DataBreast cancer screening participation data
We analyzed data from 589,879 invitations (135,772 women) from the Geneva canton’s BCS program (1999–2020), targeting women aged 50 to 7421,22 (Supplementary Fig. 1). The first invitation is sent when women turn 50, with subsequent invitations issued every two years after a screening mammogram or following the previous invitation if no response was received. Each invitation record includes the date of invitation, the participation status (Yes/No), residential address, and age group. For women who actually participated in mammography screening (i.e., participants), additional information on the chosen screening center was recorded. Additional details on the screening program’s historical context are provided in Supplementary Note 1. The study was conducted in collaboration with the Geneva Foundation for Cancer Screening, who provided access to anonymized invitation records from their administrative database. Ethical approval was obtained from the Cantonal Research Ethics Commission of Geneva, Switzerland (reference number 2023-02097, 15.01.2024). The requirement for informed consent was waived by the ethics committee under Article 34 of the Swiss Human Research Act. This waiver was granted due to the retrospective nature of the study using coded data from over 100,000 participants across 2003–2020, where obtaining individual consent would be impracticable.
We identified a systematic underestimation of participation in the data covering the early implementation phase of the BCS program (1999–2003), these years were therefore excluded from analysis. The final dataset comprised 482,318 invitations (118,232 women) covering the period 2003–2020.
Residential addresses were geocoded to obtain spatial coordinates for subsequent spatiotemporal analysis using a string-matching algorithm against the Swiss Buildings database (RegBL)23. The algorithm first standardizes addresses by removing special characters, expanding abbreviations (e.g., “ch.” to “chemin”), and extracting street numbers. It then employs a two-step matching process24: quick geocoding matches addresses within the same postal code using sequence similarity ratios, and if match quality is insufficient, slow geocoding expands the search to canton level. When street numbers were missing or not found, street-level geocoding was performed using the centroid of the street. Similarity scores are calculated using the SequenceMatcher algorithm from the difflib package (v3.13.7)25 in Python (v3.11), with matches above an 80% threshold being classified as acceptable quality. This threshold was selected to optimize the balance between false positives (incorrect matches passing the threshold) and false negatives (existing matches not being identified). Manual review and cleaning were conducted for addresses near the 80% threshold to correct false negatives (valid matches below threshold) and false positives (invalid matches above threshold). This completely offline procedure avoids the need to send information online (e.g., via APIs such as Google Maps) and therefore ensures improved data protection.
Exclusion criteria were applied to align with national age guidelines and to ensure geocoding accuracy (details in Supplementary Note 2).
Analytical data structure
Our analysis employed both individual-level and area-level approaches. Individual-level invitation records were used for descriptive analyses of participation trends across demographic and socioeconomic characteristics (Supplementary Table 1).
For spatial and spatiotemporal analyses, individual records were aggregated to neighborhood-level participation rates. Geocoded invitations were consolidated into biennial periods and aggregated in 2121 inhabited neighborhoods to ensure uniform geographic units over time using Swiss neighborhoods, developed by MicroGIS26. These neighborhoods are defined as homogeneous units corresponding to urban blocks. Importantly, these neighborhoods do not have a fixed spatial scale but are delineated based on population density and land-use patterns. We used biennial periods for consistency with participation trends, labeling each period by the even-numbered year (e.g., 2003 and 2004 as “2004”). This spatial aggregation enables detection of geographic clustering patterns and linkage with area-level covariates, but transforms the analytical unit from individual screening decisions to contextual participation patterns at the neighborhood level.
Individuals located outside populated neighborhoods (1394 out of 482,318 invitations, 0.29%) were assigned to the nearest inhabited neighborhood using the nearest_points function from the Shapely library in Python. These cases likely reflect temporal changes in population distribution, as the population data is contemporaneous to BCS invitations that span 2003–2020. Some individuals may have resided in areas that were populated during the earlier study periods but are no longer inhabited according to the more recent neighborhood data. Those at intersections of multiple neighborhoods were randomly assigned to prevent duplication.
For each neighborhood and biennial period, we calculated raw participation rates as the ratio between the number of participants who attended mammography screening (within the cantonal BCS program) and the number of invitations sent, multiplied by 100. We also computed mean values for other covariates within each spatial unit. To address unequal variances due to the disparate distribution of individuals, we used Spatial Empirical Bayes smoothing (SEBS) with 8 nearest neighbors26 implemented using the smoothing module from the Exploratory Spatial Data Analysis (ESDA, v2.7.0) package within Python Spatial Analysis Library (PySAL, v25.1)27. The 8-neighbor parameter provides conservative smoothing to stabilize rates in neighborhoods with limited invitations while minimizing spatial information borrowing.
Socioeconomic status
Socioeconomic status was determined using a neighborhood-level socioeconomic deprivation index, as individual-level SES data were not available through the BCS program. We used the same Swiss neighborhoods26, which represent the finest operational level for socioeconomic analysis. The procedure to create the deprivation index was based on the methodology from Lalloué et al.28, with adaptation to data available at the neighborhood scale in the canton of Geneva as described in our previous work18. The deprivation index was standardized to a 0–1 scale using min-max normalization and inverted such that higher values represent greater deprivation (1 = highest deprivation, 0 = lowest deprivation), ensuring consistent interpretation across all analyses.
Physical accessibility to cancer screening centers
Physical accessibility to cancer screening centers was assessed by computing the density of screening centres within a 5000-m radius for each individual, factoring in a linear decay function and street network distances29 using the Python package Pandana30. The 5-km radius was selected to capture local accessibility, primarily via walking, cycling, public transit and shorter-distance private vehicle trips in Geneva’s dense urban context, while the linear decay function ensures closer centers receive appropriate priority weighting. This approach acknowledges that women may access specialized healthcare services across substantial distances using available local transportation options. Screening center locations, operational periods, and program participation status were obtained from Geneva’s organized BCS program, with openings and closures incorporated to reflect temporal changes in accessibility within the organized screening network.
Ethical approval
The study was approved by the Cantonal Research Ethics Commission of Geneva, Switzerland (2023-02097).
Spatiotemporal analysesHot spot analysis – Getis-Ord Gi* statistic
We analyzed the spatial dependence of breast screening participation rates per neighborhood to identify geographic patterns among beneficiaries. Spatial dependence, rooted in Tobler’s first law of geography, suggests that nearby observations are more likely to exhibit similar attributes due to shared environmental, social, and economic conditions31.
Using the Getis-Ord Gi* statistic, we identified spatial clusters—or “hot spots” and “cold spots”—where local (i.e., within a specified spatial lag) screening rates were significantly higher or lower than the global average32. Our analytical approach focused directly on local cluster identification to address our applied public health objectives of identifying specific geographic areas with high and low participation for targeted intervention, as Getis-Ord Gi can identify statistically significant local clustering patterns even when global spatial autocorrelation may not be significant.
The spatial lag refers to the neighborhood definition used to calculate local statistics—specifically, the set of neighborhoods whose values are included in the local calculation for each focal neighborhood. For hot spot identification, we evaluated spatial patterns using K-nearest neighbors (KNN) weight matrices: K = 8 and K = 16 nearest neighborhoods for each spatial unit. The spatial relationship conceptualization was based on neighborhood centroids using Euclidean distance calculations. We applied binary weights where neighbors were assigned a value of 1 and non-neighbors 0, with no additional row standardization required due to equal neighbor counts across all spatial units.
We selected KNN weights over contiguity-based weights, as many units lack sufficient contiguous neighbors. We also chose KNN over distance-based weights to ensure each neighborhood has a consistent number of neighbors for spatial lag calculations, providing methodological consistency despite varying distances to neighbors across different density contexts.
Statistical significance was assessed using Monte Carlo permutation testing with 9999 permutations to empirically derive pseudo-p-values. Hot spots and cold spots were classified using multiple significance thresholds (p < 0.001, p < 0.01, p < 0.05, p < 0.1) to provide graduated significance levels, with higher statistical confidence represented by more intense colors in mapping visualizations. The spatial patterns were more pronounced with k = 16 neighbors, hence results from this neighborhood definition were retained for final analysis.
Space-time pattern mining–Emerging hot spot analysis
To discern shifts in the patterns of participation in the screening program, we utilized a space-time pattern mining (STPM) technique also known as emerging hot spot analysis33. This method is designed to spot evolving temporal trends in spatial data, including the emergence, intensification, reduction, and sporadic occurrence of hot and cold spots. This method is mainly available through the ArcGIS® pro software34, it can also be implemented in Python35.
This analysis first requires creating a 3D space-time cube composed of bins defined by space and time parameters. The bin size was set at the neighborhood scale, which enables the detection of very local urban patterns and allows a convenient spatial linkage with multiple sociodemographic descriptors (See section “Data – Socioeconomic status”). The biennial time parameter was chosen to align with the invitation schedule of the BCS program. The resultant 3D space-time cube was composed of a maximum of nine-time intervals and 2121 spatial bins (i.e., neighborhoods), yielding up to 19,089 unique space-time bins.
We utilized the Getis-Ord Gi* statistic to evaluate each bin within the space-time cube and employed a false discovery rate (FDR) correction during this process to account for multiple hypothesis testing. Like the Getis-Ord Gi* analysis, each bin was assigned a Z-score, a p-value, and a category. Overall trends were evaluated using the Mann-Kendall trend test, a non-parametric test used to identify statistically significant trends. Based on the p-values, z-scores, and Mann-Kendall statistics, each location was then categorized into a maximum of 17 predefined categories (1 no pattern, 8 cold spots, 8 hot spots) (Table 1).
Table 1 Classification of spatiotemporal patterns into 17 categories using space-time pattern mining (emerging hot spot analysis)
Pattern classification followed ESRI’s established emerging hot spot analysis methodology, including the standard 90% threshold parameter that determines temporal consistency requirements for certain categories. The terminology used in pattern categories follows ESRI’s official framework to ensure methodological consistency and reproducibility. We implemented one modification: the Historical pattern threshold was reduced from 90 to 50% of time steps to better capture meaningful historical patterns in our dataset context.
Some less populated neighborhoods did not appear in all nine biennial intervals, likely due to the intermittent presence of women within the target age range (Supplementary Fig. 2).
All maps were created using Python (matplotlib, geopandas) with base geographic layers obtained from the Geneva Territory Information System (SITG), including administrative boundaries of the canton of Geneva, municipalities, lake contours, and rivers36.
Statistical analyses
We employed the eXtreme Gradient Boosting (XGBoost) method, an established machine learning algorithm for regression applications, to assess the relative importance of selected area-level determinants in predicting neighborhood-level BCS participation rates. XGBoost (xgboost package, v3.0.5) was selected for its ability to capture complex non-linear relationships and interactions between predictors while maintaining interpretability through SHapley Additive exPlanations (SHAP) values37,38.
The analysis used neighborhood-level participation rates from the final biennial period (2019–2020) as the outcome variable. This contemporary cross-sectional approach focuses on identifying current determinants most relevant for intervention planning, given that our primary predictor variables are measured at a single time point or exhibit minimal temporal variation. In the 2019–2020 period, the analytical dataset comprised 1983 inhabited neighborhoods with screening invitation data.
Predictor variables included the neighborhood socioeconomic deprivation index, density of screening centers, age group composition of invited women (proportions in five-year categories: 50–54, 55–59, 60–64, 65–69, 70–74 years), and spatial coordinates to capture residual geographic effects not explained by measured covariates.
Data were randomly partitioned into training (80%) and test (20%) sets. Hyperparameters were optimized using Bayesian optimization with five-fold cross-validation to minimize root mean squared error using the hyperopt package (v0.2.7)39. The optimized model was trained on the training set with early stopping to prevent overfitting, then evaluated on the held-out test set to assess generalization performance.
Model interpretation employed SHAP values using the shap package (v0.48.0), which quantify each predictor’s contribution to participation rate predictions while accounting for interactions between variables39. SHAP values provide both variable importance rankings and directional effects, enabling identification of key determinants and their relationships with screening participation40.