Sepsis was uncommon after breast cancer surgery in a large United States surgical registry, but patients who developed it faced dramatically higher rates of repeat surgery and unplanned hospital readmission. A new analysis of 328,292 operations found that 838 patients developed sepsis within 30 days, an incidence of 0.26%.
The study, published in Scientific Reports on 2 October 2026, examined operations recorded in the American College of Surgeons National Surgical Quality Improvement Program, or ACS-NSQIP, between 2008 and 2022. Its scale allowed researchers to investigate a complication that is individually rare but potentially life-threatening, while also comparing conventional statistical risk modelling with machine-learning approaches.
A rare complication with substantial consequences
Breast cancer surgery generally carries a relatively low infection risk, yet patients may differ substantially in their vulnerability to severe postoperative infection. Cancer-related changes in immune function, underlying illness, surgical stress and the complexity of an operation can all contribute to postoperative risk.
Among the 328,292 adult female patients included in the analysis, 47.8% underwent partial mastectomy, 38.6% simple mastectomy and 13.6% radical mastectomy. The overall 30-day sepsis rate was 0.26%, but incidence increased with the procedure category. Sepsis occurred after 0.08% of partial mastectomies, 0.40% of simple mastectomies and 0.46% of radical mastectomies. The difference was statistically significant at p<0.001.
The low absolute incidence should not obscure the seriousness of the cases that did occur. Among patients with sepsis, 46.5% underwent reoperation, compared with 4.0% among patients without sepsis. Unplanned readmission occurred in 73.7% of patients with sepsis, compared with only 2.1% of those without it. Both differences were significant at p<0.001.
Researchers used 15 years of surgical registry data
The researchers conducted a retrospective analysis of ACS-NSQIP data covering 2008 through 2022. The primary outcome was sepsis occurring within 30 days of breast cancer surgery according to the registry’s coding criteria. Septic shock was analysed separately and descriptively.
Because only a small fraction of patients developed sepsis, the dataset presented a class-imbalance problem for prediction. The team therefore used multivariable logistic regression to identify factors independently associated with sepsis and assessed discrimination using the area under the receiver operating characteristic curve, or AUC. They also corrected the regression model for optimism through bootstrap validation and compared its performance with machine-learning models trained using explicit methods for handling the rare outcome.
This combination is important because a very large dataset does not automatically guarantee a useful prediction model. A model can identify statistically significant associations while still performing poorly when asked to distinguish which individual patients will experience a rare complication. Internal validation and held-out testing therefore provide a more demanding assessment than statistical significance alone.
Previous sepsis and functional dependence stood out
Several patient characteristics showed strong independent associations with postoperative sepsis after adjustment for the other measured variables. A history of sepsis produced the largest reported adjusted association, with an adjusted odds ratio of 4.51. Functional dependence was associated with 3.80 times the adjusted odds of sepsis, while ASA physical-status class III was associated with an adjusted odds ratio of 3.16. All three associations had p values below 0.001.
These are odds ratios rather than absolute probabilities. They indicate that the relative odds differed substantially between patient groups after statistical adjustment, but they should not be interpreted as meaning that a patient with one of these characteristics has a correspondingly high absolute probability of developing sepsis. The overall event rate remained 0.26%.
Reconstruction was associated with higher adjusted odds
Breast reconstruction was also associated with postoperative sepsis in the adjusted analysis. Autologous reconstruction, which uses a patient’s own tissue, was associated with an adjusted odds ratio of 2.15. Implant-based reconstruction was associated with an adjusted odds ratio of 1.59.
The authors explicitly caution against interpreting these associations as evidence that reconstruction itself causes sepsis. ACS-NSQIP does not contain several potentially important oncological and treatment variables, including details that could influence both the decision to reconstruct and a patient’s postoperative risk. Residual confounding is therefore possible.
That distinction matters clinically. Reconstruction is a major component of breast cancer care for many patients, and an observational association is not a reason to treat reconstruction as inherently unsafe. Instead, the result can contribute to preoperative risk assessment and discussions about how individual patient characteristics, operation type and reconstructive plans combine.
Machine learning did not beat logistic regression
The conventional logistic regression model achieved an apparent AUC of 0.752. After bootstrap correction for optimism, the AUC was 0.746. Its bootstrap-corrected calibration slope was 0.967, close to the ideal value of 1 and consistent with good agreement between predicted and observed risk within the development data.
More complex machine-learning approaches did not produce a statistically significant improvement on the held-out test set. DeLong comparisons between the models yielded p values ranging from 0.077 to 0.409. In this dataset, greater algorithmic complexity therefore did not translate into clearly superior discrimination.
This is a useful result beyond breast surgery. Machine learning can be valuable in clinical prediction, particularly when relationships are highly nonlinear or interactions are difficult to specify in advance. But large clinical datasets do not guarantee that a complex model will outperform an interpretable statistical model. When performance is similar, logistic regression has practical advantages because clinicians can more readily see how individual variables contribute to estimated risk.
What the findings could mean for surgical care
The findings suggest two complementary messages. First, sepsis after breast cancer surgery is uncommon at a population level. Second, when it occurs, it is associated with a substantial burden of further care. The large differences in reoperation and readmission underline why even a complication affecting roughly one in 400 patients deserves attention.
Risk assessment may be particularly valuable for patients with a previous history of sepsis, functional dependence or substantial systemic disease. These factors do not determine an individual’s outcome, but they may help clinicians identify patients who warrant more detailed counselling, perioperative planning or postoperative vigilance.
The results also reinforce the value of separating prediction from causation. A risk model can identify combinations of characteristics associated with an outcome without establishing that changing any single characteristic will necessarily change that outcome. This is especially relevant for reconstructive surgery, where treatment selection, cancer characteristics and patient health may be intertwined.
Important limitations remain
The study is retrospective and relies on routinely collected registry data. This allows unusually large-scale analysis but restricts the researchers to variables captured consistently by ACS-NSQIP. Important oncological and treatment-related information is absent, which limits causal interpretation of associations such as the one observed for reconstruction.
The outcome is also rare. Although the dataset contains more than 328,000 operations, there were 838 sepsis events. Rare outcomes create statistical and practical challenges for prediction, including class imbalance and uncertainty about performance in patient groups that are sparsely represented.
The model underwent internal validation and comparison on held-out data, but external validation in independent health systems would be necessary before treating it as a broadly transportable clinical tool. Surgical practice, patient characteristics and postoperative pathways can differ between hospitals and countries.
Finally, the large differences in reoperation and readmission describe associations among patients who developed sepsis. They should not be read as proof that sepsis alone caused every subsequent intervention or admission, since complicated postoperative courses can involve multiple related factors.
A large dataset clarifies a small but serious risk
By analysing more than 328,000 breast cancer operations, the study puts a rare complication into clearer numerical perspective. Sepsis occurred in only 0.26% of patients overall, yet those cases were accompanied by markedly higher reoperation and readmission rates. The strongest adjusted risk signals came from previous sepsis, functional dependence and ASA class III, while reconstructive procedures were also associated with increased odds after adjustment.
Just as importantly, the modelling comparison showed that sophisticated machine-learning methods did not significantly outperform conventional logistic regression. For clinical risk assessment, a transparent model that performs comparably to more complex alternatives may be especially useful, provided its predictions are validated in new populations before routine use.
Source Information
Study: Sepsis and septic shock after breast cancer surgery: risk modeling and machine learning validation across 328,292 cases
Authors: Jun Jiang, Julius M. Wirtz, Merih Gizlenci and colleagues
Journal: Scientific Reports
Published: 2 October 2026
DOI: 10.1038/s41598-026-73577-6
