed using R software (version 4.3.2).

The discovery cohort included 14,478 participants with 2,597 incident major NCD events and 632 deaths. The temporal validation cohort included 25,018 participants with 726 incident major NCD events and 213 deaths. The discovery cohort had a median follow-up duration of about 10.84 years, including 14,478 participants. In subtype-specific analyses, 543 cancer events, 1,601 CVD events, and 719 CRD events were identified. The temporal validation cohort had a median follow-up of 3.98 years; in subtype-specific analyses, 179 cancer events, 460 CVD events, and 120 CRD events were identified. Baseline characteristics of both cohorts are shown in Table 1 at the end of the text. Participants in the temporal validation cohort were older at baseline, whereas sex distribution was similar between cohorts.

The missingness of blood biomarkers in both cohorts is presented in Supplementary Table S1. As an overall health endpoint, we first examined associations between baseline blood biomarkers and all-cause mortality in the discovery cohort (Supplementary Figure S3). Tumor markers, inflammation-related markers, and markers of liver/kidney function and nutritional status were all significantly associated with mortality risk, suggesting that routine examination biomarkers capture long-term health risk across multiple physiological domains.

We then assessed associations between blood biomarkers and incident major NCDs and their subtypes (Figure 1). Figures 1A–C summarize significant biomarker associations across major NCD subtypes using a Circos plot and category-level counts and proportions. For cancer, tumor markers and liver function markers contributed the largest absolute counts, while several electrolytes, including chloride, sodium, and phosphate, were prominent in relative abundance. For CVD and CRD, kidney function markers such as cystatin C and creatinine showed strong overlap, whereas glucose showed a more CVD-specific pattern. Figure 1D shows adjusted Cox proportional hazards results for incident major NCDs and subtype outcomes. Overall, 25 circulating blood biomarkers were significantly associated with incident major NCDs, including 23 positively associated and 2 inversely associated biomarkers.

After applying the prespecified Cox pre-screening and three-method consensus selection framework in the discovery cohort, 19 predictors were retained for construction of the NIMO score. The final NIMO score included 15 circulating blood biomarkers, plus age, smoking status, hypertension, and daily vegetable intake. Their relevance to major NCD risk is summarized in Supplementary Table S4. Supplementary Figure S4 shows selected predictors, cut-offs, and coefficient-based point assignments used to construct the NIMO score in the discovery cohort. Participants were then classified into low-, intermediate-, and high-risk groups according to the final score.

In the discovery cohort, the NIMO score showed clear risk gradients for incident major NCDs (Supplementary Figure S5). The high-risk group represented 25.0% of participants but accounted for 44.7% (1162/2597) of incident major NCD events, with a 4.75-fold higher risk than the low-risk group (95% CI: 4.17–5.41). The NIMO score showed moderate discrimination for incident major NCDs, with an AUC of 0.70.

Similar gradients were observed for all-cause mortality. The high-risk group accounted for 55.9% (353/632) of deaths and had a 7.56-fold higher mortality risk than the low-risk group (95% CI: 5.47–10.46). The AUC for all-cause mortality was 0.73.

In the temporal validation cohort, the NIMO score retained clear risk stratification for incident major NCDs (Figure 2). The high-risk group represented 25.0% of participants but accounted for 53.9% (391/726) of incident major NCD events, with a 6.29-fold higher risk than the low-risk group (95% CI: 4.83–8.19). Similar gradients were observed for all-cause mortality: the high-risk group accounted for 59.6% (127/213) of deaths and had a 12.74-fold higher mortality risk than the low-risk group. The NIMO score showed moderate discrimination in the temporal validation cohort, with both AUCs and C-indexes of 0.70 for incident major NCDs and 0.75 for all-cause mortality.

Sensitivity analyses supported the robustness of the NIMO score, with generally consistent risk gradients across subgroup, missing-data, risk-threshold, predictor-selection, continuous-biomarker, expanded-outcome, glucose-excluded, and fixed 3-year analyses (Supplementary Figure S6; Supplementary Tables S5, S6). The composite-outcome findings were not solely driven by CVD: cancer and CRD together accounted for 42.0% and 39.8% of first events within the original composite outcome in the discovery and temporal validation cohorts, respectively, and the score retained stratification for non-CVD outcomes and for cancer, CVD, and CRD separately (Supplementary Table S7). Repeated 7:3 internal validation showed stable discrimination and risk gradients, and bootstrap analyses showed all-three consensus frequencies of 80%–100% for the selected predictors (Supplementary Table S8).

As disease-focused comparators, the cancer-, CVD-, and CRD-specific scores were compared with the NIMO score. The selected predictors for the main and disease-specific scores are summarized in Supplementary Table S9. The four scores shared overlapping predictors and high-risk population characteristics. Shared predictors included age, smoking status, daily vegetable intake, and several circulating blood biomarkers, such as carcinoembryonic antigen, cytokeratin 19 fragment, C-reactive protein, and cystatin C. Additionally, Figure 3A showed that among individuals identified as high risk by disease-specific scores, 56.1% were classified as high risk by at least two disease-specific scores, and 28.1% were classified as high risk by all three. Figure 3B further showed that the NIMO high-risk group captured 71.7%, 83.9%, and 65.2% of cancer-, CVD-, and CRD-specific high-risk individuals, respectively. Within the NIMO high-risk group, 79.2% also met high-risk criteria for at least two disease-specific scores.

In risk-stratification analyses, the NIMO score showed clear high-versus-low risk gradients across the composite and subtype outcomes, with HRs of 4.75 for major NCDs, 7.02 for cancer, 4.98 for CVD, and 3.22 for CRD (Figure 3C). Disease-specific scores generally showed stronger gradients for their target outcomes, but several non-target outcomes also showed elevated risks, suggesting cross-disease risk clustering. In discrimination analyses, disease-specific scores tended to achieve higher AUCs for their target outcomes, whereas the NIMO score showed relatively stable discrimination across the composite and subtype outcomes (Figure 3D). Together, these findings suggest that disease-specific scores may be more suitable for focused single-disease risk assessment, whereas the NIMO score provides a first-step integrated stratification approach for identifying individuals with elevated overall major NCD risk before further disease-specific evaluation.

As information-source comparators, the traditional-factor and blood-based scores were compared with the NIMO score. The selected predictors are summarized in Supplementary Table S9. In both cohorts, all three scores showed graded risk stratification for incident major NCDs and all-cause mortality (Supplementary Table S10). The blood-based and traditional-factor scores showed broadly comparable performance, suggesting that both information sources contributed to risk stratification. Discrimination analyses showed that the NIMO score had slightly higher AUCs than either comparator score, whereas the blood-based and traditional-factor scores did not differ significantly (Supplementary Figure S7). These findings suggest that integrating circulating blood biomarkers with traditional factors provided more stable risk stratification and modestly improved discrimination.

Several limitations must be acknowledged. First, blood samples were stored for varying durations before centralized testing, and some degradation might have occurred despite standard protocols. However, reliability tests of key biomarkers showed high correlations (63%–94%). Second, all-cause mortality was used only as a secondary outcome, and the temporal validation cohort had relatively short follow-up with limited death events. Third, diabetes was not included in the primary composite outcome because of its close relationship with baseline glucose. Although expanded-outcome and glucose-excluded analyses partly addressed this issue, they cannot replace dedicated modeling of diabetes and its complications. Fourth, the cut-off-based scoring approach improved interpretability but may have caused information loss and may depend on cohort-specific distributions and outcome patterns [29–32]; therefore, the stability of cut-offs and point assignments requires further external validation. Single imputation may also underestimate uncertainty, although continuous biomarker modeling, multiple-imputation, and complete-case analyses showed consistent results. Finally, some biomarkers, particularly tumor-related markers, may not be routinely measured in all health examination settings, and implementation should consider test availability, cost, and laboratory platform consistency. Participants were recruited from a single geographical area, and further external validation in different regions and populations is needed.

This study developed and temporally validated the NIMO model using a large population-based cohort, proposing a blood biomarker–based framework for major NCD risk identification. The model integrates blood biomarkers with basic epidemiological information, reflecting multisystem health status and identifying shared risk directions for major NCDs from overall blood profiles. In the temporal validation cohort, the high-risk group accounted for more than half of incident major NCD events and had an approximately sixfold higher risk than the low-risk group, while also showing substantially elevated all-cause mortality risk. Furthermore, comparisons with disease-specific scores further supported NIMO’s potential value for both composite and subtype risk stratification.