We downloaded a public dashboard dataset from the Florida Department of Health (FDOH) through publicly available FLHealthCHARTS. The dataset included all documented new HIV diagnoses between January 1, 2012, and June 30, 2022. Records including diagnoses among those aged 12 years or younger were excluded.

The Surveillance Tools and Reporting System (STARS), formerly known as PRISM (Patient Reporting Investigation Surveillance Manager), is managed by the FDOH. STARS contains relational data collected during contact tracing interviews, where people newly diagnosed with HIV identify recent partners, both through sexual contact and needle-sharing, who are at higher risk of HIV. Information on both people with HIV and their reported contacts is recorded in STARS. These contact tracing interviews are conducted by disease intervention specialists (DIS), who also obtain additional information on demographics and risk behaviors. Partners who are identified are then contacted by DIS to encourage HIV testing, and, if positive, are interviewed again to identify contacts. In general, DIS successfully contacted 90% of people newly diagnosed with HIV and contacted 65%‐70% of the named partners in Florida. STARS data elements used in the current analyses include individual ID and ID from their linked partners, demographic information (age, gender, or race or ethnicity), county of residence, and HIV risk information (male-to-male sexual contact [MMSC], hetero, IDU, etc). To extract data from STARS for network analysis, we retained only records for individuals that could be attributed to a single, specific demographic-behavioral profile (DBP; ie, without missing DBP attributes), and had at least 1 connection with another individual that could also be attributed to a single, specific DBP. We filtered out the records reporting modes of exposure not consistent with the DBPs, such as blood transfusions or prenatal exposures. The remaining modes of exposure matched the modes of exposure in the FDOH public dataset. If an individual showed multiple records over time, we used the median age from these records. Unlike the FDOH public dataset, STARS reported combinations of race and ethnicity. To be able to apply the DBPs to STARS, all people who self-identified as Hispanic were classified as Hispanic, regardless of race. Those who identified as non-Hispanic were further classified based on the race reported in STARS.

The authors abide by the Declaration of Helsinki. This study protocol was approved by the University of Florida’s Institutional Review Board (IRB) and by FDOH’s IRB (protocol #IRB201901041 and #2020‐069, respectively) as exempt. We received data extracts from FDOH’s STARS in a fully deidentified format according to the HIPAA (Health Insurance Portability and Accountability Act).

The dashboard data used in this study are publicly available at FLHealthCHART.gov. The STARS data are managed by FDOH. Researchers seeking access for replication purposes need to submit a data request to the FDOH (research@flhealth.gov) by state, federal regulations, and required ethical and privacy policies, including IRB approval by FDOH and execution of a data user agreement. Requests are independently reviewed by FDOH.

We identified a total of 4 demographic and behavioral variables common in both FDOH public data and STARS and used their combinations to create DBPs. The four variables are (1) gender: male (M), female (F); (2) age: <25, 25‐49, 50+ years; (3) race or ethnicity: Hispanic, non-Hispanic Black, non-Hispanic White, and non-Hispanic other; and (4) HIV risk group (ie, mode of exposure): MMSC, hetero, IDU, MMSC or IDU, and other. We explored the option of extracting DBPs via unsupervised clustering. Both elbow and silhouette analyses indicated 8 as the suggested number of clusters (data not shown). With the number of clusters (k=8), both agglomerative and divisive clustering approaches based on Gower distance split the data into groups perfectly and uniformly characterized by gender and race or ethnicity (ie, clusters where all the members belong to the same combination of gender and race or ethnicity). When using k>8, data were further partitioned by age (data not shown). However, none of the profile identities from the unsupervised clustering methods include the only behavioral factor (ie, HIV risk group), which we believe is important to include as it is a modifiable factor. Therefore, we decided to extract DBPs based on their frequency, as explained above, instead of using unsupervised clustering.

Descriptive analysis was used to demonstrate the frequencies for identified DBPs and our study sample characteristics. Multiple correspondence analysis (MCA) was implemented using the FactoMineR and factoextra libraries in R Studio (R Foundation). MCA is an extension of simple correspondence analysis and is used to summarize and visualize data tables with more than 2 categorical variables. This analysis was used to show a graphical representation of the similarity between identified DBPs.

Additionally, we used the identified DBPs to characterize Florida counties and HIV coordination areas. HIV coordination areas (“areas”) are groups of counties that operate under the leadership of an HIV/AIDS program coordinator who oversees the region’s HIV health care and prevention efforts within the FDOH [24]. Each county and area can be represented by an array of 37 slots, each representing the counts for a specific patient type. To compare counties and areas, we calculated the county-to-county and area-to-area rank (Spearman) correlation. We then drew the dendrogram based on the correlation matrix and found groups of similar counties and areas.

Lastly, we performed network analysis using STARS data based on the DBPs identified. The Leiden algorithm [25] is a network node partitioning method that focuses on identifying clusters characterized by maximized intracluster connections and minimized intercluster connections. In this study, the algorithm was implemented using the leidenalg library in Python (Python Software Foundation) [26], with modularity serving as the optimization function and setting the resolution parameter to 0.2 to control cluster counts.

Sample characteristics are shown in Table 1 stratified by data source. The Florida Chart public data and STARS data show similar sample characteristics: mostly male, aged 25‐49 years, exposed to HIV via MMSC. Non-Hispanic Black is the most common race or ethnicity group in both data sources.

Using the Florida public data, we identified 84 different DBPs among 44,090 people newly diagnosed with HIV between 2012 and 2022. Of these 84 DBPs, 36 cover 97% of the HIV incident cases (42,734/44,090), and each included 100 or more individuals. The remaining DBPs were combined into an “other” group, making 37 DBPs total. These DBPs are reported in Table 2. The results from the MCA, showing a graphical representation of the similarity between these DBPs, are provided in Figure S1 in Multimedia Appendix 1.

The dendrogram showing the correlation matrix between county and HIV coordination areas in Florida using the identified DBPs, is depicted in Figures 1 and 2. Many HIV areas share similar distributions of DBPs and predominant DBP types, and 2 main clusters were identified. The first area cluster included many counties in the western part of Central Florida and Southern Florida (areas 9, 14, 8A, 10, 7A, and 6) and the second represented counties in Northern Florida, the Panhandle (Northwestern Florida), and the east coast of Central Florida (areas 5, 7B, 1 8B, 4, 15, and 3/13). Area 12 (Flagler and Volusia) was dissimilar from most of the other areas. Area 11A (Miami-Dade) was similar to the Central or Southern Florida cluster but very dissimilar from the Northern Florida or Northwestern Florida cluster. Among the counties, rural counties tended to be more similar but there was no clear delineation between rural and urban counties. We identified 6 clusters of varying sizes. The largest included 12 counties: Polk, Palm Beach, Lee, Seminole, Hillsborough, Manatee, Marion, St. Lucie, Indian River, Orange, Broward, and Collier. The next largest cluster included 8 counties: Pinellas, Brevard, Sarasota, Hernando, Bay, Okaloosa, St. Johns, and Pasco. The remaining clusters included between 6 and 2 counties: Jackson, Alachua, Leon, Putnam, and Columbia; Duval, Escambia, and Clay; Flagler and Volusia; and Miami-Dade, Osceola, and Hendry. The Duval, Escambia, and Clay cluster were also similar to the 2 largest clusters, while the 1 containing Miami-Dade also shared many similarities with the counties in the largest cluster.

STARS represents contact tracing data to create networks of people with HIV and people exposed to HIV. We extracted 7944 individuals from STARS, sharing 4928 connections. The DBP counts in STARS correlate strongly with the counts of the FDOH public dataset (Spearman correlation=0.93). STARS individuals formed 3097 subnetworks (ie, networks that are not connected to one another). Of those, 2518 (~81%) are 2-node networks, 349 (~11%) are 3-node networks, 224 (~7%) networks have 4‐19 nodes, while the remaining 6 networks have 20‐246 nodes (shown in Figure 3). Notably, hetero decreases decisively with the network size, is found more commonly in smaller networks, and its representation decreases with network growth. Table 3 reports the characteristics of the STARS population based on the network size. In the 2 largest contact networks (Figures 3A and 3B), males of Black, Hispanic, and White ethnicities aged 25‐49 years predominantly constitute both networks. Interestingly, when focusing on hub nodes (nodes with more than 2 links in the network), aside from the dominant groups, young Hispanic and Black males younger than the age of 25 years also comprise a significant portion.

Using the FDOH public dataset, we characterized DBPs for all documented people newly diagnosed with HIV in Florida between 2012 and 2022. A total of 37 DBPs were identified. The attempt to extract HIV behavioral risk groups via unsupervised clustering stratified data based on demographics only, as the behavioral component (ie, the mode of exposure) was not a discriminating feature even when deliberately overclustering. Therefore, a list of all combinations of variable categories was used to characterize DBP. DBPs were further used to examine area and county similarity across Florida and characterize contact networks. The results from these analyses can be used to inform future HIV prevention and intervention efforts by guiding public health professionals in making decisions on resource allocation and policy for their strategies to end the HIV epidemic. These findings can also potentially help public health workers in the field, allowing them to better understand trends in the HIV epidemic in their areas to identify and help community members most in need. Moreover, these findings show that contact tracing surveillance data can be used to investigate the dynamics of HIV transmission and support further research and development in this area for the creation of more informative models. Integrating these models into routine surveillance would enable regular communication of results to program staff and field workers, ensuring the continued effectiveness of current efforts while also allowing for strategies to be adapted in response to evolving needs.

The 6 largest networks (30+ nodes and 511 individuals) are almost exclusively composed of individuals with MMSC (ie, MSM). While the female, non-Hispanic Black, aged 25‐49 years, and hetero profile is the fourth largest DBP, none of the large networks involve females of any race or age. This may indicate that HIV transmission to women does not involve large contact networks with many potentially at-risk partners. If this is the case, then interventions that focus solely on increasing prevention efforts among those with multiple sexual partners may still leave many women vulnerable to HIV. The implementation of past prevention efforts has made similar missteps. For example, the original guidance for prescribing PrEP focused largely on and marketed toward MSM and included language that excluded at-risk women. As a consequence, PrEP awareness and uptake have remained low among at-risk women, although the guidelines have been updated to try and address this [27,28].

Those with MMSC comprise the dominant HIV risk group among the large networks identified. This matches national trends, with 67% of new HIV diagnoses in the United States in 2021 involving MMSC [29]. Additionally, in the United States, 79% of new HIV cases were among men, 40% were among Black individuals, 29% among Hispanic individuals, and 37% among those aged 25‐34 years [29], which closely mirrors the largest transmission networks that we identified in Florida. People with IDU may appear in the network, but no major network is characterized solely by IDU. This might be because IDU is not a common behavior, with less than 2% of the US population reporting IDU in 2018 [30], and ~4% in the extracted STARS DBPs report IDU. However, IDU is still a prominent means of HIV acquisition and transmission, accounting for about 10% of new cases [1], and should still be considered in prevention and intervention efforts. Hetero exposure is also not common in the identified networks, again indicating that exposure via hetero may occur mainly in small networks.

All of the large HIV transmission networks we identified involve individuals from multiple counties. Most of these networks were localized by region or represented multiple urban centers in the state. While neighboring counties are often grouped together into HIV coordination areas to address HIV, not all of these networks are contained in 1 area. Moreover, Florida has a high population who are seasonal residents, migrants, and transient tourists from all over the world. Close communication between county health officials is essential to address the HIV epidemic as these transmission networks extend beyond immediate neighboring counties. One of the goals of the HIV coordination areas is to help facilitate between-county communication, even though some areas are only 1 county (eg, areas 7B and 11A).

We identified 2 distinct clusters of HIV areas that divide the state roughly in half (Northern or Northwestern and Central or Southern) and 6 county clusters. Another level of categorization into 2 larger HIV area groups might also help increase communication between the counties that we identified as being within similar clusters. For example, the largest county cluster, represented counties within areas 6, 7A, 8A, 9, and 15, which are all in the Central or Southern Florida cluster, plus a county in area 14. Improving both interarea and intercounty communication could help improve information sharing between counties with similar DBPs, and information shared could include successful interventions and lessons learned from interventions that were less successful to improve future decision-making and resource allocation. Increasing discussions within these clusters could also improve information sharing between neighboring counties that happen to be in different HIV areas.

There are existing evidence-based and impactful strategies for addressing HIV, such as PrEP [16,17], postexposure prophylaxis [31], condoms [32], and syringe services programs [33,34]. Our findings support increasing the use of these and other prevention strategies in areas and among groups we identified as having high transmission. For example, our findings identified those with MMSC in Miami-Dade as a high-transmission contact network, and past research has found that there is low uptake of PrEP among MSM in Miami-Dade [35]. Additionally, testing efforts among high-risk groups can be increased in geographic areas and among groups identified in this study. Our findings can enable focused and efficient resource allocation, prioritizing efforts and funding for groups at high HIV risk and areas with increased HIV transmission.

When characterizing DBPs in identified contact networks, we intentionally only include these DBPs at an aggregated level due to ethical concerns. Moreover, we would like to emphasize that many of the groups and behaviors that we have identified in these clusters are groups that already experience high levels of stigma and discrimination. The goal of this work is not to contribute to this existing stigma, but to highlight areas of need and groups who might need targeted intervention to reduce their risk of HIV. There is a continuous need to conduct ethical discussions to avoid the possible harm of advanced methodologies to vulnerable groups, especially in the context of HIV stigmatization [22].

Our results should be interpreted in consideration of some limitations. We removed STARS records without linked partners or who had missing or inconsistent records, so our network analyses were conducted on a subset of STARS data. The extracted STARS DBPs have a high correlation with DBPs identified from the Florida public dataset [13], indicating good representativeness. The data are retrospective, which could limit their usefulness for public health workers in the field if they simply confirm what they already know about the HIV epidemic in their area rather than provide new information. However, these findings support further investigation and application of contact tracing surveillance data to provide a more up-to-date and accurate understanding of transmission dynamics, allowing more targeted and impactful prevention and intervention efforts. A final consideration is that extracting DBPs from the FDOH public dataset via unsupervised clustering leads to clusters solely characterized by nonbehavioral components, such as gender and race or ethnicity, therefore missing the mode of exposure, a key feature to explore the HIV contact networks. A primary strength of this study is our public health data, with high representativeness of people with HIV in the last 10 years in Florida.

In our contact network analysis, we identified several distinct risk groups (young MSM, non-Hispanic Black MSM, Hispanic MSM, Hispanic women, and people with IDU) and clusters of transmission in Florida (divided generally into Northern or Northwestern and Central or Southern, with some similarities identified at the county level). These results can be used to inform targeted interventions aimed at reducing HIV diagnoses in the state. These groups we identified are already recognized as having a disproportionately high risk of HIV in Florida and nationwide, and our findings have shown areas in the state where these disparities might be exacerbated. Inequitable distribution of the burden of HIV among these groups contributes to existing health disparities and shows that the HIV epidemic will not be ended by a one-size-fits-all approach. Network analysis can help state and county health departments not only identify disparities and areas or groups with high transmission but also make the best use of their HIV intervention and prevention resources to help their community members.