Introduction

Surface water is an essential natural resource1. Deterioration of water quality creates enormous challenges to water availability, sustainable development and human well-being2. In assessing the health risk of poor surface water quality, carcinogenic effects are a key aspect. Relevant human evidence is considered essential for establishing surface water quality standards, that is, the minimum legislative requirements for safeguarding public health3,4. However, the association between surface water quality and cancer remains inadequately understood, especially in China, a country that accounts for nearly a quarter of global cancer cases5,6. This poses major limitations to the development of a suitable Chinese surface water quality standard, and is also deemed a driving factor in the discrepancy between standards adopted by different countries and authorities7.

Drinking water sources and disinfection by-products have been linked to the risk of several cancers (especially oesophageal8, gastric9, colorectal and renal10) in cohort or case-control studies using individual questionnaire data. However, these designs have rarely been applied to analysis of surface water pollutants because it is challenging, if not impossible, to determine individual exposure level given complex exposure routes, including ingestion, dermal absorption and inhalation3,11. Another drawback with individual-level analyses is the slowness of evidence generation. Even beyond the scope of water research, only 16 agents have been added to the Group I carcinogen list during the past 10 years (and only two since 2019)12,13. Water monitoring and population-level cancer data have enabled ecological studies to make more efficient and comprehensive inspections of the association14. For instance, Yang et al. confirmed a connection between frequency of water pollution and population mortality from digestive cancers in the Huaihe River Basin in 2014. This was influential in promoting ongoing countermeasures in this specific area15,16. At the national level, China has now established one of the world’s largest networks on water quality monitoring17 and cancer registries18, both covering all 31 province-level areas of the mainland. These efforts provide a unique opportunity to generate new information to policy-makers, scientists and the public. However, water bodies vary in their geographic background, pollutant level and combination19. The coexistence of multiple pollutants in surface water has not been examined, and the complex spatial patterns pose a methodological challenge for comprehensive appraisal of the attributable cancer burden.

In this study, we put forward a framework that integrates nationwide data from two large industries on surface water and cancer, and a design that decodes spatial associations between different pollutants and cancers. The study aimed to generate evidence on the systematic impact of surface water quality on cancer, determine the most influential water pollutants in different river basins, and forecast (and therefore possibly curb) a water quality-related cancer burden in the foreseeable future.

Results and discussion

Water quality assessment within an integrated framework

The quality of surface water in China is regularly assessed through the National Surface Water Environmental Quality Monitoring Network. During 2001–2021, this network experienced a large expansion in the number of water monitoring sections (from 454 to 3632; Supplementary Fig. 1). It now provides full coverage of all major rivers and administrative regions at prefecture level17. Figure 1a shows the locations of water monitoring sections. These were matched to areas where cancer incidence data was available through the National Cancer Registry System (Fig. 1b). The matching design relied on the natural distribution of surface water and used each water quality monitoring section as the “centre” of a buffer zone with a 30 km radius. It retained 2331 out of 3632 water monitoring sections and 486 out of the 487 qualified cancer registries, covering a population of 380 million in all 31 provinces in mainland China. This was considered a representative sample of the country for both surface water (because of a non-artificial process; Supplementary Fig. 2) and its environmental influence on population cancer.

Fig. 1: Spatial association between surface water quality and cancer incidence.
figure 1

a Spatial distribution and pollution grading (as assessed by proposed “Cluster Analysis” of Multi-pollutants in Space design) of water monitoring sections. H denotes the number of pollution indicators in the high–high clusters or high–low outliers in the area: Grade 0 pollution (H = 0), Grade 1 pollution (H = 1–2), Grade 2 pollution (H = 3–5 H) and Grade 3 pollution (H ≥ 6). b Distribution of total incidence of 11 selected cancers in 486 cancer registry areas of China. The coloured boxes show cancer incidence in cases per 100,000 population. c Incidence and rate ratios (RRs) with 95% confidence intervals for each cancer in the area exposed to the different grades of surface water quality.

Full size image

We considered all nine river basins in mainland China (Songhua & Liaohe, Haihe, Huaihe, Yellow, Continental, Southwest, Yangtze, Southeast and Pearl River Basins), and used this division to stratify the complex water network after a few modifications (see Fig. 1a). We sub-divided the Yangtze River Basin (the world’s third longest river, the mainstream covering 11 provinces in China) into up-, mid- and down-stream basins to reflect discrepancies in both environment and population health20. We combined the Continental and Southwest River Basins for a similar reason (both located in western China and mainly include inland rivers). Annual summary data on the surface water quality in the monitored sections of these river basins in 2021 are given in Supplementary Table 1. The analysis included 21 basic pollution indicators because of their testability, comprehensiveness and high pollution share rate in China.

To make a comprehensive assessment, we applied the current Chinese standard, the Environmental Quality Standard for Surface Water (EQSSW version 2002)21. This sets a maximum of six levels for each pollution indicator, and the overall water quality is determined by the highest (i.e., worst) level. The water quality in 86.14% of the monitoring sections met the entry-level standard for sources of drinking water (Level III21). This was a marked improvement in surface water quality over time, compared with previous reports that used the same criterion in China17 (Supplementary Fig. 1). However, this quality standard, which has been in use for 20 years, was not designed for cancer risk assessment. The high number of monitoring sections meeting the standard across China raises questions about whether this criterion is useful in catalyzing further environment improvement or to benchmark between monitored sections.

To overcome these issues, we used the 75th percentile of each pollution indicator value as a threshold for defining inferior water quality level. This was stricter than EQSSW Level III for all the indicators except total nitrogen (Supplementary Table 1). Against these new standards (Fig. 1a), Western and Central China were largely free of inferior-quality surface water. However, the Eastern area had several areas with multiple indicators of inferior quality, particularly in the Huaihe, Haihe and Songhua & Liaohe Basins. Per-capita water availability in China is only a quarter of the world average22, and it is particularly low in these areas (e.g., 37.5, 62.0 and 74.6 cubic metres per head in Beijing, Tianjin and Hebei in the Haihe River Basin) compared to the national average (2156.3 cubic metres per head in 2020)23. The varying distribution pattern reflects the validity of our proposed 75th percentile criteria and, more importantly, the difficulty of protecting and improving access to clean water.

Association between multiple pollutants and cancers

We ran a negative binomial regression analysis24 for each pollution indicator, including any river basins in which over 20% of the monitored sections of water were of inferior-level quality. This value was used to avoid a “dilution” effect in analysing association and was selected because it approximates to the national average level using the 75th percentile threshold. Of the 21 pollution indicators, 11 (total nitrogen, petroleum, total phosphorus, permanganate index, chemical oxygen demand (COD), volatile phenol, fluoride (F−), ammonia nitrogen (NH3-N), arsenic, selenium, and zinc) were positively related to at least one cancer from the following sites (Supplementary Fig. 3): oesophagus (International Statistical Classification of Diseases 10th revision code: C15), stomach (C16), colorectum (C18–C21), liver (C22), gallbladder (C23–C24), pancreas (C25), lung (C33–C34), bone (C40–C41), breast (C50), kidney (C64–C66, C68) and brain (C70–C72, C32–C33, D42–D43). These 11 cancers were included in our analysis because of their significance in the Chinese population, assessed using incidence, mortality and 5-year survival rate25,26.

Spatial autocorrelation was found for each of the cancer-related pollution indicators, evaluated using the global Moran’s method (Supplementary Table 2; all p < 0.05). We identified high–high (HH) and low–low (LL) clusters (i.e., neighbouring water sections with similar superior/inferior water quality level), and high–low (HL, high amongst low) and low–high (LH, low amongst high) outliers of the monitored water sections, using local Moran’s I, to show the distribution pattern of each indicator (Supplementary Fig. 4). Overall, more HH clusters and HL outliers were observed in North China than the South, but this varied for specific pollution indicators. The pollution indicators therefore form a complex pairwise correlation structure (Supplementary Fig. 5).

We were interested in whether multiple pollutants coexisted in the same area and their joint effect on cancers. We therefore proposed a design, “Cluster Analysis of Multi-pollutants in Space (CAMS)”. We graded each monitored water section by the number of pollution indicators in HH clusters or HL outliers. For example, if there were two pollution indicators in the HH clusters and one in the HL outliers, then the H value for this section would be 3. This gave a CAMS grading: Grade 0 (H = 0; 26.9% of the water sections), Grade 1 (H = 1–2; 38.4%), Grade 2 (H = 3–5; 24.0%), Grade 3 (H ≥ 6; 10.7%). Compared with existing water quality classification rules (e.g., the EQSSW worst-level approach), the CAMS grading has advantages because it takes into account the coexistence of multiple pollutants in a specific location, and the impact of each area’s relationships with neighbouring areas. This reflects the impact of both point and non-point source water pollution27.

We found an approximate “dose–response” relationship between the CAMS grading and population incidence data for the selected cancers (Fig. 1c). Grades 1, 2 and 3 were associated with a relative increase in population incidence of 10%, 17% and 17% in stomach cancer, 9%, 36% and 29% in pancreatic cancer, and 9%, 35%, and 33% in kidney cancer compared with Grade 0 (all p < 0.05); no statistical significance when comparing the two highest levels. This suggests that these cancers are very sensitive to water quality, especially when there is more than one pollutant present. There was no significant increase in population incidence in oesophageal, breast, gallbladder or lung cancer when comparing Grade 1 to Grade 0. However, a joint effect became apparent at higher grades. For instance, there was a relative increase in incidence in oesophageal cancer of 18% (Grade 2) and 47% (Grade 3). A significant increase (18% and 11%) was seen in bone and brain cancer incidence in Grade 3 areas compared to Grade 0. This may only have been visible in Grade 3 areas because of the low incidence of these cancers. When combined, the overall incidence of the 11 selected cancers increased from 197.42 per 100,000 in Grade 0 to 222.55 per 100,000 in Grade 3 (a relative increase of 12.7%).

These results provide evidence of a nationwide spatial association between water quality and cancer in China. This study simultaneously demonstrates such an association for multiple cancers. The multiple-test results of the pairwise association found between each pollution indicator and each cancer site should be treated with caution. However, the consistency in the “dose–response” relationship for the majority of the cancer sites reinforces our confidence that the contribution of surface water quality to cancer is widespread.

It may be helpful to consider some insights into the CAMS grading to help explain its observed relationship with cancer incidence. For nearly all the specific pollutants, the quality was poorer in water sections of higher CAMS grades (Supplementary Fig. 6a–d): take pollutant F− as an example, the rates of inferior quality (i.e., exceeding the 75th percentile standard) were 1.0% (Grade 0), 12.5% (Grade 1), 52.8% (Grade 2) and 90.8% (Grade 3). Besides that, there were widespread correlations between the pollutants, and these became more complex (positively or negatively correlated) and more apparent (higher absolute values of correlation coefficient) in Grades 2 or 3 (Supplementary Fig. 6e–h). These findings suggest that the CAMS grading could be used as a comprehensive measure to summarize both single roles of pollutants and the complex effects of their interactions. Another spatial analysis in the United States showed water-borne chemicals not currently recognized as carcinogens may contribute to the population cancer risk28. When interpreted with our results, this suggests that specific pollutants may not necessarily be carcinogenic on their own. Rather than assessing the health risk of individual pollutants in surface water, we may need to look more comprehensively.

Key cancer-related pollutants in specific river basins

There was a dramatic variation in cancer incidence between cancers and across river basins (Fig. 2a). To assess the environmental impact (water quality and other socioeconomic factors), we labelled the top three river basins with the highest average incidence per cancer, and applied a machine learning technique (SHapley Additive exPlanations (SHAP) values generated from an extreme gradient boosting algorithm) to provide deeper insights into river basin-specific contributions of the pollution indicators (Fig. 2b).

Fig. 2: Key cancer-related pollutants in river basin-specific contexts.
figure 2

a Ranked incidence of 11 selected cancers in 10 river basins with different grades of surface water quality. CAMS, proposed “Cluster Analysis” of Multi-pollutants in Space design. b Key pollution indicators affecting total cancer incidence in each river basin, assessed by SHapley Additive exPlanations (SHAP) values generated from an extreme gradient boosting algorithm.

Full size image

River basins that generally had poorer-quality water (Huaihe, Haihe, Songhua & Liaohe, and downstream Yangtze River Basins) ranked high in incidence of multiple cancers. The Huaihe River Basin in particular is renowned for its high levels of pollution19,29,30. It ranked highest for oesophageal cancer and third for both gallbladder and bone cancers, even though the Chinese government has invested heavily in environmental restoration and cancer prevention in this area in recent years31,32. This finding emphasizes the difficulty in completely eliminating the impact of water pollution, especially in a short time.

Permanganate index, petroleum and chemical oxygen demand were important pollution indicators that could explain the high level of cancer incidence in the Huaihe River Basin. We obtained similar findings from the Haihe and Songhua & Liaohe River Basins. The Haihe River Basin is next to the Huaihe River Basin both geographically and in its water quality, and ranked second for stomach cancer and third for oesophageal cancer. In this basin in particular, F− was related to cancer incidence. A previous study reported that the high level of F− in this basin was related to the long-term over-exploitation of groundwater for agricultural irrigation in the area, resulting in a large amount of F− in clay soils entering the groundwater and surface water33. Recent systematic reviews have not found a link between water F− and cancer34,35, but further work is needed to explain our observations in the Haihe River Basin. Another interesting finding was that selenium was related to cancer incidence in the Songhua & Liaohe River Basin. Previous observational studies have suggested that selenium has a protective effect against cancer36, but recent randomized controlled trials found no beneficial effect of selenium supplements on cancer risk37. The relationship between selenium and cancer is therefore unclear. Consistent with our study, a long-term cohort study (28-year follow-up) among Italian citizens observed an increased risk of cancer with higher selenium concentration in the water supply system38. We therefore need to consider the inorganic form of selenium generally found in water because its biological properties may be markedly different from those of other chemical forms36. Laboratory studies have shown that the toxicity of inorganic (tetravalent) selenium greatly exceeds that of organic selenium39.

Notably, the Songhua & Liaohe River Basin had better overall water quality than the Huaihe and Haihe River Basins but had the highest incidence of liver and kidney cancer, second-highest levels of pancreatic, lung and breast cancer, and third-highest colorectal cancer levels. Air pollution has also been reported to be associated with cancer40, probably resulting from open burning of straw and other causes41. All these provide an explanation for this complex finding. Surface water pollution is likely one, but not the only, critical environmental driver of increased cancer.

One especially thought-provoking finding was from the Yangtze River Basin, which supports more than 40% of China’s population and nearly half of its economy. Along the west-to-east course of the Yangtze River, the per-capita gross domestic product increases (e.g., 57,532 CNY up-stream, 65,481 CNY mid-stream, and 104,031 CNY down-stream in year 2020)23, but the water quality decreases (possibly due to cumulative pollution from the branches, Supplementary Fig. 7) and cancer incidence increases. This contributes to very different pictures of the mode of impact of water quality on cancer across these three sub-basins, even though their water systems are closely connected. The up- and mid-stream basins had no cancer ranked among the country’s top three for population incidence, and there was little or no association between cancer and water quality. However, the incidence of all cancers except for liver and bone cancers in the downstream basin was among the top three (resulting in the highest combined incidence, 288.50 per 100,000). There were multiple water pollution indicators that could explain this, including total nitrogen, petroleum, permanganate index, zinc, F−, and volatile phenol (in descending order of influence). These patterns can be viewed as a mirror of China’s east–west development discrepancy, and convincing proof of the impact of the environment. It strongly suggests a need to end the approach that has traded environmental quality for economic growth, and instead makes systematic efforts towards environmentally-friendly and sustainable development. In 2022, the Chinese government issued the Action Plan for Deepening the Yangtze River Protection and Restoration Campaign, clarifying the outstanding ecological and environmental problems that need to be addressed by 202542. Our study provides a model to identify key targets that would allow effective, tailored countermeasures.

Rising to current and future challenges

Following our proposed 75th percentile thresholds and risk estimates corresponding to the CAMS grades, we estimated the population attributable fraction (PAF) and the number of water quality-related cancers (Supplementary Table 3). On average, 5.0% of all new cancers in current cancer registry areas can be attributed to poor quality surface water, with the most marked fractions in the Huaihe (9.3%), Songhua & Liaohe (7.3%) and Yellow River Basins (5.8%).

Our estimates provide a deep look at the impact of water quality on cancers, but they show only part of the environmental influence, which can be modified through proactive measures and government commitment. Another driving force in the cancer burden in China and many parts of the world, population aging, provides a dual challenge, and the impact is unlikely to cease in the foreseeable future43,44. The number of cancer cases worldwide is projected to double from the 2020 figures by 207045. Assuming no further improvement in water quality across China, we estimate that the number of new cancer cases will rise by 7.8%, to 874,218 in the present cancer registry areas by 2030, and an increasing number each year will be attributable to poor water quality during 2022–2030 (Fig. 3a). We estimate that this will amount to a total of 388,431 new cases. This is hypothetical, but highlights the urgency of acting to stop a rising tide of potentially preventable cancer.

Fig. 3: Rising to current and future challenges in surface water-related cancer.
figure 3

a Projected excess cases of total cancer caused by the degree of surface water pollution in cancer registry areas of China, 2016–2030. b Distribution of the amount of fertilizers applied in different provinces of China. c Distribution of wastewater discharge in different provinces of China.

Full size image

Finding and controlling the sources of pollution are likely to be the key solutions. Figure 3b, c shows the distributions of the amount of fertilizers applied and wastewater discharge in different provinces. This gives an insight into some of the sources of water pollution. The result suggests that there are multi-faceted, interrelated causes. First, to increase yields to supply the growing population, farmlands in China (e.g., in the Huaihe River Basin) used to apply large amounts of fertilizers (highly soluble ammonia fertilizers in particular), which drained into waterways and increased total nitrogen pollution46. Second, products (including these chemical fertilizers) are manufactured in industrial areas (mainly in the Haihe, Songhua & Liaohe, and Yangtze downstream River Basins), where large numbers of chemical compounds are discharged into water as part of their production processes47. Third, domestic sewage discharge is also high in these areas because of the high population density48. The total amount of wastewater in China reached a peak in 2015, having grown by an average annual rate of 2.8%, and has only recently slowed down by an annual rate of −2.5%23,49,50. Good management of chemical fertilizers and wastewater outfalls is suggested as the best way to improve water quality in China.

According to a survey in 2022, water pollution has replaced air pollution as the primary public concern in China51. This reflects both the urgency and societal will for change. However, just as pollution problems do not develop in a day, the effects cannot be eliminated rapidly. Realizing ambitious goals in environmental protection and population health requires a generational time frame and sustained input into policy, research, and most importantly, creation of an environmentally-friendly culture. Another implication for developing countries is that pollution source governance in industry and agriculture may have an impact on economic development. Investment in green technology is vital in balancing development and environmental protection52.

Breaking barriers in environmental health research

This paper proposes a paradigm in environmental health research, and therefore makes an important step in exploring the complex association between water quality and multiple cancers. It has integrated data from multiple sources in a unified spatial scale of river basins, a notable difference with the administrative system typically used in previous cancer research. One important barrier that hampered advances in both this field, and other environmental health fields, is the absence of a harmonized data framework. The environmental industry and the health industry have different focuses and operational structures (Supplementary Fig. 8), making the linkage between different data sources particularly challenging53. More cross-industry coordination is required to foster a more solid basis to build advances in data sciences and keep up with the increased concerns of the public about the environment.

A unique methodological challenge in assessing environment quality and its impact on health outcomes is the coexistence of multiple pollutants. Pollution often has a spatial autocorrelation, making multiple pollutants inter-correlated and giving different combination patterns in different regions. This phenomenon means that assessing health risks for individual pollutants is difficult because their interactions provide additional hazards that may not be explicable with current knowledge. To address these issues and a limit in the pairwise analysis of correlation in current methods, we proposed the CAMS design, which is comprehensive because it retains both the quality and spatial variation information. It is also versatile, transparent, easy-to-interpret and validated by our results. This design allowed us to restore the complicated real-world associations from previously separate findings, and generate knowledge about joint effects in a clear way. We hope this design may be useful for future studies and action to improve environmental and public health.

The EQSSW standards have been in use for 20 years in China, and many other countries have also not updated similar standards during the past two decades or more7. Using the current 75th percentile values, we found several pollution indicators were related to population cancer risk. We suggest that the EQSSW is not suitable for cancer risk assessment, and recommend that the 75th percentile thresholds should be updated annually to provide a dynamic goal for ongoing progress. This approach would also motivate local officers to benchmark against national levels. We used the national 75th percentile (i.e., the upper quartile) as a threshold because it allows for variation across river basins and is considered suitable to indicate high levels. Yet we do not suggest these values are necessarily the best cut-off values for safeguarding public health (for instance, the findings were similar if 80th percentile values were used; Supplementary Fig. 9). The negative findings for the other pollution indicators may be the result of not knowing the correct cut-off values to use. There are therefore many challenges in updating EQSSW and harmonizing standards across the world7.

In this study, we used annual average values of the water quality monitoring data to smooth temporal fluctuations. Similarly, population incidences, as parameters in each cancer registry, are relatively robust over time. Population-level data (i.e., parameter) has such good properties that help to address the issue of time lag between exposure and cancer in individual-level studies, and preserve spatial variations (see Supplementary Table 1) that are important to explore law of the nature. Analysis of these data has been widely embraced in a number of study fields, but very little in medicine. This may be because of the traditional focus on individuals, instead of a larger perspective embedded in a broader ecosystem, where each individual is a ‘cell’ that is deeply influenced by the surroundings. A shift from human-centred to ecosystem- or nature-centred perspectives would help to break traditional boundaries in health-related studies, and ultimately, create a better living environment for all.

Methods

Analytic workflow

Supplementary Fig. 10 gives an overview of the analytical framework used in this study.

Data source

Surface water quality data (2021) were derived from all (3632) monitored sections of the National Surface Water Environmental Quality Monitoring Network, Ministry of Ecology and Environment of China54. The surface water monitoring sections took into account the natural attributes of the basin such as basin area, river network density, runoff supply and hydrological characteristics.

Cancer incidence data were extracted from the Annual Report of the China Cancer Registry (2019), which was drawn from 487 registries in 31 provincial-level places in mainland China55. It covered a total population in China of 381.6 million people (193,632,323 males and 187,933,099 females), accounting for 27.60% (24.3% for urban areas and 32.0% for rural areas) of the national population at the end of 2016.

The data on the amount of applied fertilizers and wastewater discharge were extracted from the China Statistical Yearbooks and provincial statistical yearbooks (2019–2021)23. The data on forecast population size were obtained from the United Nations World Population Prospects (2019)56. Map data were obtained from the China Resource and Environment Science and Data Center.

Spatial analysis

Spatial mapping was used to show the spatial distribution and pattern of cancer incidence and surface water pollution. For spatial autocorrelation of each pollution indicator, we used the global Moran’s I (see formula (1)) and local Moran’s I (see formula (2)) to confirm the overall spatial autocorrelation and identify the pattern of local spatial clusters57,58,59.

$$I=\frac{n}{\mathop{\sum }\nolimits_{i=1}^{n}\mathop{\sum }\nolimits_{j=1}^{n}{w}_{ij}}\times \frac{\mathop{\sum }\nolimits_{i=1}^{n}\mathop{\sum }\nolimits_{j=1}^{n}{w}_{ij}({x}_{i}-\bar{x})({x}_{j}-\bar{x})}{\mathop{\sum }\nolimits_{i=1}^{n}{({x}_{i}-\bar{x})}^{2}}$$
(1)
$${I}_{i}=\frac{{x}_{i}-\bar{x}}{{{S}_{i}}^{2}}\times \mathop{\sum }\nolimits_{j=1,j\ne i}^{n}{w}_{ij}({x}_{j}-\bar{x}),{{S}_{i}}^{2}=\frac{\mathop{\sum }\nolimits_{j=1,j\ne i}^{n}{({x}_{j}-\bar{x})}^{2}}{n-1}$$
(2)

where \(n\) denotes the number of monitored surface water sections;\({x}_{i}\) the value of the pollution indicator for the \({i}^{th}\) section (\(i=1,2,\cdots ,n\)), \({x}_{j}\) the value of the pollution indicator for the \({j}^{th}\) neighbouring section (\(j=1,2,\cdots ,n;i\,\ne\, j\)); \(\bar{x}\) the average value of each pollution indicator; and \({w}_{ij}\) the neighbouring relations of sections \(i\) and \(j\).

We assumed that as the distance from a river increases, a concomitant decrease occurs in the probability that a population will be influenced by water pollution in the river. We selected 30 km as the distance beyond which the health effect from surface water pollution is negligible. This was determined using a pre-defined buffer analysis informed by previous studies on interaction between riverine chemistry and land use60 and on spatial distribution of persistent organic pollutant autocorrelation61. The buffer analysis process57 involves generating buffers around existing monitored surface water sections and then identifying or selecting areas based on whether they fall inside or outside the boundary of the buffer. The cancer incidence and rate ratios (RR) in areas of different surface water quality grade were calculated based on buffer analysis, and 95% confidence intervals (CI) of the RRs were derived from negative binomial regression models.

Statistical analysis

Categorical variables are shown as frequencies and percentages, and continuous variables as medians (25th percentile, 75th percentile) if they did not satisfy a normal distribution. Pairwise correlation cross-pollution indicators were examined using Spearman’s correlation analysis. Negative binomial regression24 was used to test the effect of single pollution indicators on different cancers.

To make full use of the original values of the pollution indicators and accommodate the complex correlations between them, we conducted extreme gradient boosting (XGBoost), a scalable tree boosting and effective machine learning algorithm, to confirm our findings62. However, one of the disadvantages of this approach is the limited interpretability of the complex underlying feature interaction and non-linear structure. On the basis of game theory, SHAP values allowed us to estimate the mean impact on model output magnitude for each of the input features63. SHAP values were generated from the XGBoost algorithm (parameters tuned: learning rate = 0.04, max depth = 3, subsample ratio = 0.8), to allow us to identify the most important pollution indicators affecting multiple cancers and the primary pollution indicators affecting cancers in each river basin.

The excess cases (EC), i.e., expected cancer reductions, were computed via population attributable fractions (PAFs) (see formula (3))64,65.

$${\rm{PAF}}=\frac{{\sum }_{i=0}^{3}{p}_{{e}_{i}}{{\rm{RR}}}_{i}-1}{{\sum }_{i=0}^{3}{p}_{{e}_{i}}{{\rm{RR}}}_{i}}$$
(3)

where pei denotes the prevalence of exposure i among the total population, that is, the proportion of population at Grade i, and RRi the relative risk of cancer incidence at Grade i. The number of EC was given by \({\rm{EC}}={\rm{No.}}\,{\rm{of}}\,{\rm{incidence}}\times {\rm{PAF}}\). The prevalence of graded exposure levels was evaluated by Thiessen polygon analysis57. Considering the trend of cancer incidence in 2010–201625,55, we used linear regression to project an annual total incidence of the selected cancers in 2017–2030.

We conducted the data integration and negative binomial regression in SAS 9.4 software (SAS Institute Inc., Cary, NC, USA). The XGBoost algorithm and SHAP plots used the R package XGBoost66 and SHAPforxgboost67. All spatial analyses used the ArcMap module in ArcGIS 10.8 software (ESRI, Redland, CA).