Streptococcus pyogenes, also referred to as Group A Streptococcus (GAS), is a Gram-positive human pathogen that can cause severe invasive diseases such as bacteraemia, necrotising fasciitis, and streptococcal toxic shock syndrome, as well as milder non-invasive diseases such as pharyngitis and impetigo. The case fatality rate of invasive GAS (iGAS) infection is high, around 10% overall and more than 30% among patients with necrotising fasciitis and streptococcal toxic shock syndrome in the USA.^1,2 In recent years, incidence of iGAS disease in the USA increased substantially, from 3·2–4·3 cases per 100 000 people during 2002–12 to 7·6 cases per 100 000 people in 2019.² The incidence of iGAS decreased to 6·1–6·3 per 100 000 people in 2020–21 but rapidly rebounded to 8·1 cases per 100 000 people in 2022.² Previous studies have documented frequent occurrence of iGAS outbreaks among long-term care facility residents, people experiencing homelessness, and people who inject drugs, potentially due to increased disease susceptibility, crowded living, and suboptimal hygiene conditions among these disadvantaged populations.^3,4 During an outbreak, iGAS isolates from patients within the closely connected transmission network show highly similar genomic sequences (genomic clustering) because of their close relatedness to their common ancestor. Two iGAS cases can be defined as genomically clustered if genomic distance between case isolates is lower than a small single-nucleotide variant (SNV) threshold.^5,6

Genomic clusters of iGAS disproportionally affect disadvantaged communities⁷ and contribute heavily to the expansion of antimicrobial resistance in this pathogen.⁸ Conventional iGAS cluster detection mainly uses epidemiological links including timing, location, and contact of suspected cases.³ In the past 10 years, microbial whole-genome sequencing (WGS) data have been increasingly used to support cluster identification in both health-care and community settings.^6,9,10 An iGAS case that is not yet part of existing clusters could be later identified as part of a new genomic cluster (cluster formation) due to the transmission of related isolates from a persistent bacterial lineage reaching a detectable level in surveillance of invasive infections. However, the risk of cluster formation and factors influencing such risk are largely unknown.

Knowledge of iGAS cluster formation risk could be valuable in understanding disease transmission mechanisms, informing chemoprophylaxis and other infection control strategies, and prioritising prevention resources. Here, we aim to assess the risk of an iGAS infection becoming a new transmission cluster (cluster formation) and identify pathogen and host features associated with increased risk.

Between Jan 1, 2015, and Dec 31, 2019, 11 279 iGAS cases were identified through ABCs. 9699 (86%) of 11 279 iGAS isolates were characterised by WGS and passed the study’s quality filter. 6143 iGAS cases from 2016 to 2018 were assessed for at-risk status of forming a new genomic cluster (figure 1), and 2764 cases were identified as at risk due to no evidence of clustering with any previous cases. The main analysis focused on these 2764 cases at risk (905 [33%] for 2016, 943 [34%] for 2017, and 916 [33%] for 2018) and their case features on detection date (appendix p 16).

The median age of cases at risk was 56 years (IQR 37–70); and 1266 (46%) of the 2764 cases were female and 1498 (54%) were male. Many cases had comorbidities or established risk factors for iGAS infection: 802 (29%) had diabetes; 537 (19%) had a BMI of 30 kg/m² or higher; 318 (12%) had chronic skin disease; 269 (10%) were either people experiencing homelessness (141 [5%]) or people who inject drugs (204 [7%]), 76 (3%) of whom were both; and 229 (8%) had alcohol abuse. The three most common syndromes among cases at risk were cellulitis (975 [35%]), bacteraemia without focus (594 [21%]), and pneumonia (378 [14%]). 1978 (72%) of the 2764 cases at risk were infected by one of the ten most common emm types: type 1 (398 [14%]), type 89 (358 [13%]), type 12 (274 [10%]), type 28 (242 [9%]), type 4 (195 [7%]), type 77 (125 [5%]), type 11 (125 [5%]), type 3 (116 [4%]), type 6 (85 [3%]), or type 92 (60 [2%]). WGS-based assessment of antimicrobial resistance found that levofloxacin non-susceptible case isolates were uncommon (56 [2%]), but there were notable rates of erythromycin non-susceptibility (543 [20%]), clindamycin non-susceptibility (502 [18%]), and tetracycline non-susceptibility (585 [21%]).

By the end of 2019, cluster formation events were observed among 656 of 2764 cases at risk (23·7% [95% CI 22·2–25·4]) while the other 2108 cases at risk remained unclustered (figure 2A). The median trace-forward time for all cases at risk was 732 days (IQR 410–1087). Among the 656 cases at risk that had observed cluster formation events, the median time to event was 94 days (32–235). 587 of 656 (89·5% [95% CI 86·9–91·6]) cluster formation events occurred between a case at risk and a subsequent iGAS case in the same surveillance site. The Kaplan–Meier survival curve in figure 2B shows the cumulative incidence of cluster formation over time. The estimated cumulative incidences at 30 days were 0·057 (95% CI 0·048–0·066), at 90 days were 0·12 (0·11–0·13), and at 180 days were 0·16 (0·15–0·18). The estimated instantaneous event rate (hazard) appeared to be high immediately after detection of cases at risk and decreased consistently afterwards (figure 2C), which was consistent with the diminishingly ascending cumulative incidence (figure 2B).

The cluster formation events across 72 subgroups, each defined by a specific case feature (eg, infection by the emm1 strain or homelessness), are shown in the appendix (p 16). Subgroups with the highest cluster formation proportions were in individuals infected by emm92 (26 of 60 [43·3%; 95% CI 31·6–55·9]), people who inject drugs (86 of 204 [42·4%; 35·6–49·0]), and people experiencing homelessness (54 of 141 [38·3%; 30·7–46·5]). Conversely, the lowest proportions were observed in individuals infected by emm28 (32 of 242 [13·2%; 9·5–18·1]), emm4 (27 of 195 [13·8%; 9–7–19·4]), and a strain positive for Alp family adhesion (R28; 65 of 427 [15·2%; 12·1%–18·9%]).

For 29 of the 72 subgroups, the survival (cluster formation event free) function within the subgroup (feature-positive cases) differed significantly from that outside the subgroup (feature-negative cases) based on the log-rank test (corrected p<0·05; appendix p 16), indicating the 29 binary case features were significantly associated with cluster formation risk in the univariable analysis. We did not observe that cluster formation risk varied by age as a continuous variable (appendix p 43).

Multivariable Cox regression confirmed that emm type, patient residence, injection drug use, and season of detection were significant risk factors after adjusting for all covariates in the Cox model (table). No evidence of violation of the proportional-hazards assumption was observed (appendix p 42; p>0·05 for all covariates). Compared with emm89, higher risk of cluster formation was associated with emm3 (adjusted hazard ratio 2·72 [95% CI 1·82–4·06]), emm1 (2·37 [1·71–3·30]), emm6 (2·28 [1·49–3·51]), emm92 (2·21 [1·38–3·56]), and emm12 (1·47 [1·05–2·06]). Increased risk was observed among long-term care facility residents (1·78 [1·29–2·45]) and people experiencing homelessness (1·42 [1·01–1·99]), compared with cases at risk who lived in private residence before the onset of iGAS infection, as well as in people who inject drugs 2·08 (1·59–2·72) compared with people who do not inject drugs. Cases at risk detected in the autumn–winter seasons (October to March) were significantly more likely to form clusters (1·34 [1·14–1·57]), compared with cases at risk detected in the spring–summer seasons (April to September). We show the adjusted probability of cluster formation based on the fitted Cox model for the four independent risk factors: emm type (figure 3A), residence (figure 3B), injection drug use (figure 3C), and autumn–winter season (figure 3D).

Additional analysis based on one SNV and five SNV genomic cluster thresholds indicated similar risk factors (appendix pp 19–27), including infection by emm types 1 and 12, residence in a long-term care facility, injection drug use, and the autumn–winter season; homelessness was not a significant risk factor only when the less stringent SNV threshold (five SNV) was evaluated (appendix p 27). These factors, including homelessness, were also observed in the mixed Cox model (appendix p 31).

The machine-learning approach for predicting cluster formation risk is shown in figure 4A. In model comparison analysis, the random survival forests model exhibited the highest performance (mean C-index 0·6071 [95% CI 0·5811–0·6330]), with four other models also surpassing the Cox proportional hazards model, albeit modestly (appendix p 33). Consequently, the random survival forests model was chosen for this study and trained using data from the derivation cohort (n=1382). Notably, variables of highest importance to the random survival forests model (appendix p 44) largely overlapped with significant variables in the univariable analysis (appendix p 16).

The model’s predictive performance was assessed in the validation cohort (n=1382), achieving a C-index of 0·63 (95% CI 0·60–0·66; figure 4B). The model output categorised the cohort into three risk groups (figure 4C). The three risk groups (low, moderate, and high) were defined using the 25th (12·8) and 75th (34·5) percentiles of derivation cohort risk scores as thresholds (appendix p 45). The proportion of cases with observed cluster formation was 14% (95% CI 10–18; 37 of 274) in the group at low risk, 22% (19–25; 161 of 738) in the group at moderate risk, and 35% (30–40; 130 of 370) in the group at high risk. The cumulative incidence curves representing moderate and high risk diverged sharply from the curve representing low risk in the first month (figure 4D). Continued separation of the high-risk curves from moderate-risk curves appeared to continue for another 3 months. The 90-day risk of cluster formation was 0·03 (95% CI 0·01–0·05) for the group at low risk, 0·10 (0·08–0·13) for the group at moderate risk, and 0·21 (0·17–0·25) for the group at high risk.

Leave-one-out cross-validation on the derivation cohort (appendix pp 46–49) and additional analyses based on one SNV and five SNV cluster thresholds (appendix p 34) confirmed the random survival forests model’s consistent predictive performance.

This study, using population-based and laboratory-based surveillance data in ABCs, showed that 24% of iGAS cases who were not part of an existing genomic cluster on detection date subsequently became part of a new genomic cluster, typically within a year. This proportion was much higher in certain subgroups (eg, people experiencing homelessness, people who inject drugs, and people infected by some emm types) and in the autumn–winter season. Using case information available on the detection date (before any cluster formation) the predictive model consistently identified iGAS cases with the highest future cluster formation risk. This insight could inform potential targeted intervention strategies. This study revealed previously unknown predictive values of known risk factors and, to our knowledge, estimated cluster formation probabilities for the first time. This study also identified specific strain features associated with transmission risk, providing a basis for generating hypotheses for future research on risk prediction and understanding mechanisms.

There are three potential scenarios for the formation of an iGAS cluster: (1) direct transmission of GAS from one invasive case to another; (2) a local spread of non-invasive GAS following the detection of the initial iGAS case; and (3) occasional invasive disease resulting from local GAS carriage. The formation of an iGAS cluster is more probable under scenarios 1 and 2, whereby the iGAS cases are linked by few transmission events. Scenario 3 is less probable because invasive disease is a rare outcome of GAS carriage. In cases of sporadic iGAS, the isolates would be separated by numerous carriage instances before reconnecting to their most recent common ancestor, resulting in a substantial SNV distance between them. However, a higher than usual level of GAS transmission due to high prevalence of GAS carriage in a population could result in the appearance of iGAS genomic clusters by chance. Our findings and the predictive model can be interpreted as assessing the combined effects from these scenarios, influenced by their respective likelihoods.

GAS spreads mainly through individuals with milder GAS diseases (such as pharyngitis and impetigo) or asymptomatic carriers. GAS displays greater emm type diversity and a higher occurrence of non-throat-tropic strains in developing countries or disadvantaged populations in developed nations.^4,23,24 Our study found a link between higher iGAS cluster formation risk and disadvantaged groups in the USA. High-risk emm types, such as emm1 and emm92, might be more effective at causing iGAS after initial infections or more transmissible in disadvantaged groups. Regional variations in cluster formation risk (appendix p 16) could be due to complex population immunity and economic conditions such as poverty rates, housing density, and health-care access.

Our study has several limitations. First, GAS strain distribution and disease epidemiology vary globally,^25,26 making the generalisability of identified risk factors and groups uncertain. Furthermore, clusters might form outside the surveillance area, potentially underestimating events. Second, the machine-learning model, derived from half the data and validated with the rest, represents an ideal scenario for prediction accuracy, but results might vary with datasets of different characteristics and structures, possibly affecting accuracy. Additionally, key predictive features such as direct contact with infected individuals and host immune responses to previous GAS exposures are missing from the data currently available, contributing to modest predictive accuracy (C-index around 0·63) in machine-learning models. Incorporating more relevant demographic and immunological information could enhance performance. Despite its modest accuracy, the model offers some use in risk stratification and serves as a baseline for future improvements. Third, genomic clusters defined by different SNV threshold values still share varying degrees of ancestry with each other. In the univariable and multivariable analyses, the shared ancestry between genomic clusters were not adjusted for. In this study, iGAS genomic clustering was defined by bacterial genomic distance, yet the clustering was influenced by more than just bacterial factors. New cluster predictors—such as residence in a long-term care facility, homelessness, and injection drug use—are inherently epidemiological. The iGAS genomic clustering often reflected the closer proximities of people that varied seasonally and was linked to underlying epidemiological transmission dynamics.

In conclusion, we used specific definitions of laboratory confirmed iGAS disease genomic clusters to quantify the risk of cluster formation and identify subgroups of patients with significantly higher risks. Predicting iGAS cluster formation risk might enhance prevention strategies by identifying cases at high risk before clusters occur, allowing for more effective allocation of public health resources. This approach supports targeted surveillance, improved infection control in health care, and focused educational campaigns in high-risk areas.