As of March 1, 2024, all data used in our study were publicly accessible and sourced from the NHANES, managed by the National Center for Health Statistics (NCHS). Since 1999, data collection from participants has been conducted through questionnaire interviews, physical examinations, and laboratory tests. NHANES received approval from the NCHS Institutional Review Board, and all participants have provided written informed consent. Survival status data of NHANES participants were derived from the National Death Index (NDI).

Our study selected 69 congeners of persistent organic pollutants detected and validated in the NHANES database (1999–2004), including dioxins, furans, and PCBs. Through data screening and analysis, we identified seven PCBs—PCB074, PCB170, PCB178, PCB180, PCB156, PCB157, and PCB146—as noteworthy compounds. We have provided specific nomenclature and abbreviations for these compounds:

LBX074, representing PCB074(2,4,4′,5-Tetrachlorobiphenyl).

LBX170, representing PCB170(2,2′,3,3′,4,4′,5-Heptachlorobiphenyl).

LBX178, representing PCB178(2,2′,3,3′,5,5′,6-Heptachlorobiphenyl).

LBX180, representing PCB180(2,2′,3,4,4′,5,5′-Heptachlorobiphenyl).

LBX156, representing PCB156(2,3,3′,4,4′,5-Hexachlorobiphenyl).

LBX157, representing PCB157(2,3,3′,4,4′,5′-Hexachlorobiphenyl).

LBX146, representing PCB146(2,3,3′,4,4′,5′-Hexachlorobiphenyl).

It is important to note that these seven PCBs were all non-dioxin-like PCBs. The completeness of the data and the limitations of the detection equipment and methods at the time may have contributed to the underrepresentation of dioxin-like PCBs in the data analysis.

We matched the NHANES data from 1999 to 2004 with mortality data from the NDI website, excluding participants with missing mortality data and incomplete data for the seven PCBs. Additionally, the NHANES database provides a weight for each participant to account for the complex survey design, including oversampling, non-response adjustments, and post-stratification to align with the total census population. When weighted, the sample in NHANES represents the civilian, non-institutionalized population of the United States. Each participant is assigned a sample weight, which reflects the number of individuals represented by that sample. Therefore, we also excluded participants with missing weights. However, for comprehensive network analysis, we refrained from further data exclusion. This meticulous curation resulted in a final cohort of 10,961 participants. At different analysis stages, data were selected based on varying data requirements.

The analyses were quantified in serum using high-resolution gas chromatography and isotope-dilution high-resolution mass spectrometry (HRGS/ID-HRMS). Serum samples were fortified with 13C12-labeled internal standards and extracted through either C18 solid-phase extraction (SPE) or liquid-liquid extraction. Chromatographic separation occurred on a DB-5ms capillary column employing a Hewlett-Packard 6890 gas chromatograph. Quantification was achieved by ID-HRMS using selected ion monitoring (SIM) at a resolving power of 10,000 with either a Micromass AutoSpec ULTIMA or Finnigan MAT95 mass spectrometer in electron ionization (EI) mode. Detection limits were reported for each sample, accounting for sample weight and analyte recovery.

From the entire persistent organic pollutant (POP) library, we ultimately identified 18 PCBs, with data for these substances in the NHANES cycles from 1999–2000, 2001–2002, and 2003–2004 accounting for over 75% of the total data. Further analysis of the 18 POPs using weighted quantile sum (WQS) analysis identified seven PCBs with the greatest impact on mortality risk, cumulatively contributing to 95% of the total weight. As per NHANES guidelines, values falling below the limit of detection (LOD) were imputed with a value equivalent to the LOD divided by the square root of 2.

As NHANES does not directly record mortality data, mortality information in this study was obtained through a probabilistic match between NHANES and NDI, following procedures validated by the National Center for Health Statistics (NCHS). Mortality data from the NDI were available for analysis until December 31, 2019.

All diseases available in NHANES were included in our analysis, encompassing Angina, Heart attack (HA), Heart disease (HD), Heart failure (HF), Hypertension, Stroke, Alcoholic fatty liver disease (ALD), Hepatitis B virus (HBV), Hepatitis C virus (HCV), Hepatitis D virus (HDV), Non-alcoholic fatty liver disease (NAFLD), Hyperlipidemia, Diabetes, Osteoporosis, Hyperuricemia, Thyroid disease (TD), Human Immunodeficiency Virus (HIV), Arthritis, Depression, Cancer, Chronic bronchitis (CB), Emphysema, Asthma, Chronic kidney disease (CKD), and Proteinuria, totaling 25 diseases. Diabetes was defined as self-reported diabetes diagnosis, use of oral antidiabetic drugs or insulin, glycated hemoglobin (HbA1c) levels ≥6.5%, plasma glucose levels ≥200 mg/dL 2 h after an oral glucose tolerance test (OGTT), or fasting plasma glucose levels ≥126 mg/dL (23). Hypertension was determined based on self-reported hypertension or NHANES-measured data: an average systolic blood pressure ≥130 mm Hg or diastolic blood pressure ≥80 mm Hg from three measurements (24, 25). To assess CKD, essential indicators including estimated glomerular filtration rate (eGFR) and urine albumin-to-creatinine ratio (UACR) were relied upon. UACR (mg/g) was calculated as the ratio of urine albumin (mg/dL) to urine creatinine (g/dL), with a UACR value exceeding 30 mg/g indicating “proteinuria.” eGFR was computed using the CKD-EPI formula, expressed as:

GFR = 175 × standardized serum creatinine^(−1.154) × age^(−0.203) × 1.212 [if Black] × 0.742 [if female], where serum creatinine is measured in mg/dL.

CKD was defined as eGFR <60 mL/min/1.73 m² the presence of renal damage markers (such as proteinuria), or both, persisting for at least 3 months, regardless of the underlying etiology (26). In line with previous publications, NAFLD was defined by hepatic steatosis index (HSI) and US fatty liver index (USFLI). The formulas are as follows:

HIS = 8 × (alanine aminotransferase/aspartate aminotransferase ratio) + body mass index (+2 for female; +2 for diabetes);

USFLI = (e^−0.8073 × Non–Hispanic Black+0.3458 × Mexican American+0.0093 × Age+0.6151 × loge (Gamma glutamyltransferase) +0.0249 × Waist Circumference+1.1792 × loge (Insulin)+0.8242 × loge (Glucose)−14.7812)/ (1 + e^−0.8073 × Non–Hispanic Black+0.3458 × Mexican American+0.0093 × Age+0.6151 × loge (Gamma glutamyltransferase) +0.0249 × waist circumference+1.1792 × loge (Insulin)+0.8242 × loge (Glucose) – 14.7812) × 100.

USFLI cutoff value ≥ 30 or HSI value > 36 was diagnosed as NAFLD (27). ALD was defined by a combination of an evidence of excessive alcohol consumption (≥ 210 g/week for men and ≥ 140 g/week for women) and an ALD/NAFLD index > 0, which was calculated as:

−58.5 + 0.637 (Mean Corpuscular Volume) + 3.91 (Aspartate Aminotransferase [AST]/Alanine Aminotransferase [ALT]) – 0.406 (Body Mass Index) + 6.35 for Male Gender.

In the subpopulation with ALD, the AST-to-platelet ratio index (APRI) and FIB-4 score were used to evaluate ALD FIB. The formula is as follows:

APRI = (AST/Upper Limit of Normal/Platelet Count [109/L]) × 100, where the upper limits of normal AST levels were set at 37 IU/L for men and 29 IU/L for women;

FIB-4 = Age × AST/[Platelets in 109/L × (ALT)^1/2].

Cut-off values for advanced fibrosis (≥ F3) were set at 1.5 for APRI and 3.25 for FIB-4 (28, 29). HBV, HCV, HDV, and HIV-positive patients were determined based on antigen measurements and quantification of relevant viral DNA or RNA levels in NHANES laboratories. Hyperlipidemia was defined as fasting triglyceride values ≥ 200 ng/dl. Smokers were defined as individuals who have smoked more than 100 cigarettes in their lifetime and currently smoke on some days or every day, while nonsmokers are those who have smoked <100 cigarettes in their lifetime (30). Hyperuricemia was defined dichotomously with serum uric acid (SUA) levels ≥416 μmol/L (7.0 mg/dL) for males and ≥357 μmol/L (6.0 mg/dL) for females (31). Apart from the diseases mentioned above, data on Heart attack, Heart disease, Heart failure, Stroke, Osteoporosis, Thyroid disease, Emphysema, Arthritis, Depression, Cancer, Chronic bronchitis, and Asthma were obtained from questionnaire data provided by NHANES. These indices were derived from comprehensive full blood cell count tests, the details of which can be found in the “Questionnaire” data within the NHANES dataset. Participants who answered “Yes” were considered to have that condition. Finally, we categorized these diseases into seven major classes based on disease type: Circulatory system diseases, Digestive system diseases, Endocrine/Metabolic diseases, Immune system diseases, Respiratory system diseases, Urinary system diseases, and Others.

The DII assesses the inflammatory effect of diet using 45 dietary parameters, normalizing individual intake of each food parameter to global intake. Standardized intake scores (Z-scores) are converted to proportions and centered. The centered proportions of these specific food intakes are multiplied by their inflammation effect scores and summed to obtain an individual's overall DII score. Participants' DII scores represent the sum of each DII score. Higher DII scores indicate a pro-inflammatory diet, while lower scores indicate an anti-inflammatory diet. In this study, 28 out of 45 food parameters were utilized for DII calculation: carbohydrates, protein, total fat, alcohol, fiber, cholesterol, saturated fatty acids, monounsaturated fatty acids, polyunsaturated fatty acids, n-3 fatty acids, n-6 fatty acids, niacin, vitamin A, thiamine, vitamin B2, vitamin B6, vitamin B12, vitamin C, vitamin D, vitamin E, iron, magnesium, zinc, selenium, folic acid, carotene, caffeine, and energy (32, 33).

We included C-reactive protein (CRP) and the systemic immune-inflammation index (SII) as covariates in our analysis. We selected and utilized data on platelet count (PC), neutrophil count (NC), and lymphocyte count (LC) in the computation, with SII calculated as SII = PC ^* (NC/LC) (34).

Other covariates included Age, Poverty Income Ratio (PIR), Body Mass Index (BMI), Gender, Race, Education, Smoking Exposure, and Alcohol Exposure. BMI was categorized into three groups: normal (BMI <25 kg/m²), overweight (25 ≤ BMI ≤ 30 kg/m²), and obese (BMI > 30 kg/m²), based on participants' BMI values (35). Alcohol consumption was assessed using data from NHANES questionnaires. Participants who had consumed fewer than 12 alcoholic drinks in their lifetime were classified as non-drinkers. Former drinkers were individuals who had consumed ≥12 drinks at any point in their lifetime but had not consumed alcohol in the past year. To minimize recall bias, smoking exposure was evaluated based on serum cotinine levels rather than relying solely on the “smoking history questionnaire”. Current smokers were identified by serum cotinine levels >10 ng/mg, former smokers had serum cotinine levels ≤ 10 ng/mg, and non-smokers exhibited serum cotinine levels <0.011 ng/mg (36).

In accordance with Centers for Disease Control and Prevention (CDC) guidelines, our statistical analyses adhered to stipulated principles. To address the complex multi-stage cluster survey design inherent to NHANES, appropriate sample weights were meticulously applied to each participant. Categorical variables were expressed as proportions, while continuous variables were presented as means (mean ± standard deviation). Descriptive statistics comprehensively summarized participants' demographic characteristics and biomarker concentrations. Specifically, for each selected PCB, analysis was conducted after stratifying into three groups based on quartiles, and for substances exhibiting highly right-skewed distributions, analysis was stratified into two groups based on the median.

A total of 69 POPs were initially identified in NHANES data (1999–2004), encompassing dioxins, furans, and PCBs. From these, 18 persistent organic pollutants with valid data representing 75% of the total data were selected for further analysis. WQS regression, focusing on mortality risk, was performed on these 18 substances, with the top seven substances selected based on their cumulative contribution rate to the preceding 95%, all of which were PCBs.

Correlation heatmaps were employed to illustrate the interrelationships among the 7 PCBs and their associations with various diseases. Principal component analysis (PCA) was utilized to visualize the relationship between PCBs and mortality risk in two dimensions. Additionally, five diseases highly correlated with PCBs (Hyperuricemia, Hypertension, Diabetes, CKD, and Arthritis) were included in the machine learning algorithm model, while Cancer, Osteoporosis, and Hepatitis C (HCV) were excluded due to either their broad spectrum or insufficient data volume. Further, multivariable logistic regression was employed to evaluate the associations between these substances and 25 diseases adjusted for Age, Gender, Race, BMI, Social Inequality Index (SII), smoking exposure, and alcohol exposure. We analyzed each of the seven PCBs separately for each disease outcome, calculating the effect of each PCB on a specific disease outcome individually. To illustrate the interrelationships among different diseases, logistic regression and comorbidity network analysis were also performed on the 25 diseases, with the most relevant diseases determined based on Odds Ratios (ORs) and p-values, and results presented via network analysis diagrams.

Furthermore, based on the above analysis, we used both traditional Cox regression and machine learning models to investigate whether PCBs act as independent risk factors for mortality. In the Cox regression analysis, after adjusting for baseline and disease data, we found a significant positive correlation between each PCB and mortality risk (HR > 1). Simultaneously, we developed a high-precision mortality risk prediction model. After incorporating PCB data into the baseline and disease data, the model showed a significant improvement in predictive accuracy. The study further demonstrated that PCBs are independent risk factors for mortality, even after adjusting for confounding factors such as diseases. We constructed five individual machine learning models: Support Vector Machine (SVM), Naïve Bayes, Decision Tree, Stochastic Gradient Descent (SGD), and Gradient Boosting Decision Tree (GBDT), as well as four ensemble models: Random Forest, Histogram-based GBDT (Hist GBDT), Bagging, and Neural Networks. Additionally, a voting algorithm was developed for outputting results, which were evaluated using ROC curves and confusion matrices. Moreover, we divided the participants into three groups based on their level of pollutant exposure and then further stratified each group into two subgroups according to whether their DII was >0. Survival curves were plotted for each subgroup, and multivariable Cox regression models were constructed for each group to explore the differential Hazard Ratios (HR) of mortality outcomes associated with the DII, after adjusting for baseline data. Our results effectively demonstrated that dietary patterns modulated by the DII can mitigate the adverse health effects of high PCB exposure in populations with elevated levels of PCB exposure.

Table 1, Supplementary Table 1 present the essential characteristics of our study cohort. The mean age of participants was 40.99 years. Regarding gender distribution, females slightly outnumbered males, constituting 50.2% compared to 49.8%. Additionally, we categorized the 10,961 participants into three groups–low, moderate, and high–based on tertiles reflecting PCB concentrations. Notably, individuals in the high LBX074 group displayed distinct characteristics compared to those in the low and moderate groups (Table 1). Specifically, the high LBX074 group exhibited a higher proportion of females, a significantly greater mean age, and a substantially elevated percentage of active smokers. Similar trends were observed for other comprehensive PCB data, as outlined in Supplementary Table 1.

Figures 1A, B depict the results of WQS and weighting analysis for the 18 PCBs and 25 diseases. By calculating the cumulative contribution rate, we identified the top seven substances, ranked in descending order based on their contribution rate, accounting for 95% of the cumulative contribution rate:

LBX074, representing PCB074(denoting 2,4,4′,5-Tetrachlorobiphenyl).

LBX170, representing PCB170(2,2′,3,3′,4,4′,5-Heptachlorobiphenyl).

LBX178, representing PCB178(2,2′,3,3′,5,5′,6-Heptachlorobiphenyl).

LBX180, representing PCB180(2,2′,3,4,4′,5,5′-Heptachlorobiphenyl).

LBX156, representing PCB156(2,3,3′,4,4′,5-Hexachlorobiphenyl).

LBX157, representing PCB157(2,3,3′,4,4′,5′-Hexachlorobiphenyl).

LBX146, representing PCB146(2,3,3′,4,4′,5′-Hexachlorobiphenyl).

In line with this, the main text primarily focused on LBX074 data, aimed at representing this category of substances, while additional data were predominantly included in the Supplementary material.

Figures 1C, D illustrate, in the form of heatmaps, the interrelationships among the seven PCBs and between PCBs and the 25 diseases, respectively. Significant positive correlations were observed among all seven PCBs (p < 0.0001), with LBX146 exhibiting the strongest correlation with the other six substances. The seven PCBs demonstrated significant positive correlations with most diseases, including arthritis, cardiovascular diseases, respiratory disorders, endocrine diseases, chronic hepatitis, and chronic kidney disease (p < 0.0001). However, all seven PCBs were negatively correlated with Asthma, and LBX178 and LBX167 were negatively correlated with Chronic bronchitis and Hyperlipidemia. Notably, Alcoholic fatty liver showed almost no correlation with any of the PCBs. These findings suggest that PCBs might be important environmental factors associated with the development of most diseases, affecting multiple systems in the body.

Figure 2 employed principal component analysis (PCA) to illustrate the relationship between PCBs and the risk of death in a two-dimensional format. Pairwise combinations of the seven PCBs were analyzed, and the PCA two-dimensional plot indicates that for each pair of PCBs, higher PCB levels correspond to increased risk of death. Thus, all seven PCBs are important factors in increasing the mortality risk.

Multivariable logistic regression was employed to analyze the summarized 25 diseases, adjusting for Age, Gender, Race, BMI, SII, smoking exposure, and alcohol exposure. Each of the seven PCBs was individually analyzed in relation to each specific disease outcome, with the impact of each PCB on a given disease outcome calculated separately. Among them, 12 diseases showed significant associations with PCBs (p < 0.05), with six diseases significantly positively correlated with PCBs (OR>1, p < 0.05) (Figure 3A), including Hyperuricemia (OR > 6.0, p < 0.001), Diabetes (OR > 6.0, p < 0.0001), HCV (OR > 6, p = 0.0071), Hyperlipidemia (OR > 6, p = 0.016), HIV (OR > 6, p = 0.0005), and Arthritis (OR = 5.42, p = 0.0065). Adjusting for covariates, it could be demonstrated that PCBs were independent influencing factors for multiple diseases.

Logistic regression was used to compute the odds ratio (OR) and p-values for each pair of the 25 diseases to determine disease associations. Each disease outcome was paired with the remaining 24 disease outcomes, and by continuing to adjust for covariates consistent with the previous models, 600 different paired models were generated. We calculated the OR values for all disease pairs (Supplementary Table 2) and presented the disease pairs most closely related to each disease (p < 0.05) (Supplementary Table 3). Among these pairs, Depression and Stroke emerged as the most correlated diseases (OR [95%CI]: 40.21[5.83, 794.00], p = 0.001), while the association between Thyroid disease and Angina was the least significant (OR [95%CI]: 3.60 [2.76, 4.66], p < 0.0001). Diseases of the circulatory system tend to be interrelated, with diseases highly correlated with diabetes and proteinuria, typical of the urinary system, accounting for the majority of associations. Depression is strongly associated with some common chronic clinical conditions. Based on comorbidity network analysis, we identified 283 potential links, with the number of related links, OR values, and comorbidity network analysis results shown in Figure 3B. Each node represents a medical condition, and the thickness of the connecting lines reflects the strength of the disease pairs' association. Nodes closer to the network center have stronger centrality, indicating a greater number of connections with other diseases. Hypertension exhibits the strongest centrality and the highest number of connections, significantly influencing most other diseases. It is also notable that diseases of the circulatory system are often closely associated with other diseases. Depression, ALD, HIV, and HPL are distanced from the center, showing fewer associations with other diseases and smaller impact.

Before conducting machine learning predictions, we first applied the traditional Cox regression method for analysis (Supplementary Figure 2). The model was adjusted for baseline demographic and disease data, consistent with the data included in the predictive model. This analysis successfully validated the independent effect of PCBs on mortality risk. A significant positive correlation was observed between most PCBs and mortality risk, with increased PCB exposure corresponding to a notable rise in mortality risk. For example, PCB74 (HR [95% CI]: 1.990 [1.288, 3.074], p = 0.002, group of “High”), PCB170 (HR [95% CI]: 2.015 [1.039, 3.909], p = 0.038, group of “High”), and PCB180 (HR [95% CI]: 3.090 [1.667, 5.729], p = 0.0003, group of “High”).

After that, we constructed nine learning models, including five individual models: Support Vector Machine (SVM), Naïve Bayes, Decision Tree (Tree), Stochastic Gradient Descent (SGD), and Gradient Boosting Decision Tree (GBDT). Additionally, we developed four ensemble models: Random Forest, Histogram Gradient Boosting Decision Tree (hist GBDT), Bagging, and Neural Network. Furthermore, we incorporated a Voting algorithm for result output. Two types of models were added to the algorithm: one containing Age, gender, race, hyperuricemia, hypertension, diabetes, CKD, and arthritis, and the other adding PCBs data. The AUC values of these two types of models, representing predictive accuracy, were compared (Figures 4A, B). It is evident that models incorporating PCBs data exhibited varying degrees of improvement in accuracy compared to the baseline models. Overall, ensemble models outperformed individual models, with Random Forests showing significant advantages in prediction. Prior to incorporating PCBs data, the accuracy rates were as follows: SVM (0.86), SGD (0.88), Naïve Bayes (0.85), Decision Tree (0.88), GBDT (0.91), hist GBDT (0.89), Random Forests (0.98), Bagging (0.92), Neural Network (0.91), and the final Voting model (0.94). After incorporating PCBs data, the accuracy rates were: SVM (0.89), SGD (0.90), Naïve Bayes (0.86), Decision Tree (0.90), GBDT (0.94), hist GBDT (0.91), Random Forests (1.0), Bagging (0.95), Neural Network (0.95), and the final Voting model (0.96). Figures 4C, D illustrate the comparison of ROC curves before and after applying the Voting algorithm. Figures 4E, F illustrate the comparison of the confusion matrices for the Voting algorithm before and after. These results strongly demonstrate a substantive positive correlation between PCBs and mortality risk, even after controlling for baseline data and confounding factors such as diseases. The ROC curves and confusion matrices of the other algorithms are presented in Supplementary Figure 1.

Furthermore, in Figure 5, we demonstrated through survival curves that we could mitigate the impact of DII by adjusting lifestyle habits. We first continued to utilize the previously built machine learning models to calculate the accuracy and errors of the predictive model containing population baseline data (Age, gender, race) and 5 diseases (hyperuricemia, hypertension, diabetes, CKD, arthritis), along with 7 PCBs data (Figure 5A). Furthermore, we incorporated DII data of each participant into the aforementioned model for prediction and recalculated the accuracy and errors of the new model (Figure 5B). The results indicated that the DII index effectively enhanced the accuracy of the model, highlighting its significance as a contributing factor to increased mortality risk. In Figure 5C, participants were categorized into low, medium, and high groups based on the quartiles of the total amount of seven PCBs in their bodies, and survival curves were plotted accordingly. Participants with high levels of PCBs showed a significantly increased risk of mortality, while participants with medium and low levels exhibited progressively lower risks of mortality (p < 0.0001). In Figure 5D, participants with high levels of PCBs were further classified based on their DII scores: those with DII scores >0 were defined as “positive,” while those with DII scores ≤ 0 were defined as “negative.” In comparison with Figure 5C, it is evident that a DII score >0 (indicating reduced inflammation) effectively reduces the risk of mortality among participants with higher levels of PCBs. Figures 5E, F depict similar classifications for participants with medium and low levels of PCBs, respectively. However, the conclusions drawn from participants with high levels of PCBs were not statistically significant in these two groups, indicating no significant differences.

In Figures 5G–I, we further constructed multifactor Cox regression models to examine the association between DII index and mortality, adjusting for age, gender, race, education, BMI, smoking exposure, and SII covariates. Among the high PCBs population, a DII index <1 (indicating inflammation suppression) was significantly negatively correlated with mortality (OR [95% CI] = 0.7151 [0.6683, 0.7651], p < 0.05). However, in the medium PCBs and low PCBs populations, the association between a DII index <1 and mortality was not statistically significant (OR [95% CI] = 1.05689 [0.94773, 1.1786], p > 0.05) (OR [95% CI] = 1.125 [0.87141, 1.4525], p > 0.05).

Through the above analysis, we effectively demonstrated that among the high PCBs population, modulation of dietary habits and adjustment of DII can effectively counteract the adverse effects of PCBs and similar environmental factors on human health. However, in the medium PCBs and low PCBs populations, the impact of DII adjustment on mortality is not evident.