Abstract
Background
Identifying communities disproportionately affected by hepatitis C infection is essential for targeted prevention and resource allocation. This study aimed to identify zip code tabulation areas at elevated risk of hepatitis C infection in Los Angeles County and to identify area-level variables of importance in estimating hepatitis C infection rates.
Methods
This was an ecological cross-sectional study integrating Elastic Net and Random Forest machine learning approaches. The outcome was reported first-time RNA-positive hepatitis C infection among persons aged 18–39 years residing in Los Angeles County from January 2020 to December 2023. 31 area-level variables representing socioeconomic conditions, healthcare access, and opioid-related morbidity were included as model features. Model-predicted hepatitis C infection rates were used as proxies for underlying community-level risk. Zip code tabulation areas with model-predicted hepatitis C rates above the 80th percentile were classified as very high risk. Area-level variables of importance were identified through Elastic Net feature selection, and importance scores and rankings were derived from the Random Forest model.
Results
Forty-nine of 245 zip code tabulation areas were classified as very high risk and were distributed across the county. Nine area-level variables of importance were identified. In descending order of importance, these included: proportions of families living below the federal poverty level, residents aged 18–39 years, and households without a vehicle; annual rates of opioid-related emergency department visits, newly diagnosed HIV infections, and opioid-related hospitalizations; distance to substance use treatment facilities; and proportions of Asian and male residents.
Conclusion
Integrating machine learning methods with public health surveillance data enabled the identification of communities at elevated risk of hepatitis C infection and area-level variables of importance in estimating hepatitis C infection rates. The findings could help identify priority areas that warrant further assessment for targeted hepatitis C elimination interventions, including harm-reduction programs, testing, and linkage-to-care efforts.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12889-026-28108-3.
Keywords: Hepatitis C, Spatial epidemiology, Social determinants of health, Machine learning, Health disparities, Public health surveillance
Introduction
Hepatitis C infection is the most prevalent blood-borne infection in the United States (U.S.), with approximately 2.4 million people living with chronic hepatitis C infection [1, 2]. It’s often asymptomatic, leading many people to remain undiagnosed. If untreated, chronic hepatitis C infection can result in severe liver complications, including cirrhosis, hepatocellular carcinoma, and increased risk of mortality. The epidemiology of hepatitis C infection in the U.S. has undergone a notable demographic shift, with a rising prevalence observed in individuals under 40 years, primarily due to unsafe injection drug use [3–5]. Approximately 70% of acute hepatitis C infections are attributed to unsafe injection drug use, underscoring its central role in ongoing transmission [3]. Additionally, transmission occurs within a broader risk environment shaped by socioeconomic conditions and healthcare access, which ultimately shapes the vulnerability of different areas to the spread of hepatitis C infection [6–9].
Despite the availability of highly effective direct-acting antiviral therapies for treating hepatitis C infection and preventing transmission, substantial gaps in diagnosis and linkage to care persist [5, 10, 11]. Approximately 40% of individuals with hepatitis C infection in the U.S. are unaware of their infection, and fewer than 30% have achieved viral clearance [10, 11]. In response to this public health challenge, the Centers for Disease Control and Prevention expanded its recommendations in 2020 to include universal one-time screening for all adults aged 18 years and older [2]. However, even with these efforts, the uptake of testing and treatment remains inadequate, leading to missed opportunities for early diagnosis, timely treatment, and prevention of ongoing transmission across communities, essential steps necessary to achieve the hepatitis C elimination targets established by the World Health Organization [6, 12, 13].
For instance, Los Angeles County, which has the highest estimated number of individuals diagnosed with chronic hepatitis C infection in California, approximately 100,000 people, many of whom remain unaware of their infection status, represents a setting of particular concern [14]. Moreover, only about one in three residents diagnosed with chronic hepatitis C have received treatment, underscoring persistent gaps in linkage to care and treatment completion [15]. The variability in access to harm reduction services across communities, coupled with structural barriers such as socioeconomic disadvantage and limited healthcare access, further complicates the landscape [16]. These ongoing unmet needs highlight the importance of identifying communities where the burden of hepatitis C may be disproportionately high and where targeted assessment and resource prioritization may be most beneficial.
Area-level risk modeling has increasingly been used to identify communities with higher relative risk for the spread of hepatitis C infection due to unsafe injection drug use. This is often determined by estimating the area-level rate of acute or newly reported hepatitis C infections among individuals under 40 years as a proxy for the elevated area-level risk for the spread of hepatitis C infection [17–21]. For instance, in 2016, the Centers for Disease Control and Prevention identified counties across the U.S. at risk of outbreaks of hepatitis C or HIV due to injection drug use [17]. Similarly, a study conducted in California in 2017 revealed that two-thirds of counties were at risk for the spread of hepatitis C or HIV infection due to unsafe injection drug use, with Los Angeles County being one of the high-risk counties [14].
However, recent research assessing areas at risk for the spread of hepatitis C infection has predominantly used county-level geographic units of analysis [14, 17, 19, 20]. A larger geographic scale obscures disparities in risk within bigger, socioeconomically heterogeneous jurisdictions. Using a smaller geographic scale, such as zip code tabulation areas, could provide a more precise understanding of disparities across communities within counties. Furthermore, existing approaches estimating area-level risk primarily relied on linear statistical algorithms, which insufficiently account for the complex interactions and nonlinear nature of risk-environment indicators that shape variation in area-level hepatitis C infection rates [22, 23]. More nuanced methodologies are necessary to explore area-level risk for the spread of hepatitis C infection within local health jurisdictions.
This study sought to address these gaps by leveraging machine learning techniques to estimate the area-level rate of hepatitis C infection and identify communities at elevated risk of hepatitis C infection at smaller geographic units. Prior research has demonstrated the utility of machine learning approaches in generating area-level model-estimated hepatitis C infection rate, and identifying areas at elevated risk of hepatitis C infection transmission that could be targeted for implementation of testing, linkage to care, and localized syringe services programs [23, 24].
Therefore, this analysis used machine learning approaches to estimate the zip code tabulation area-level hepatitis C infection rate among residents aged 18–39 years and to identify areas at elevated risk of hepatitis C virus infection in Los Angeles County. Zip code tabulation areas at elevated risk for the spread of hepatitis C infection were defined as those with model-predicted RNA-positive hepatitis C infection rates above the 80th percentile. The study also highlighted the model-selected area-level factors of importance in estimating hepatitis C infection rates.
Methodology
Study design, study area, study population
This analysis employed an ecological cross-sectional study design, using data from Los Angeles (LA) County (Fig. 1). LA County has the highest estimated number of individuals living with chronic hepatitis C infection in California, with around 100,000 people affected, many of whom are unaware of their infection status [14]. Furthermore, LA County is the most populous county in the U.S., home to more than 10 million residents and covering roughly 4,300 square miles. It contains 88 incorporated cities and is characterized by substantial socioeconomic and demographic heterogeneity, with significant variation in income, racial/ethnic composition, and healthcare access across communities. This makes it an area of interest for exploring communities at elevated risk of hepatitis C infection that could benefit from hepatitis C microelimination public health interventions.
Fig. 1.
Overview map of study area (LA County): Highlighting zip code tabulation area boundaries, incorporated and unincorporated areas, and overlayed with service planning areas (SPA). SPAs are eight geographic regions created by the LA County Department of Public Health to organize health service delivery. Incorporated areas: cities with their own local governance structures; unincorporated areas: regions administered directly by Los Angeles County
The geographic unit of analysis in this study was the zip code tabulation area for LA County, based on the 2020 Census TIGER/Line Shapefiles published by the United States Census Bureau. This unit of analysis enabled the identification of communities at elevated risk of hepatitis C infection at a granular geographic level, while ensuring the confidentiality of sensitive protected health information.
Data and sources
Outcome variable
The outcome of interest was the number of first-time reported RNA-positive hepatitis C infections among LA County residents aged 18 to 39, reported to the LA County Department of Public Health between January 2020 and December 2023, aggregated at the zip code tabulation area level. Individual cases were assigned to a zip code tabulation area based on the latitude and longitude of residential addresses. Only cases among non-institutionalized persons with a residential address within LA County were included in the analysis. The LA County Department of Public Health does not receive reports for residents of Long Beach and Pasadena; therefore, data from these areas were not included in the analysis.
The reported cases of first-time RNA-positive hepatitis C infection aged 18 to 39 were used as a proxy indicator for unsafe injection drug use patterns and the risk for the spread of hepatitis C infection. Prior models in recent studies utilized reported positive hepatitis C cases as a proxy for unsafe injection drug use, mainly because of the absence of nationwide, statewide, or county surveillance data to represent injection drug use risk behaviors [17, 20, 23, 24]. Although not all recently reported hepatitis C cases aged below 40 years are due to unsafe injection drug use, unsafe injection drug use accounts for approximately 75% of new hepatitis C infections among people aged 18–40 years and is a driving force of the epidemic in the U.S [3, 4, 17, 25]. Additionally, a model-based study by Trickey et al. (2019) [26] estimated that 77% of hepatitis C infections would be prevented if additional hepatitis C transmission due to injection drug use was removed between 2018 and 2030 in the U.S.
Independent variables
Potential area-level variables included in this study were based on a previous systematic review that examined area-level predictors of hepatitis C infection rate [27], contingent on data availability. The model included 31 variables aggregated at the zip code tabulation area level, as summarized below. Supplementary Table 1 provides a detailed description of all variables and their respective data sources.
To represent the socioeconomic and physical environment, we included variables related to the ZCTA-level age composition, ethnicity composition, education level, poverty rate, median household income, vehicle ownership, insurance coverage, housing, and population density. Those variables were sourced from the 5-year estimates of the American Community Survey (2018–2022).
To represent the healthcare service environment, we incorporated indicators related to health outcomes that co-occur with hepatitis C infection, such as ZCTA-level HIV prevalence and reported annual rates of newly diagnosed HIV infections, as well as opioid-related outcomes, such as ZCTA-level annual hospitalization and emergency department visit rates. Data on Opioid-related and HIV-related metrics were obtained from the California Department of Public Health, Substance and Addiction Prevention Branch-Overdose Surveillance database [28] and the AIDSVu data portal [29], respectively, for the year 2022. Additionally, variables related to geographic distance (in miles) to primary healthcare facilities, substance use treatment centers, primary care providers, and hepatitis C treatment specialists were estimated and included in the analysis. The geographic distance from the population-weighted centroid of each zip code tabulation area to the nearest healthcare services of interest was estimated and subsequently used as a variable in the model. Distance estimation was calculated using the Network Analyst tool in ArcGIS Pro version 3.4. Data on the location of healthcare facilities and providers were obtained from the California Department of Health Care Access and Information database, the National Plan and Provider Enumeration System [30], and the Medi-Cal Managed Care Provider Listing dataset for 2022 [31].
Analytical approach
Figure 2 illustrates the analytical workflow from data preparation and model development to Geovisualization in stepwise detail.
Fig. 2.
Analytical workflow summarizing steps followed to identify areas at elevated risk of hepatitis C infection in Los Angeles County
Model development
Hybrid elastic net and random forest model
A novel hybrid machine learning model was developed to predict first-time RNA-positive hepatitis C infections at a zip code tabulation area level by combining Elastic Net and Random Forest techniques using 31 predictor variables (see Supplementary Table 1). A hybrid Elastic Net and Random Forest modeling approach was used because it combines the complementary strengths of penalized regression and ensemble tree-based learning. A Poisson log-likelihood function was incorporated into the hybrid model to accurately model the outcome variable, which followed a Poisson distribution. Additionally, an offset, defined as the log of the total number of residents aged 18–39 years in each zip code tabulation area, was included in the model, based on 2018–2022 American Community Survey 5-year estimates.
In the hybrid model, the Elastic Net model was used to select the most important features while addressing multicollinearity among predictor variables [32]. The model shrinks coefficients toward zero; those that approach zero indicate a lack of substantial predictive value and are excluded, a process known as regularization [32]. The features selected by the Elastic net were then used in the Random Forest model. Random forest uses a combination of decision trees to produce more accurate and stable predictions [22, 33]. It is a highly flexible machine learning approach that can handle nonlinear and nonparametric variables [34]. Additionally, the model accounts for higher-order interactions among variables [33].
The five-fold nested cross-validation technique was used in the hybrid Elastic Net and Random Forest model to reduce the risk of overfitting and accurately estimate the coefficients of the selected variables. This process involved randomly partitioning the data into five subsets, training the model on four subsets, and testing it on the remaining subset. This procedure was repeated five times to thoroughly evaluate the model’s performance.
Model performance
Model performance was assessed using 5-fold cross-validation based on root mean squared error (RMSE), which quantifies the average difference between observed and predicted values, lower values indicate better predictive performance. The performance of the hybrid Elastic Net and Random Forest model was compared with five alternative models: a traditional Poisson model, standard Elastic Net and LASSO Poisson models, a standard Random Forest model, and a hybrid LASSO and Random Forest model.
The hybrid Elastic Net and Random Forest model performed similarly to the best-performing model (standard Random Forest) (Supplementary Tables 2 and Supplementary Fig. 1). Despite this similarity in performance, the hybrid model was selected to balance predictive accuracy with interpretability. Specifically, the Elastic Net component enabled the selection of a parsimonious subset of predictors, reducing the total number of variables from 31 to 9, while the Random Forest component captured nonlinear relationships using this reduced feature set. This approach resulted in a more interpretable and stable model with minimal loss in predictive performance compared to the standard Random Forest model.
Variables of importance
The importance rankings of the parsimonious, more interpretable subset of area-level variables selected by the Elastic Net model were evaluated using the Random Forest model (Supp. Table 3). The results are shown as ranked importance scores for each selected variable. The importance scores reflect the mean decrease in model accuracy when a predictor’s values are shuffled, indicating how the predictor affects the model’s predictive performance [34]. Furthermore, the standardized coefficients from the Elastic Net model for the selected variables were included for exploratory purposes and transparency.
Model-estimated rate of hepatitis C infection and risk categories
Finally, the hybrid elastic net and random forest model was used to predict first-time hepatitis C infections per 100,000 population among persons aged 18–39 years for each zip code tabulation area. The predicted estimates for each zip code tabulation area served as a proxy elevated risk of hepatitis C infection due to unsafe injection drug use. The predicted estimates are deemed more robust than the raw rates because they provide a more stable and precise representation of the underlying risk [35]. By incorporating a parsimonious set of variables that capture the structural and contextual risk markers of hepatitis C infection rate, the model predictions smooth out noise, substantially reducing estimation error. This is an essential advantage in geographic areas with small populations or sparse case counts, where crude rates could be highly unstable.
Furthermore, given the spatial nature of this analysis, spatial autocorrelation in model residuals of the outcome variable (hepatitis C infection rate) was assessed using Global Moran’s I statistic to investigate whether there was significant clustering in the outcome variable (Moran’s I = 0.183,p = 0.832. Upon investigation, the results showed no statistically significant spatial autocorrelation in the residuals, indicating no spatial clustering and no need to account for spatial autocorrelation in the model.
To facilitate interpretation, predicted hepatitis C infection rates were classified into five categories based on their percentile distribution: very high (≥ 80th percentile), high (60th–<80th percentile), moderate (40th–<60th percentile), low (20th–<40th percentile), and very low (< 20th percentile). This quantile-based approach was used to enable relative comparison and ranking of ZCTAs within the study area in the absence of established epidemiologic thresholds, and to ensure balanced group sizes across categories [24, 36]. Accordingly, the “very high risk” category represents areas at the upper end of the model-estimated hepatitis C infection rate distribution rather than an absolute epidemiological threshold.
Furthermore, the agreement between risk categories derived from the model-predicted hepatitis C infection rate and those from the observed hepatitis C infection rate across ZCTAs was evaluated using a confusion matrix, and results were reported as overall accuracy, Cohen’s kappa, and class-specific performance metrics (sensitivity, specificity, and balanced accuracy).
Geovisualization of percentiles of the predicted rate of hepatitis C infection
The model results were visualized by mapping the percentiles (risk categories) of the model-predicted hepatitis C infection rate to highlight the spatial distribution of high- and low-risk ZCTAs in LA County. Further, summary characteristics of variables of importance for ZCTAs categorized as very-high and very-low risk were estimated. The summary statistics were reported as mean (standard deviation) and median (interquartile range).
Software
Machine learning models were fitted using the Caret glmnet and Caret packages in R version 4.3.2. Geovisualization was done using ArcGIS Pro version 3.4.
Ethical approval
The study was approved by the Los Angeles County Department of Public Health Institutional Review Board (IRB Project No. 2024-07-002) and the University of Southern California IRB (Study ID: UP-24-01157).
Results
The analysis included 3,169 of 5585 reported first-time RNA-positive hepatitis C cases (January 2020 to December 2023) from 245 zip code tabulation areas among non-institutionalized LA County residents. The overall ZCTA-level mean and median of reported first-time RNA-positive hepatitis C infection rates among persons aged 18–39 were 156 cases per 100,000 and 82 cases per 100,000, respectively. Supplementary Table 1 provides detailed summary statistics for all variables included in this analysis.
Model performance and area-level variables of importance
The hybrid Elastic Net and Random Forest model achieved a root mean squared error (RMSE) of 11, indicating that predicted values deviated from observed hepatitis C cases by an average of approximately 11 cases (Table 1, Supplementary Fig. 1). The model explained 34% of the variance in hepatitis C infection rates (Table 1). Predictive performance was comparable to that of the best-performing model (standard Random Forest). Given this similarity, the hybrid model was selected to provide a more parsimonious and interpretable set of variables of importance while maintaining predictive accuracy.
Table 1.
Summary of model performance metrics
| Model | RMSE | MAE | Pseudo R² | Pearson Correlation |
|---|---|---|---|---|
| Random Forest | 11.1 | 6.4 | 0.39 | 0.63 |
| Hybrid Elastic Net + Random Forest | 11.4 | 6.9 | 0.34 | 0.58 |
| Hybrid LASSO + Random Forest | 11.6 | 7.0 | 0.32 | 0.56 |
RMSE Root mean squared error, MAE mean absolute error
Of the 31 area-level variables, the hybrid elastic net and random forest model identified 9 variables of importance in estimating the first-time hepatitis C infection rate among persons aged 18–39 years. Figure 3 summarizes the variables of importance, arranged in descending order by their importance scores, along with their respective estimated effect sizes (included for exploratory purposes and transparency). The selected variables of importance included: the proportions of families with incomes below the federal poverty level, residents aged 18–39 years, and households without a vehicle; the annual rates of opioid-related emergency department visits, newly diagnosed HIV infections, and opioid-related hospitalizations; the distance to a substance use treatment facility; and the proportions of Asian and male residents. Notably, variables with high Elastic Net effect sizes did not correspond to those with high Random Forest importance scores, indicating that the two methods capture different aspects of variable importance. Elastic Net quantifies linear effects after regularization, whereas Random Forest identifies variables that improve prediction through nonlinear relationships and interactions, hence the difference.
Fig. 3.
Variables of Importance in predicting first-time RNA-positive hepatitis C infection rate among individuals aged 18–19 at a zip code tabulation area level. Note: Variables were selected using the Elastic Net model, and their importance scores were derived from the Random Forest model
Geovisualization of percentiles of the predicted rate of hepatitis C infection
The hybrid Elastic Net and Random Forest model was used to predict the first-time RNA-positive hepatitis C infections per 100,000 population among persons aged 18–39 years for each zip code tabulation area. The model’s predicted hepatitis C infection rates were classified into risk categories based on the percentile distribution of the model-predicted rates to highlight areas at elevated risk of hepatitis C infection. Figure 4 displays the spatial distribution of the risk categories across zip code tabulation areas: very high risk (≥ 80th percentile), high (60th–<80th percentile), moderate (40th–<60th percentile), low (20th–<40th percentile), and very low (< 20th percentile). ZCTAs identified as very high- and high-risk were distributed across the county, with notably high-risk ZCTAs located in the Metro and South service planning areas of Los Angeles County (Fig. 4).
Fig. 4.
Risk areas for hepatitis C infection based on the percentile distribution of predicted first-time RNA-positive hepatitis C infections per 100,000 persons aged 18–39 across zip codes in LA County. The modeled outcome was the reported first-time RNA-positive hepatitis C infections among residents aged 18–39 (2020–2023). ZIP codes in Long Beach and Pasadena are outside the LA County Department of Public Health’s jurisdiction; hence, the data is missing
Furthermore, risk categories from model-estimated hepatitis C infection rates across ZCTAs were compared with those derived from observed hepatitis C infection rates. The agreement between risk categories derived from model-estimated and observed hepatitis C infection rates was modest but significantly better than chance (accuracy = 34.6%, 95% CI: 28.6–40.9; κ = 0.18; p < 0.001) (Supp Table 4). Sensitivity for identifying very high-risk ZCTAs was 68%, while specificity remained high across all risk categories (> 80%), indicating limited model over-classification (Supp Table 4). Ranking the ZCTAs by the model-estimated hepatitis C infection rates and visualizing the top and bottom 20 ZCTAs showed no overlap in the confidence intervals between these groups, despite the top 20 ZCTAs having relatively wider confidence intervals (Supp Fig. 2).
Table 2 summarizes the characteristics of the model-selected variables of importance across very high-risk and very low-risk zip code tabulation areas. Notably, zip codes categorized as very high risk for the spread of hepatitis Cinfection had a higher proportion of families with incomes below the federal poverty level, higher rates of opioid-related emergency department visits and hospitalizations, and higher rates of newly reported HIV diagnoses and HIV prevalence than zip code tabulation areas categorized as very low risk.
Table 2.
Summary characteristics of variables of importance across zip code tabulation areas categorized as very-high and very-low risk for the spread of hepatitis C infection, Los Angeles County
| Variables of importance | Risk categories based on percentiles of the predicted rate of hepatitis C infection across zip code tabulation areas (N = 245) | |||
|---|---|---|---|---|
| Very low (predicted hepatitis C rate below the 20th percentile) (N = 49) | Very High (predicted hepatitis C rate above the 80th percentile) (N = 49) | |||
| Mean (SD) | Median (Q1,Q3) | Mean (SD) | Median (Q1,Q3) | |
| Annual rate of reported hepatitis C cases/ 100,000 * | 55 (9) | 57(51,62) | 477 (581) | 276 (209,454) |
| % aged 18–39 years | 31 (8) | 29 (27,33) | 35 (10) | 34 (29,40) |
| % of Male residents | 49 (2) | 49 (48,50) | 50 (4) | 50 (48,51) |
| % of families with income below the poverty level | 9 (4) | 9 (6,12) | 16 (10) | 12 (9,19) |
| Opioid-related Emergency Department visits/100,000 | 31 (19) | 26 (18,42) | 94 (135) | 66 (31,112) |
| Newly reported HIV diagnoses/100,000 | 57 (33) | 45 (33,76) | 155 (168) | 81 (60,187) |
| % of households without a vehicle | 5 (3) | 4 (4,6) | 14 (13) | 9 (5,20) |
| % of Asians | 25 (19) | 18 (11,35) | 12 (11) | 9 (5,16) |
| HIV prevalence, cases/100,000 | 316 (243) | 252 (180,393) | 1024 (1219) | 569 (271,1243) |
| Distance to SUD treatment facility (miles) | 1.65(1.52) | 0.98 (0.47,2.65) | 1.27(2.51) | 0.61 (0.32,0.95) |
| Opioid-related Hospitalization/100,000 | 5 (4) | 4 (3,7) | 22 (42) | 10 (3,14) |
*Based on model-predicted first-time RNA-positive hepatitis C infection cases among people aged 18–39 (2020–2023).; N: number of Zip code tabulation areas; SD: standard deviation; Q1: first quartile; Q3: Third quartile
Discussion
In this ecological study, we used a hybrid Elastic Net-Random Forest model to estimate zip code tabulation area-level hepatitis C infection rates among residents aged 18–39 years and to identify areas at elevated risk of hepatitis C virus infection in Los Angeles County. Forty-nine of 245 ZCTAs were classified as very high risk, with these areas distributed across the county. Nine area-level variables of importance were selected by the model. In descending order of importance, these included the proportions of families living below the federal poverty level, the proportion of residents aged 18–39 years, and households without a vehicle; annual rates of opioid-related emergency department visits, newly diagnosed HIV infections, and opioid-related hospitalizations; and the proportions of Asian and male residents. Together, these findings demonstrate geographic heterogeneity in area-level risk of hepatitis C infection and highlight important area-level variables associated with the area-level hepatitis C infection rate.
The findings revealed that high-risk zip code tabulation areas were dispersed throughout Los Angeles County, with most located in the Metro and South service planning areas. These patterns may reflect varying underlying contextual conditions, including gaps in access to prevention and treatment services, as well as a higher prevalence of behaviors associated with hepatitis C transmission at the population level. Given the rising prevalence of acute hepatitis C among individuals aged 18–39 years, particularly in the context of injection drug use [3, 4], integrated, community-level approaches, including expanded testing and improved linkage to treatment, remain important for addressing hepatitis C infection in high-risk areas.
The model-selected area-level variables of importance included socioeconomic factors, such as the percentage of families with incomes below the federal poverty level and the number of households without vehicles. These findings are consistent with prior studies highlighting the role of the socioeconomic environment in shaping the geographic distribution of hepatitis C burden [27, 37]. These results suggest that communities characterized by lower socioeconomic status may warrant prioritization for hepatitis C prevention interventions, including testing, linkage to care, and harm reduction services [38].
Furthermore, the analysis revealed that area-level prevalence of HIV, annual rates of newly diagnosed HIV infections, opioid-related emergency department visits, and opioid-related hospitalizations were among the model-selected variables of importance. These findings underscore the interconnected relationship among the epidemics of HIV, the opioid crisis, and hepatitis C infection [6, 39]. Prior research suggests that areas experiencing overlapping burdens of HIV, opioid-related harms, and socioeconomic disadvantage may also exhibit higher levels of hepatitis C infection [6]. In this context, integrated, community-level approaches that address the co-occurrence of hepatitis C infection, HIV, and opioid-related harm may be important for comprehensive prevention efforts in high-risk areas.
The hybrid Elastic Net and Random Forest model demonstrated moderate predictive performance, explaining a portion of the variability in hepatitis C infection rates (pseudo R² = 0.34) and the agreement between predicted and observed hepatitis C risk categories was also modest. Despite this, the model’s ability to identify ZCTAs at the extremes of the hepatitis C infection rate distribution suggests it may be useful for flagging priority areas for further assessment. Accordingly, the resulting risk map could be most appropriately used to support population-level planning, prioritization, and hypothesis generation, rather than to guide fine-scale resource allocation decisions in isolation. Methodologically, combining Random Forest—a robust, non-linear predictive algorithm—with Elastic Net for feature selection allowed for the development of a parsimonious and interpretable model that prioritizes the most informative predictors. However, future work incorporating additional data sources and a formal assessment of predictive calibration would improve model performance and strengthen its application for public health decision-making.
Public health Implications of study findings
The findings of this study have potential implications for supporting hepatitis C microelimination efforts, particularly in communities with relatively elevated risk of hepatitis C infection. While the model results cannot replace local surveillance and monitoring systems, they could serve as a starting point for identifying priority areas for further assessment. Local health departments could use this information to guide enhanced surveillance, community needs assessments, and validation with additional data sources such as testing patterns and healthcare service utilization. Priority areas may then be considered for context-specific public health responses, such as expanded testing, linkage-to-care initiatives, and needle exchange programs. Recent studies in other settings highlighted how identifying high-risk areas could help guide targeted hepatitis C linkage to care and prevention programs [23, 24]. However, such public health responses are most effective when integrated with the local context and stakeholder input. The model-selected variables of importance provide additional insight into the role of contextual factors in shaping the burden of hepatitis C infection across communities. Contextual factors such as socioeconomic conditions, newly diagnosed HIV infections rate, and opioid-related emergency department visits and hospitalization rates were identified as variables of importance in the model. These findings could inform context-specific integrated care initiatives in priority areas such as the implementation of programs that integrate hepatitis C care with harm reduction and HIV care services to effectively address the hepatitis C burden within communities [40, 41].
Limitations
This analysis has some limitations. First, the model relied on reported cases of first-time RNA-positive hepatitis C infections among individuals aged 18–39 years, aggregated at the ZCTA level, as a proxy for unsafe injection drug use patterns and, hence, as an indicator of area-level risk of hepatitis C infection. While injection drug use is a major driver of hepatitis C transmission in the study setting, not all infections are attributable to this pathway. Sexual transmission of hepatitis C is not explicitly captured in this model. Future models could incorporate area-level proxy variables that better capture the contribution of alternative transmission pathways. Second, the classification of ZCTAs into risk categories was based on the distribution of model-predicted hepatitis C infection rates and reflects a relative ranking rather than an absolute measure of risk. As such, the designation of “very high risk” (≥ 80th percentile) does not correspond to a specific epidemiologic threshold, and alternative cutoffs could yield different categorizations of ZCTAs [24, 36, 42], although the overall spatial patterns and identified higher-risk areas are unlikely to change substantially. Therefore, these categories should be interpreted as indicative of relative differences across ZCTAs and used to support comparative prioritization rather than as definitive criteria for public health decision-making. While not performed in this study, sensitivity analyses using alternative classification thresholds could provide additional insight into the robustness of these categorizations and represent an area for future research. Third, for simplicity and ease of interpretation, ZCTAs were treated as independent units, and the model did not account for human mobility, which may influence patterns of hepatitis C transmission [43]. Furthermore, although the spatial autocorrelation results for the model residuals revealed no statistically significant spatial clustering, this does not preclude localized spatial patterns or interactions among neighboring ZCTAs. As such, caution is warranted when interpreting the data and designing interventions to target high-risk zip code tabulation areas, as surrounding zip codes may also require consideration. Future studies could extend this work by incorporating spatial dependence structures and measures of human mobility to enhance model performance. Fourth, using ZCTAs instead of smaller geographic units could have reduced spatial precision because ZCTAs are larger, less demographically homogeneous units that can mask significant neighborhood-level socioeconomic and health disparities [44, 45]. However, given the small number of reported RNA-positive hepatitis C infections, analyzing our data at the zip code level enabled analysis at a relatively granular spatial scale while safeguarding the confidentiality of sensitive protected health information. Fifth, this was an ecological cross-sectional study; therefore, the variables of importance and their corresponding effect sizes should be interpreted with caution and should not be used to infer causation. However, the parsimonious subset of model-selected variables of importance provides an opportunity to generate future hypotheses focused on the role of contextual factors in shaping area-level risk of hepatitis C infection. Sixth, the analysis relied on reported cases of hepatitis C infection, which are subject to under-ascertainment. In addition to asymptomatic infections going undetected, the study period (2020–2023) coincided with the COVID-19 pandemic, during which disruptions to healthcare utilization, screening, and diagnostic services likely reduced hepatitis C testing. As a result, the reported first-time RNA-positive hepatitis C cases may underestimate the true burden of infection, and this under-ascertainment may have varied across ZCTAs, potentially introducing spatial bias in the model estimates. Therefore, it is essential to consider high-risk areas within the context of local testing service utilization. Furthermore, the lack of data from Pasadena and Long Beach may have resulted in a missed opportunity to identify high-risk ZCTAs in those cities, potentially leading to an underestimation of risk, particularly in adjacent areas. Future research could benefit from incorporating data from these areas to create a comprehensive risk map for the entire County. Finally, this model was developed and evaluated within Los Angeles County, and its generalizability to other jurisdictions is uncertain. Differences in population characteristics, healthcare infrastructure, and surveillance systems may affect model performance and the relevance of the model-selected variables of importance in other local health jurisdictions. Nonetheless, this analysis provides an approach for identifying high-risk areas for hepatitis C infection within a local health jurisdiction that could be prioritized for further assessment to inform targeted Hepatitis C elimination interventions.
Conclusions
This study demonstrated the application of machine learning approaches to estimate area-level hepatitis C infection rates among individuals aged 18–39 years and to identify ZCTAs with relatively elevated risk of hepatitis C infection in Los Angeles County. Very high-risk areas were distributed throughout the county, with greater concentration in the Metro and South service planning areas. Key model-selected area-level variables of importance included rate of opioid-related hospitalizations and emergency department visits, newly diagnosed HIV infections, and socioeconomic indicators. While the model results cannot replace local surveillance and monitoring efforts, they may help identify priority areas that warrant further assessment for targeted hepatitis C elimination interventions, including harm-reduction programs, testing initiatives, and linkage-to-care efforts. These findings are best interpreted as supporting community-level prioritization, particularly in the context of addressing hepatitis C and related syndemics in vulnerable communities. Future research could refine and extend this approach to incorporate spatiotemporal dynamics and improve the characterization of the area-level risk of hepatitis C infection.
Supplementary Information
Acknowledgements
We gratefully acknowledge the Hepatitis Division, Acute Communicable Diseases Control at the Los Angeles Department of Public Health, who generously provided the data that made this research possible.
Abbreviations
- IDU
Injection Drug Use
- LA
Los Angeles County
- LADPH
Los Angeles County Department of Public Health
- SPA
Service Planning Areas
- ACS
American Community Survey
- LASSO
Least Absolute Shrinkage and Selection Operator
- ZCTA
Zip Code Tabulation Area
Authors’ contributions
J.M.G: conceptualization, methodology, formal analysis, visualization, and writing – original draft. J.D.K., and D.R.: supervision, writing, review, and editing.
Funding
Not Applicable.
Data availability
The Hepatitis C case data used in this study are available from the Los Angeles County Department of Public Health, but restrictions apply to their availability. Other datasets supporting the findings are publicly available on the U.S. American Community Survey data portal and other sources listed in the methods section. A partial dataset and R code used in this study can be found on GitHub via this link: [ https://github.com/Gizamba/Machine-learning-and-community-risk-for-spread-of-HCV-infection ]
Declarations
Ethics approval and consent to participate
This study was approved by the Los Angeles County Department of Public Health Institutional Review Board (IRB Project No. 2024-07-002) and the University of Southern California Institutional Review Board (Study ID: UP-24-01157). The requirement for informed consent was waived by both Institutional Review Boards because the study used secondary, de-identified surveillance and area-level data. All procedures were conducted in accordance with relevant institutional guidelines and regulations.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Hofmeister MG, Rosenthal EM, Barker LK, Rosenberg ES, Barranco MA, Hall EW, et al. Estimating prevalence of hepatitis C virus infection in the United States, 2013-2016. Hepatology. 2019;69(3):1020–31. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Schillie S, Wester C, Osborne M, Wesolowski L, Ryerson AB. CDC Recommendations for Hepatitis C Screening Among Adults - United States, 2020. MMWR Recomm Rep. 2020;69(2):1–17. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Ryerson AB, Schillie S, Barker LK, Kupronis BA, Wester C. Vital signs: newly reported acute and chronic hepatitis C cases―United States, 2009–2018. Morb Mortal Wkly Rep. 2020;69(14):399. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Zibbell JE, Asher AK, Patel RC, Kupronis B, Iqbal K, Ward JW, et al. Increases in acute hepatitis C virus infection related to a growing opioid epidemic and associated injection drug use, United States, 2004 to 2014. Am J Public Health. 2018;108(2):175–81. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Bhattacharya D, Aronsohn A, Price J, Lo Re V. The American Association for the Study of Liver Diseases–Infectious Diseases Society of America HCV Guidance Panel , Hepatitis C Guidance 2023 Update: american association for the study of liver diseases– infectious diseases society of america recommendations for testing, managing, and treating hepatitis C virus infection. Clin Infect Dis. 2023;ciad319. 10.1093/cid/ciad319.
- 6.Perlman DC, Jordan AE. The Syndemic of Opioid Misuse, Overdose, HCV, and HIV: Structural-Level Causes and Interventions. Curr HIV/AIDS Rep. 2018;15(2):96–112. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Singer M, Bulled N, Ostrach B, Mendenhall E. Syndemics and the biosocial conception of health. Lancet. 2017;389(10072):941–50. [DOI] [PubMed] [Google Scholar]
- 8.Rhodes T. Risk environments and drug harms: a social science for harm reduction approach. Int J Drug Policy. 2009;20(3):193–201. [DOI] [PubMed] [Google Scholar]
- 9.Kolak MA, Chen YT, Joyce S, Ellis K, Defever K, McLuckie C et al. Rural risk environments, opioid-related overdose, and infectious diseases: A multidimensional, spatial perspective. Int J Drug Policy. 2020;85:102727. 10.1016/j.drugpo.2020.102727. [DOI] [PMC free article] [PubMed]
- 10.Clarisse A, Tsang J, Tonzel H, Symum HW, Kaufman WA, Meyer III, Osinubi A et al. State-Specific Hepatitis C Virus Clearance Cascades — United States, 2013–22. 2024. 10.15585/mmwr.mm7321a4. [DOI] [PMC free article] [PubMed]
- 11.Gnanapandithan K, Ghali MP. Self-awareness of hepatitis C infection in the United States: A cross-sectional study based on the National Health Nutrition and Examination Survey. PLoS ONE. 2023;18(10):e0293315. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.The White House. National HIV/AIDS Strategy for the United States 2022–2025. Office of National AIDS Policy. 2021. https://files.hiv.gov/s3fs-public/NHAS-2022-2025.pdf.
- 13.US Department of Health and Human Services. Viral hepatitis national strategic plan for the United States: a roadmap to elimination (2021–2025). Washington, DC, USA: US Department of Health and Human Services; 2020 https://www.hhs.gov/sites/default/files/Viral-Hepatitis-National-Strategic-Plan-2021-2025.pdf.
- 14.California Department of Public Health. California Fatal Opioid Overdose and HIV or Hepatitis C Virus (HCV) Vulnerability Assessment. California Department of Public Health; 2018. https://www.cdph.ca.gov/Programs/CID/DCDC/CDPH%20Document%20Library/VA_PublicReport.pdf.
- 15.Hernandez-Tamayo C, Stafylis C, Klausner JD. Lack of Hepatitis C Virus Elimination by 2030 in Los Angeles County at Current Treatment Rate. Open Forum Infect Dis. 2023;10(4):ofad125. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Acevedo A, Panas L, Garnick D, Acevedo-Garcia D, Miles J, Ritter G, et al. Disparities in the treatment of substance use disorders: does where you live matter? J Behav Health Serv Res. 2018;45(4):533–49. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Van Handel MM, Rose CE, Hallisey EJ, Kolling JL, Zibbell JE, Lewis B, et al. County-Level Vulnerability Assessment for Rapid Dissemination of HIV or HCV Infections Among Persons Who Inject Drugs, United States. J Acquir Immune Defic Syndr. 2016;73(3):323–31. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Shrestha S, Bayly R, Pustz J, Sawyer J, Van Handel M, Lingwall C, et al. Methods for jurisdictional vulnerability assessment of opioid-related outcomes. Prev Med. 2023;170:107490. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Sharareh N, Hess R, White S, Dunn A, Singer PM, Cochran J. A vulnerability assessment for the HCV infections associated with injection drug use. Prev Med. 2020;134:106040. [DOI] [PubMed] [Google Scholar]
- 20.Rickles M, Rebeiro PF, Sizemore L, Juarez P, Mutter M, Wester C, et al. Tennessee’s In-state Vulnerability Assessment for a Rapid Dissemination of Human Immunodeficiency Virus or Hepatitis C Virus Infection Event Utilizing Data About the Opioid Epidemic. Clin Infect Dis. 2018;66(11):1722–32. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Bergo CJ, Epstein JR, Hoferka S, Kolak MA, Pho MT. A vulnerability assessment for a future HIV outbreak associated with injection drug use in Illinois, 2017–2018. Front Sociol. 2021;6:652672. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Bi Q, Goodman KE, Kaminsky J, Lessler J. What is machine learning? A primer for the epidemiologist. Am J Epidemiol. 2019;188(12):2222–39. [DOI] [PubMed] [Google Scholar]
- 23.Yedinak JL, Li Y, Krieger MS, Howe K, Ndoye CD, Lee H et al. Machine learning takes a village: Assessing neighbourhood-level vulnerability for an overdose and infectious disease outbreak. Int J Drug Policy. 2021;96:103395. 10.1016/j.drugpo.2021.103395. [DOI] [PMC free article] [PubMed]
- 24.Bartholomew TS, Tookes HE, Spencer EC, Feaster DJ. Application of machine learning algorithms for localized syringe services program policy implementation - Florida, 2017. Ann Med. 2022;54(1):2137–50. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Buckley GJ, Strom BL. Eliminating the public health problem of hepatitis B and C in the United States: Phase one report. National Academies; 2016. p. 1–186. [PubMed]
- 26.Trickey A, Fraser H, Lim AG, Peacock A, Colledge S, Walker JG, et al. The contribution of injection drug use to hepatitis C virus transmission globally, regionally, and at country level: a modelling study. Lancet Gastroenterol Hepatol. 2019;4(6):435–44. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Gizamba JM, Finch BK, Wang S, Klausner JD. Insights into the spatial epidemiology of hepatitis C infection: systematic synthesis of area-level determinants and spatiotemporal analyses. BMC Public Health. 2025;25(1):687. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.California Overdose Surveillance Dashboard. Prepared by California Department of Public Health (CDPH) - Substance and Addiction Prevention Branch (SAPB). Available from: https://skylab.cdph.ca.gov/ODdash/. cited 01/16/25.
- 29.Sullivan PS, Woodyatt C, Koski C, Pembleton E, McGuinness P, Taussig J, et al. A data visualization and dissemination resource to support HIV prevention and care at the local level: analysis and uses of the AIDSVu public data resource. J Med Internet Res. 2020;22(10):e23173. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Centers for Medicare & Medicaid Services. National Plan and Provider Enumeration System (NPPES) national provider identifier (NPI) downloadable file. In: Services CfMM, editor. Baltimore, MD, 2023. https://download.cms.gov/nppes/NPI_Files.html.
- 31.California Department of Health Care Services. Medi-Cal Managed Care Provider Listing Sacramento. CA: California Department of Health Care Services; 2023. [Google Scholar]
- 32.Zou H, Hastie T. Regularization and variable selection via the elastic net. J Royal Stat Soc Ser B: Stat Methodol. 2005;67(2):301–20. [Google Scholar]
- 33.Breiman L. Random forests. Mach Learn. 2001;45:5–32. [Google Scholar]
- 34.Liaw A, Wiener M. Classification and regression by randomForest. R news. 2002;2(3):18–22. [Google Scholar]
- 35.Wakefield J, Okonek T, Pedersen J. Small area estimation for disease prevalence mapping. Int Stat Rev. 2020;88(2):398–418. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Mabikwa OV, Greenwood DC, Baxter PD, Fleming SJ. Assessing the reporting of categorised quantitative variables in observational epidemiological studies. BMC Health Serv Res. 2017;17(1):201. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Walters SM, Frank D, Felsher M, Jaiswal J, Fletcher S, Bennett AS, et al. How the rural risk environment underpins hepatitis C risk: qualitative findings from rural southern Illinois, United States. Int J Drug Policy. 2023;112:103930. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Alenzi M, Almeqdadi M. Bridging the gap: Addressing disparities in hepatitis C screening, access to care, and treatment outcomes. World J Hepatol. 2024;16(8):1091. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.King B, Rishworth A. Infectious addictions: Geographies of colliding epidemics. Prog Hum Geogr. 2022;46(1):139–55. [Google Scholar]
- 40.Di Ciaccio M, Villes V, Perfect C, El Kaim J, Donatelli M, James C, et al. Need for integration of hepatitis C (HCV) services in community-based settings for people who inject drugs: results from a global values and preferences survey. Harm Reduct J. 2023;20(1):15. [DOI] [PMC free article] [PubMed]
- 41.Solomon SS, Quinn TC, Solomon S, McFall AM, Srikrishnan AK, Verma V, et al. Integrating HCV testing with HIV programs improves hepatitis C outcomes in people who inject drugs: a cluster-randomized trial. J Hepatol. 2020;72(1):67–74. [DOI] [PMC free article] [PubMed]
- 42.Bennette C, Vickers A. Against quantiles: categorization of continuous variables in epidemiologic research, and its discontents. BMC Med Res Methodol. 2012;12(1):21. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Boodram B, Hotton AL, Shekhtman L, Gutfraind A, Dahari H. High-risk geographic mobility patterns among young urban and suburban persons who inject drugs and their injection network members. Urban Health. 2018;95(1):71–82. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Goetschius LG, Henderson M, Han F, Mahmoudi D, Perman C, Haft H, et al. Assessing performance of ZCTA-level and Census Tract-level social and environmental risk factors in a model predicting hospital events. Soc Sci Med. 2023;326:115943. [DOI] [PubMed] [Google Scholar]
- 45.Krieger N, Chen JT, Waterman PD, Soobader M-J, Subramanian S, Carson R. Geocoding and monitoring of US socioeconomic inequalities in mortality and cancer incidence: does the choice of area-based measure and geographic level matter? the Public Health Disparities Geocoding Project. Am J Epidemiol. 2002;156(5):471–82. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The Hepatitis C case data used in this study are available from the Los Angeles County Department of Public Health, but restrictions apply to their availability. Other datasets supporting the findings are publicly available on the U.S. American Community Survey data portal and other sources listed in the methods section. A partial dataset and R code used in this study can be found on GitHub via this link: [ https://github.com/Gizamba/Machine-learning-and-community-risk-for-spread-of-HCV-infection ]