A prognostic study published in JAMA Network Open evaluates the use of machine learning to predict major complications following post-mastectomy breast reconstruction, highlighting the potential role of artificial intelligence in supporting personalised surgical decision-making in breast cancer care.
Post-mastectomy breast reconstruction is an important component of survivorship for many patients with breast cancer; however, it is associated with a risk of complications that may require reoperation or hospital readmission. Accurately identifying patients at higher risk remains challenging, particularly given the interplay of clinical, treatment-related, and patient-specific factors. In this context, the study explores whether machine learning models can improve risk prediction by leveraging both structured and unstructured electronic health record data.
CLINICAL SUMMARY
What was examined
A retrospective prognostic study evaluating machine learning models to predict major complications within one year of post-mastectomy breast reconstruction using electronic health record data.
Key findings
- An XGBoost model demonstrated good predictive performance (AUROC 0.83; accuracy ~81%) for major complications.
- Model performance was consistent across implant-based and autologous reconstruction.
- Key predictors included smoking, radiotherapy, body mass index, age, and diabetes.
Clinical implications
- Machine learning models may support personalised risk assessment and patient counselling in breast reconstruction.
- Integration of structured and unstructured health record data may enhance predictive accuracy.
- External validation and prospective evaluation are required before clinical implementation.
The analysis included 411 patients who underwent post-mastectomy breast reconstruction, with models developed to predict major complications within one year, defined as reoperation or rehospitalisation. Multiple machine learning approaches were evaluated, with an extreme gradient boosting (XGBoost) model demonstrating the best performance. This model achieved an area under the receiver operating characteristic curve of 0.83 and an overall accuracy of approximately 81%, indicating good discriminative ability in identifying patients at increased risk.
Importantly, model performance was consistent across both implant-based and autologous reconstruction approaches, suggesting potential applicability across common surgical pathways. Key predictors of complications included established clinical risk factors such as smoking status, receipt of radiotherapy, body mass index, age, and diabetes, supporting the clinical plausibility of the model.
Featured Event
The integration of unstructured data, including clinical notes, was a notable feature of the study and contributed to improved model performance. This approach reflects a broader trend in healthcare toward harnessing the full breadth of electronic health record data to enhance predictive modelling.
Despite these promising findings, several limitations warrant consideration. The study was retrospective and conducted within a single health system, with no external validation cohort. As such, the generalisability of the model remains uncertain, and prospective validation will be required before clinical implementation can be considered. In addition, while the model demonstrated good predictive performance, its impact on clinical decision-making and patient outcomes has not yet been established.
Overall, the findings suggest that machine learning–based risk prediction may offer a valuable adjunct to clinical assessment in post-mastectomy breast reconstruction, with potential applications in patient counselling, surgical planning, and shared decision-making. However, further validation and evaluation in real-world settings will be essential to determine its role in routine practice.
Paper: Shaheen MS, McManus BT, Cullen CM, et al. Machine Learning Model to Predict Post-Mastectomy Breast Reconstruction Complications. JAMA Netw Open. 2026;9(4):e267232. doi:10.1001/jamanetworkopen.2026.7232 Access online here.
