Street view images help to reveal the impact of noisy environments on the survival duration of stroke patients

Abstract

Background

While road traffic noise is an emerging environmental risk for cardiovascular mortality, its age-group-specific effects on stroke mortality remain unclear. This study further explored socioeconomic disparities in this association.

Methods

We conducted a retrospective cohort study (2011–2019) with 36,240 hospitalized stroke patients in Fuxin, China. Residential noise levels were estimated using street view imagery analyzed by a novel and multimodal deep learning model. Age-grouped cox proportional hazards models adjusted for NO₂, NDVI (Normalized Difference Vegetation Index), and sociodemographic covariates were applied to assess mortality risks.

Results

Among elderly patients aged 60 years with lower medical insurance, each 5-dB increase in residential road noise was associated with a 93.6% increase in stroke mortality risk (HR = 1.936, 95% CI: 1.024-3.660; p = 0.042). The estimated exposure prevalence in this subgroup was 3%, yet the population attributable fraction reached 1.7%. In contrast, no significant associations were found among patients with higher insurance coverage. Younger Males had a 51.3% higher mortality risk than females (adjusted HR=1.513, 95% CI: 1.142-2.005), independent of environmental exposures. NO₂ and NDVI were not significantly associated with mortality across subgroups.

Conclusions

These findings highlight the need for noise mitigation strategies that prioritize vulnerable populations, particularly the elderly and those with limited healthcare access.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12942-025-00416-8.

Introduction

Chinese society has experienced notable aging in recent years. According to the latest data from the China National Bureau of Statistics(NBSC), by the end of 2024, the number of people aged 60 and older will exceed 310 million, Making up approximately 22% of the total population, which categorizes China as a moderately aging society [1]. China is expected to become a severely aging society by 2035 when almost 30% of the population will be 60 years of age or older. This demographic shift highlights the urgent need for comprehensive policies and infrastructure to support the aging population, particularly in healthcare areas. As the population ages, the prevalence of age-related health issues, including cardiovascular diseases and stroke, is expected to rise [2].

Stroke is one of the leading causes of death and disability globally and the estimated global cost of stroke in 2022 is more than US $721 billion (0.66% of the global GDP) [3]. In 2019, the number of fatalities attributed to stroke surpassed 6 million, constituting 11% of all deaths worldwide [4]. China is particularly burdened by the prevalence of cerebral infarction. In 2018, the mortality rate for cerebrovascular diseases in China reached 149.49 per 100,000 individuals, resulting in 1.57 million deaths [5]. The total cost of hospitalization for cerebrovascular diseases is 136.028 billion Yuan, of which cerebral infarction accounts for 81.197 billion Yuan, and cerebral hemorrhage accounts for 29.633 billion Yuan [6]. These statistics underscore the profound impact of cerebral infarction on the health and economic security of China. Current research identifies several modifiable risk factors that significantly contribute to the incidence of stroke and related disabilities. These include smoking, a diet high in sodium, elevated body mass index (BMI), increased systolic blood pressure (SBP), elevated levels of low-density lipoprotein (LDL) cholesterol, renal dysfunction, and high fasting plasma glucose (FPG) [7]. Furthermore, emerging evidence suggests that environmental factors, such as noise pollution, may also play a key role in the stroke [8–11].

Noise pollution, defined as unwanted or harmful sound that disrupts normal activities and reduces the quality of life, is an omnipresent issue in urban environments. Sources include vehicular traffic, industrial activity, and recreational events. Among these, traffic noise is one of the most prevalent and persistent, affecting millions of people daily [12]. WHO has recognized environmental noise as a substantial public health threat, linking it to a variety of adverse health outcomes, including hearing loss, sleep disturbances, and mental health [13–15].

The streetscape plays a critical role in the shaping of urban environments and influencing public health. The Street View encompasses elements such as the width of the road, the number of cars, the height of the building, the green space, and the pedestrian pathways, which can significantly affect the propagation of noise and the levels of exposure [16]. For instance, narrower streets with high-rise buildings tend to create canyons, leading to higher levels of exposure for residents and pedestrians [17]. Conversely, green spaces and vegetation can mitigate noise levels through sound absorption, providing a natural barrier against noise pollution [18]. In recent years, map services and geo-tagged imagery, such as Baidu Street View, have emerged as a vital resource for street-level imaging, containing high-resolution visual data across a wide range of urban streets and neighborhoods [19]. Significant progress has been made in automated semantic information extraction, visual element classification, and large-scale noise pollution studies thanks to street-level photos [20–23]. Street View is cost-effective and widely accessible, allowing for extensive noise assessments across large less developed areas without the need for expensive equipment or extensive fieldwork.

Given these limitations in the literature, our study aims to investigate the relationship between long-term residential exposure to traffic-induced air pollution (measured as nitrogen dioxide, NO₂), green space (measured by NDVI) and road traffic noise with stroke mortality in a large cohort. We seek to explore the effects of noise pollution independently and in conjunction with traffic-induced air pollution and green space, examining whether combined exposure exacerbates the risk of stroke mortality. Moreover, we aim to determine if elderly individuals are more susceptible to the adverse effects of noise pollution compared to younger populations, potentially leading to higher mortality rates in this vulnerable group. By shedding light on this critical environmental health issue, we hope to contribute to the development of effective public health strategies and urban planning policies that mitigate the adverse effects of noise pollution on vulnerable populations.

Methods

Study region

We selected Fuxin, a city at the prefecture level in Liaoning Province, northeast China, as our study region. Covering an area of 10,355 km, Fuxin had a population of 1.837 million at the end of 2019. The sex distribution was nearly equal, with 909,000 men (49.5%) and 928,000 women (50.5%). In particular, the population aged 60 and above was 451,000, Making up 24.6% of the total population, and Fuxin experienced a natural population growth rate of −2.05 per 1,000, showing a minor population fall and highlighting the transition to a super-aged society. Fuxin’s infrastructure includes a total road length of 7,753 km, with 311 km of highways. The city had 287,100 civilian cars, comprising 247,000 passenger cars and 39,400 freight vehicles. As a provincial garden city, Fuxin has a forest reserve of 2,728 km, with a forest cover of 26.34%.

Fuxin, encompasses five districts: Haizhou, Xihe, Taiping, Xinqiu, and Qinghemen, as well as two counties: Fuxin Mongol Autonomous County and Changwu County. This diverse urban and transportation landscape allowed us to analyze various street environments, providing a comprehensive view of the impact of noise pollution. Figure 1 illustrates the research region, highlighting the different districts and counties within Fuxin. This map serves as a visual aid to understand the geographical position of our study area.

Fig. 1.

Geographical location and administrative divisions of the study area, Fuxin City, Liaoning Province.​​​​ a Position of Liaoning Province within China; ​​​b Position of Fuxin City within Liaoning Province; ​​c Detailed map of Fuxin City, showing its central districts and surrounding counties

Data

Road network and street view

To analyze the impact of noise pollution on stroke risk, we used OpenStreetMap (OSM), a publicly accessible and collaborative spatial database, to collect detailed road network data for Fuxin. OSM provides comprehensive annotations at the street level, including road types and locations, essential for our analysis. Using data from OSM road network, we generated sampling points at 50-meter intervals throughout the city. The precise latitude and longitude of these points were then used to gather Street View photos via the Baidu Street View API.

The Baidu Map Open Platform’s Panorama Static View API enabled us to capture detailed panoramic views of streets along roadways. The Baidu Street View (BSV) service offers extensive geographic coverage and fast and high quality images, making it an ideal tool for visualizing urban areas in China. Each high-resolution panorama image (1024 by 360 pixels) was meticulously annotated with a unique place ID number and precise latitude and longitude coordinates. The images were captured with azimuth and elevation angles set at 0 degrees, ensuring consistency across all visuals. We acquired 17,467 Street View images in 2019 for the daytime hours in Fuxin.

Referring to the Street View to Noise model developed by our group [24], we estimated noise levels based on these street views. This approach allowed us to accurately assess road noise pollution in Fuxin. This extensive noise dataset forms the foundation of our research, enabling us to investigate the impact of noise pollution on stroke risk with high precision.

Patient profile data

Patient profile data was obtained from medical records at Fuxin General Mining Hospital and the Municipal Hospital, both renowned Class A tertiary hospitals where nearly all neurological inpatients in Fuxin seek treatment. Stroke deaths were identified through a comprehensive link with national health registries, hospital records, and death certificates. Stroke cases were defined according to the criteria of the World Health Organization and validated by trained medical professionals. All personally identifiable information, including names, ID numbers, and contact details were irreversibly anonymized. Therefore, ethical approval is exempted.

From the initial data set of 42,866 stroke patients, drawn from these two hospitals, we applied stringent screening criteria to refine the cohort. Patients without street sampling points within 100 m and those with incomplete disease data were excluded. This meticulous process resulted in a refined cohort of 36,240 patients. This cohort comprised residents admitted for stroke events between January 2011 and May 2019, with profile data including age, sex, and residence at the time of admission. Information on patients’ medical payment methods was recorded, including out-of-pocket and various insurance types (e.g., urban employee insurance, public medical care). These categories served as socioeconomic status (SES) proxies. Out-of-pocket payment was classified as the low-insurance group (coded 1), while all other payment types comprised the high-insurance group (coded 0, reference). Residence addresses were geocoded to obtain longitude and latitude coordinates using a geocoder in the WGS coordinate system. To ensure patient confidentiality, we introduced a random perturbation of 100 m to the patients’ addresses (Figure 2). We assumed that the patient’s residence address remained the same throughout the survey period.

Fig. 2.

Spatial distribution of stroke patients’ residences. a Panoramic view showing Fuxin's administrative boundary; b Case distribution in Changwu County; ​​c Case distribution in central urban area;​​ ​​d Case distribution in Qinghemen District​ (A 100-m random perturbation was applied to protect privacy)

Noise properties

The noise dataset (n=17,467 points) was generated using the deep learning method detailed in Section 2.3. The derived values exhibited a normal distribution with a long tail >80 dB (Figure 3), ranging 64-91 dB (mean=74.3 dB, SD=3.8 dB). The presence of this tail indicates that there are outliers or extreme values that raise noise levels above the average, typically due to major roads or intersections with heavy traffic. Spatial autocorrelation analysis confirmed significant clustering (Moran’s I=0.848, p<0.001).

Fig. 3.

Frequency distribution of estimated road traffic noise in Fuxin

Covariates

We used NDVI to quantify the green space around residential areas. This index was derived from data captured by the Moderate-Resolution Imaging Spectroradiometer (MODIS) at a resolution of 1000 m by 1000 m. We computed the average NDVI for the period from January 1, 2019, to December 31, 2019.

To measure traffic-induced air pollution, we used Sentinel-5P data for the year 2019. Sentinel-5P, launched on October 13, 2017, is equipped with the Tropospheric Monitoring Instrument (TROPOMI), which provides daily global coverage and detailed information on atmospheric gases and pollutants. We collected data on nitrogen dioxide with a resolution of 1000 m by 1000 m and calculated the average NO₂ level for the entire year of 2019 to represent traffic-induced air pollution.

Based on patients’ residential addresses, we extracted the corresponding NDVI and NO₂ values from the raster data. These environmental data points were then incorporated into our analysis as confounders. By including NDVI and NO₂ values in the Cox regression analysis, we aimed to isolate the specific influence of noise pollution on stroke incidence, accounting for the potential impact of greenspace and air quality on health outcomes. The study assumes that the spatial distribution and the values of NDVI and NO₂ remain constant throughout the study period.

Noise estimation

This section details the methodology for generating the noise dataset characterized in Section 2.2.3. Using street view images, we inferred traffic noise levels along the road network by employing an innovative framework that integrates classification and regression techniques [24]. This approach provides end-to-end output from street view photos to road traffic noise levels, measured in decibels. The DCNN-RF model within the street-view-based noise estimation framework achieves an R of 0.64, with MaE and RMSE values of 2.01 dB and 2.71 dB. This method can estimate noise levels for any given street view image, successfully balancing efficiency and accuracy, as confirmed by previous studies [25, 26].

In this approach, a deep convolutional neural network (DCNN) is trained to recognize traffic noise-related features in the road environment. A SoftMax Function retrieves the output probability vector, generating a 5-dimensional vector (Q = q1, q2, q3, q4, q5) that represents the estimation probability for each image category. For example, q1=0.2 implies that the image has a 20% chance of falling into the category ’low noise’(level 1). We then extract the Q training feature vector and their related labels and use regression algorithms to generate numerical estimates of road traffic noise in decibels.

To apply this innovation to obtain road noise values around the patients, we first converted the captured road street scenes into noise points. A 100-meter buffer was then created around each patient’s location. The lowest noise level within this buffer was calculated as the baseline noise level perceived by the patients around their homes. The road traffic noise estimate model can be obtained from this repository: https://github.com/kellyhuang-gis/traffic-noise-estimation.

Statistical methods

This observational cohort study analyzed 36,240 stroke patients from Fuxin using age-grouped Cox proportional hazards models. The final analytical sample is classified by payment method as a proxy for socioeconomic status: 2,604 in the low-insurance group (out-of-pocket payment, coded 1) and 33,636 in the high-insurance group (public-funded coverage, coded 0, reference). The survival time was defined from hospital admission to death. Road traffic noise exposure was analyzed as a continuous variable scaled per 5 dB increment. Sex was treated as a binary covariate (male=1, female=0), with females serving as the reference group. Environmental covariates including NO₂ and NDVI were standardized using z-score normalization (mean = 0, standard deviation = 1).

The analytical workflow progressed through three sequential phases. First, age-dependent effect modification was evaluated by introducing interaction terms (age  noise and age  sex) into a Cox proportional hazards model. A statistically significant age  noise interaction () justified subsequent grouping at the clinically relevant 60-year threshold as defined by WHO [27].

Second, age-grouped Cox models for 60y and >60y groups were fitted with two adjustment sets: a reduced model (noise and sex) and a full model (noise, sex, NO₂, NDVI, insurance). For the reduced model, proportional hazards assumptions were satisfied (p>0.05; Supplement Table 1), but Schoenfeld residual tests indicated violation of PH in the >60y stratum (p=0.032) for the full model (Supplement Table S2), prompting further stratification by payment type for this subgroup. Consequently, separate Cox models were fitted: one for the low-insurance group and another for the high-insurance group, both satisfying PH assumptions (p>0.05; Supplement Table S3).

Table 1.

Baseline characteristics of the study population by age groups

Characteristic	<60 years N=11,404	60 years N=24,836	Overall N=36,240
Age (Mean ± SD)	52.5 ± 5.3	72.8 ± 7.9	66.4 ± 11.9
Noise level (dB) (Mean ± SD)	73.1 ± 2.0	73.0 ± 2.1	73.0 ± 2.1
NO2 () (Mean ± SD)	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0
NDVI (Mean ± SD)	0.2 ± 0.0	0.2 ± 0.0	0.2 ± 0.0
Sex
Male (%)	8,147 (71.4)	13,434 (54.1)	21,581 (59.6)
Female (%)	3,257 (28.6)	11,402 (45.9)	14,659 (40.4)
Insurance Group
High insurance (%)	10,458 (91.7)	23,178 (93.3)	33,636 (92.8)
Low insurance (%)	946 (8.3)	1,658 (6.7)	2,604 (7.2)
Mortality outcomes
Events (%)	157 (1.4)	868 (3.5)	1,025 (2.8)
Censored (%)	11,247 (98.6)	23,968 (96.5)	35,215 (97.2)

Finally, sensitivity analyses evaluated robustness: E-values quantified unmeasured confounding strength needed to nullify observed effects, and population attributable fractions (PAF) were calculated for significant exposures using stratum-specific prevalence. All analyses used Python 3.9 (lifelines 0.27.7).

Results

Description of the cohort

The study cohort comprised 36,240 participants with distinct age-grouped characteristics (Table 1). Younger adults (<60 years, n=11,404) demonstrated male predominance (71.4% vs 54.1% in elderly, ²=785.6, p<0.001) with mean age 52.5±5.3 years, while elderly participants ( 60 years, n=24,836) had higher crude mortality rates (3.5% vs 1.4%, p<0.001). Insurance coverage differed, with higher high-insurance proportions in the elderly (93.3% vs 91.7% in younger adults). Noise exposure levels were consistent across groups (overall 73.0±2.1 dB), with environmental indicators (NO₂ and NDVI) showing homogeneous distributions. The cohort accumulated 1,025 mortality events during follow-up, with elderly participants accounting for 84.7% (868/1,025) of total events.

Descriptive statistics of estimated road traffic noise

Figure 4 illustrates the distribution of noise levels in Fuxin. The road traffic noise in the study area was systematically divided into small squares of 200 m 200 m. Road noise is graded into four categories: 70, 70.1-75, 75.1-80, and >80 decibels. The Moran’s I index for noise in Fuxin is 0.848, indicating a clear spatial clustering effect. This means that noise pollution is not randomly distributed, but tends to be concentrated in specific areas.

Fig. 4.

The spatial distribution of noise. a Panoramic view of noise distribution across Fuxin's administrative area; ​​b Noise distribution in ChangWu County; ​​c Noise distribution in central urban area; ​​d Noise distribution in Qinghemen District

The noise distribution in Changwu is shown in the figure 4(b). One major source of excessive noise levels in Changwu is primary north-south roads, such as the G304 and G101 National Highways. These are primary urban routes and freeways with heavy traffic, which increases noise levels due to their volume and speed. The figure 4(c) depicts the noise distribution in Xihe and the adjacent areas, which include Haizhou, Taiping, Xinqiu, and Fuxin Mongol Autonomous. Major urban roads and expressways, such as S204, and G2513, are significant sources of noise. These highways are primarily metropolitan major roads and high-traffic secondary roads that connect important commercial and residential areas. The figure 4(d) depicts the noise dispersion in Qinghemen. The noise in Qinghemen is gathered, particularly in the center sections and around major traffic nodes. The district’s primary and secondary roads, which are used for urban traffic and logistical transportation and have a high traffic volume and speed, are the main sources of noise.

Figure 5 below shows the transition from street level to noise level using the innovative Street View to Noise model. The first BSV, characterized by abundant green trees minimal traffic, and the substantial presence of vegetation likely mitigates noise pollution by absorbing and deflecting sound waves, resulting in a quieter environment. It has the lowest noise level at 65.3 dB. In contrast, the second BSV has less vegetation, and denser residential buildings compared to the first BSV. The estimated noise for the second street view is 71.6 dB. The third BSV is a scenario in a residential area near an intersection with more economic activity, resulting in a noise level of 75.9 dB. The rise in noise is caused by increasing traffic flow, higher building complexes, and more intense commercial activity... The fourth BSV, which depicts an expressway, exhibits the highest noise level at 82.4 dB. The absence of vegetation and the presence of high traffic volumes with fast-moving vehicles significantly contribute to the elevated noise levels. This analysis underscores the impact of urban planning on noise levels, demonstrating that areas with more greenery and less traffic tend to be quieter. In contrast, regions with high traffic volumes, commercial activities, and fewer natural noise barriers are significantly noisier.

Fig. 5.

Representative street-view images and noise estimation examples. a Tree-lined street (Pred: 65.3 dB); b Urban arterial road under clear conditions (Pred: 71.6 dB); c High-traffic intersection (Pred: 75.9 dB); d Major roadway with construction activity (Pred: 82.4 dB)

Association of noise on stroke mortality

Age significantly modified the association between road-traffic noise exposure and stroke mortality, with a 2.5% annual increase in hazard ratio per year of aging (HR=1.025/year, 95%CI=1.01–1.04; ), culminating in a 28% elevated risk per decade (Table 2). In contrast, age did not alter sex disparities (age sex HR=1.002, 95% CI=0.86–1.16; ), maintaining stable male predominance across all age strata.

Table 2.

Age Interactions of noise exposure and sex with stroke mortality risk

Variable	HR (95% CI)	HR per decade	p-value
Age Noise	1.025 (1.01–1.04)	1.28	<0.001
Age Sex	1.002 (0.86–1.16)	1.02	0.982

Age-grouped analyses revealed divergent risk profiles (Table 3–4). Within the full model framework (adjusted for noise, sex, NO₂, NDVI, and insurance), distinct mortality risk patterns emerged across age groups. Among participants aged <60 years, Male exhibited a 51.3% higher mortality risk (95% CI=1.142–2.005; ) with PAF of 24.2%, while noise effects remained non-significant (HR=1.002 per 5 dB, 95% CI=0.737–1.361; ). Conversely, in the 60 years cohort, per 5 dB noise elevation was associated with a 21.1% increased mortality risk (95% CI = 1.038-1.414; p = 0.015), while sex was not a significant factor (HR = 0.956, 95% CI = 0.849-1.076; p = 0.457). Environmental covariates remained non-significant across all models.

Due to violation of the proportional hazards assumption for payment type (p < 0.005) in supplement table 2, stratified analyses by medical payment method were conducted within the 60 years group (Table 5). In the low insurance subgroup, traffic noise was significantly associated with mortality (HR = 1.936 per 5 dB, 95% CI = 1.024-3.660; p = 0.042), with an estimated E-value of 3.28 and a PAF of 1.7% (exposure rate = 3%). In contrast, no significant association was observed in the high insurance group (HR = 1.156, 95% CI = 0.986-1.356; p = 0.074). These findings suggest a potential vulnerability to traffic noise among older adults with lower medical insurance.

Table 5.

Cox regression results grouped by medical payment type among patients aged 60 years

	Covariate	HR (95% CI)	p
High insurance	Sex (Male)	0.954 (0.844–1.078)	0.447
	Noise (5dB)	1.156 (0.986–1.356)	0.074
	NO2 (scaled)	1.03 (0.96–1.106)	0.41
	NDVI (scaled)	0.978 (0.917–1.042)	0.494
Low insurance	Sex (Male)	1.126 (0.694–1.83)	0.63
	Noise (5dB)	1.936 (1.024–3.66)	0.042
	NO2 (scaled)	0.903 (0.753–1.082)	0.267
	NDVI (scaled)	1.066 (0.861–1.32)	0.558

Multivariable validation revealed exceptional model stability. Sensitivity analyses confirmed resistance to unmeasured confounding (E-values 3.28 for noise) and validated proportional hazards assumptions (Schoenfeld test ). Complete test statistics and p-values stratified by insurance groups are provided in supplement tables 3. Population-attributable fractions aligned with exposure prevalence, highlighting age-specific intervention priorities.

Table 3.

Relationships between exposure to road traffic noise and stroke mortality

Age	Covariate	HR (95% CI)	p
<60	Sex (Male)	1.516 (1.144–2.008)	0.004
	Noise (5 dB)	1.023 (0.762–1.372)	0.882
60	Sex (Male)	0.950 (0.844–1.069)	0.397
	Noise (5 dB)	1.217 (1.054–1.405)	0.007

Table 4.

Age-stratified Cox regression results on mortality risk factors

	Covariate	HR(95% CI)	p
<60	Sex (Male)	1.513 (1.142–2.005)	0.004
	Noise (5dB)	1.002 (0.737–1.361)	0.992
	NO2 (scaled)	1.061 (0.936–1.202)	0.357
	NDVI (scaled)	0.969 (0.855–1.098)	0.619
	Low insurance	1.012 (0.649–1.579)	0.957
60	Sex (Male)	0.956 (0.849–1.076)	0.457
	Noise (5dB)	1.211 (1.038–1.414)	0.015
	NO2 (scaled)	1.013 (0.947–1.083)	0.712
	NDVI (scaled)	0.990 (0.931–1.053)	0.754
	Low insurance	0.823 (0.65–1.043)	0.107

Discussion

Comparative analysis of noise-induced stroke mortality risk

This study reveals that road traffic noise significantly impacts the risk of stroke mortality in individuals aged 60 and above. The Cox regression analysis indicates that for every 5-decibel increase in noise among hospitalized elderly, the stroke mortality risk increases by approximately 21.3%(HR=1.213, 95% CI: 1.039-1.416, p=0.014).

Our HRs were relatively high compared to other studies. For example, one study reported an ischemic stroke mortality HR of 1.05 (95% CI: 1.002-1.099) per 10 dB increase for participants aged 30 years and older [28]. Similarly, a Danish cohort study found a 10-year traffic noise-weighted stroke mortality risk ratio of 1.12 (95% CI: 1.00-1.26) for individuals aged 50-64 in areas with higher noise exposure (IQR: 10.4 dB) [29]. A meta-analysis of 311,878 stroke cases reported a modest 3% elevated mortality risk per 10 dB(A) noise increment (pooled HR=1.03, 95% CI:1.00-1.07), though the confidence intervals spanning unity suggest limited statistical certainty in this association [30]. However, a study also showed a relative risk (RR) of 1.02 (95% CI: 0.98-1.06) for stroke mortality associated with noise exposure in adults aged 25 years and older, which was not statistically significant [31].

These differences may be attributed to several factors. Firstly, the study focused on hospitalized patients, a population that is inherently more susceptible to environmental stressors. This increased sensitivity may lead to a relatively higher observed risk associated with noise exposure in this group. Additionally, the total number of hospitalized stroke patients isn’t large enough, resulting in a greater risk of stroke mortality compared to the general public. This dynamic may also amplify the association between noise exposure and stroke mortality. Secondly, variations in noise exposure assessment methods can lead to different outcomes. Our study utilized a novel noise estimation model based on street view images, which may provide a more efficient and comprehensive assessment of road noise compared to traditional methods used in previous studies. This innovative approach allows for a more detailed spatial analysis of noise pollution, potentially capturing nuances that might be overlooked by conventional measurement techniques. Thirdly, the urban context and population demographics of Fuxin differ significantly from those in European settings, which might influence the observed associations between noise and stroke mortality. For instance, the higher proportion of elderly individuals and distinct urbanization patterns in Chinese small cities could exacerbate the impact of noise pollution on health outcomes. The rapid urban development and high population density in Fuxin might result in greater exposure to traffic noise, particularly among vulnerable groups such as the elderly, who are more susceptible to its adverse health effects.

Mechanisms of noise-induced stroke risk

The results of our investigation showed that noise significantly increases the risk of stroke in the hospitalized elderly. The Cox regression analysis, focusing on hospitalized elderly patients aged 60 and above, demonstrates that the likelihood of stroke-related mortality escalates by approximately 21.3% (95% CI: 1.004-1.568, p=0.047) with each 5-decibel increment in ambient noise levels, explained by several factors that have an impact on the physical and mental health of the elderly.

Noise exposure has been shown to trigger oxidative stress and inflammation, contributing to a range of neurodegenerative diseases. The underlying mechanisms involve the activation of the hypothalamic-pituitary-adrenal (HPA) axis and the sympathetic nervous system, which lead to increased levels of inflammation and oxidative stress. These stress responses elevate the production of reactive oxygen species (ROS) and inflammatory cytokines, creating an environment that fosters cellular damage and neurodegeneration [32, 33]. Furthermore, long-term exposure to high noise levels can cause endothelial dysfunction and increase the likelihood of arteriosclerosis. This is particularly concerning as endothelial dysfunction impairs the ability of blood vessels to dilate properly, contributing to hypertension and atherosclerotic plaque formation, which are significant risk factors for stroke [34].

In addition to its physiological effects, noise exposure also negatively impacts mental health. Prolonged exposure to high noise levels can lead to anxiety, depression, and other negative emotional states. Psychological stress and anxiety can influence the nervous and cardiovascular systems through various pathways, ultimately contributing to stroke. Moreover, noise negatively affects cognitive functions, especially in the elderly population. Declining cognitive function can increase individual psychological stress, further raising the risk of stroke [35].

Large-scale estimation of urban road traffic noise exposure made possible

Our study employs an advanced multimodal model that translates 17,467 BSV into auditory noise level predictions. This cutting-edge approach integrates machine learning techniques to analyze street view images, subsequently estimating the associated noise levels. By converting visual information into auditory insights, our model offers several significant advantages. Firstly, this model is highly cost-effective as it utilizes existing street view data to estimate noise levels, thereby eliminating the need for extensive new data collection. Traditional on-site noise monitoring, while precise, demands substantial investments in equipment, time, and personnel, making large-scale or continuous monitoring impractical. Similarly, physical models, despite their detailed output, require complex environmental data, making their development and validation labor-intensive and expensive. Surveys, although capable of capturing subjective noise perceptions, often suffer from biases and limited sample sizes, which affects data reliability. Consequently, large-scale noise prediction has traditionally been a formidable challenge due to these extensive demands on equipment and manpower. The high efficiency of the Street View to Noise model effectively overcomes these limitations by providing a scalable and practical solution.

Secondly, the model has strong transferability. This transferability is supported by its training on a varied dataset that containsins a wide range of road types, including rural roads, suburban routes, acity-centerter streets. Because of its extensive diversity, the noise prediction model can successfully adapt to different urban environments, ensuring accurate noise level estimations across a wide range of geographical situations. By incorporating diverse road conditions during the training phase, the model shows increased robustness and the ability to transfer findings to other sites, confirming its application in a variety of contexts.

Lastly, the model is adept at estimating historical noise levels. By leveraging past local street view, it can effectively infer historical noise levels and analyze long-term noise fluctuations within the city. This capability helps for the noise identification of specific causes behind these fluctuations. Such historical noise data cannot be captured through conventional on-site measurements, making this model indispensable for retrospective noise analysis.

Limitations

This study has certain limitations that must be acknowledged. Firstly, we lacked data on socio-economic status (SES), smoking, and alcohol consumption. These variables could impact the Cox regression analysis as potential confounders. Although medical payment methods served as a practical proxy for socioeconomic status (SES), this single-dimensional indicator cannot fully capture multifaceted SES constructs (e.g., education, occupation, income). Future studies should incorporate comprehensive SES and lifestyle data to provide a more nuanced analysis. In addition, the street view used for noise estimation was captured during the day, excluding nightly levels. Nighttime noise can have different health impacts compared to daytime noise. Subsequent studies should complement daytime data with nighttime noise measurements to provide a more comprehensive understanding of noise exposure and its health effects. Another limitation is that our data on urban traffic-induced air pollution was limited to nitrogen dioxide (NO₂). Recent research suggests that NO_x is often associated with other pollutants related to transportation such as PM₂, CO, and VOCs, which are highly correlated. Since motor vehicle emissions are an important source of these pollutants, NO₂ concentrations can reflect the general levels of traffic pollution in many urban environments[36]. Future studies should consider a broader range of air pollutants to better capture the complexity of urban air pollution and its health impacts. Finally, our study did not account for temporal variations in environmental exposures, such as seasonal fluctuations in noise/NO₂ levels or longitudinal changes over the 8-year cohort period (2011-2019), which may attenuate estimates of health effects. To mitigate these limitations in future research, we will apply land - use regression models with time series satellite imagery. The satellite imagery offers a broad - scale view of long - term environmental changes. Integrating these data into the models helps estimate long-term exposure changes, especially when ground-based measurements are scarce.

Conclusions

Using an innovative street view-to-noise model, we achieved a precise assessment of noise exposure levels. Stratified analyses by medical payment type revealed that, among hospitalized elderly patients aged 60 and older with lower medical insurance, a 5-decibel increase in road noise was significantly associated with a 93.6% increase in stroke mortality risk (HR = 1.936, 95% CI = 1.024-3.660; p = 0.042). Although the exposure prevalence in this subgroup was relatively low (3%), the estimated population attributable fraction (PAF) reached 1.7%, suggesting a non-negligible public health burden concentrated among socioeconomically disadvantaged individuals. These findings underscore the critical need for targeted noise abatement strategies and equitable healthcare access for vulnerable populations. Moreover, this study exemplifies the utility of advanced noise modeling techniques in epidemiological research.

Author contributions

JX: Writing-original draft, visualization, writing-review & editing, formal Analysis. TF: Supervision, methodology, writing-review & editing. BY: Resources, methodology. YH: Resources, methodology. YD: Study oversight, funding acquisition.

Funding

This work was supported by the National Natural Science Foundation of China General Program (Grant No. 42571486 & 42271476), the National Key Research and Development Program of China (Grant No. 2024YFC3307601; 2024), and the Open Funding from State Key Laboratory of Resources and Environmental Information System.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was granted ethical exemption by the School of Resource and Environmental Science, Wuhan University. The research utilized anonymized public datasets: medical records stripped of personal identifiers and street view images with geographic coordinates perturbed by 100 meters. Informed consent was not required under national regulations for non-identifiable secondary data analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.