Utilizing Artificial Intelligence for Predicting Postoperative Complications in Breast Reduction Surgery: A Comprehensive Retrospective Analysis of Predictive Features and Outcomes

Abstract

What if we could predict complications from breast reduction surgery with unprecedented accuracy? In 2023, over 75,000 women in the United States of America underwent bilateral breast reduction (BBR) surgery,¹ a procedure that, while largely successful, still carries risks.^2,3 Complications range from minor issues like delayed wound healing to more severe outcomes such as infection, hematoma, and necrosis. Publications showed a complication incidence rate between 6.2% and 43%^2,4-6.

Breast reduction is aimed at alleviating the physical and psychological burdens associated with macromastia. Accurately predicting these complications can significantly enhance patient care by enabling tailored preoperative counseling and individualized postoperative management strategies.

Previous studies have been done to identify complication predictors; these studies have shown a correlation between features such as increased BMI, smoking, diabetes, and age, contributing factors that increase the complication rate.^2,4-6 Moreover, several publications have suggested scoring methods to predict risk for surgical-site morbidity after BBR.⁷

In this context, integrating artificial intelligence (AI), specifically machine learning (ML), into medical practice offers a promising avenue for enhancing predictive accuracy and optimizing patient outcomes. AI refers to the capability of machines to perform tasks that typically require human intelligence, such as pattern recognition and decision making.⁸ ML is a subset of AI that enables computers to learn from data without being explicitly programmed. It is particularly valuable in healthcare because it allows for the analysis of large datasets to identify patterns and correlations that may not be evident through traditional methods.

In contrast to conventional statistical approaches, ML models do not rely on predefined hypotheses; they adapt to the data, often leading to more accurate predictions. This article explores the development of an ML-based model designed to predict complications following breast reduction surgery, providing surgeons with a data-driven tool to identify patients at higher risk for postoperative complications. ML models have been developed in different fields of medicine, demonstrating improved accuracy; moreover, they enable better understanding of the contributing features to the prediction.⁹

By leveraging clinical and operative data, the model aims to provide surgeons with a powerful tool to identify patients at higher risk for postoperative complications. This predictive capability has the potential not only to improve the surgical outcomes but also to improve patient education and manage expectations, as well as reduce healthcare costs by potentially preventing complications before they arise.

METHODS

Data Collection and Statistical Analysis

Data were collected from electronic medical records (EMRs) of breast reduction procedures performed at the Tel Aviv Medical Center (TLVMC) from July 2017 to July 2024 as a retrospective study. Clinical features, operative reports, and postoperative follow-up sections were gathered for each procedure.

This research was approved by the TLVMC ethics committee (TLV-0031-23).

We documented a complication as any signs of infection, dehiscence, seroma, hematoma, fat necrosis, or nipple–areola complex necrosis that required treatment and that occurred up to 30 days following the operation. Major complications were defined as those requiring admission of the patient for further treatment or surgical revision.¹⁰

Following data collection and annotation, statistical analysis of procedure complications and their association with discrete predictor features (χ² testing) or continuous features (2-sample t test).

Classification Model

We employed a gradient-boosting decision tree (GBDT) for the classification model, constructed using XGBoost (version 1.2).¹¹ XGBoost, an open-source optimized distributed gradient-boosting library, is widely recognized as the most effective method for structured problems in a variety of real-world applications.¹² It shares features with other decision tree ensemble algorithms, such as the ability to handle both numerical and categorical data. Additionally, XGBoost is particularly suitable for these models because of its sequential learning of trees, where each tree represents a gradient step with respect to a loss function and the inclusion of a randomization parameter to decorrelate individual trees.¹³

Model Evaluation

We applied 5-fold cross-validation to evaluate our ML model’s performance and minimize the risk of overfitting. Overfitting occurs when a model learns too closely from the training data, capturing noise or patterns specific to that dataset rather than general trends. As a result, the model performs well on the training data but poorly when applied to new, unseen data. Cross-validation helps address this issue by testing the model on multiple subsets of data, thus assessing its ability to generalize.

In 5-fold cross-validation, the dataset is divided into 5 equally sized “folds” or subsets. For each iteration, 4 folds are used to train the model, and the remaining fold is used to test it. To illustrate, in the first iteration, Folds 1 through 4 (80% of the data) are used for training the model, and Fold 5 (20% of the data) is used for testing.

In the second iteration, Folds 1, 2, 3, and 5 are used for training, and Fold 4 is used for testing. This process is repeated 5 times, with each fold serving as the test set once. The results from each iteration are averaged, providing a more reliable and unbiased estimate of the model's performance on new data. This approach reduces the likelihood of overfitting, ensuring that our model is better equipped to make accurate predictions on future, unseen cases.

Evaluation Metrics

The area under the receiver operating characteristic curve (AUC-ROC) was selected as the evaluation metric for the models. ROC graphs are invaluable for assessing classifiers; the AUC-ROC reflects the probability that the classifier will rank a randomly chosen positive instance higher than a randomly chosen negative instance.¹⁴ To adhere to the reporting standards for ML predictive models in biomedical research, we also reported accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).¹⁵ All ML models were implemented using Python (version 3.7).

Model for Clinician Use

We aim for clinicians to utilize our models while also understanding the impact of features at the surgeon's discretion on the model output. Thus, we created a model that estimates the specimen weight based on the collected clinical features and plots the probability of complications as a function of specimen weight.

The outcome will be a user interface in which the input will be clinical data and output would be a plot of the probability of complication according to the resected specimen weight for a specific predicted range.

RESULTS

During the mentioned period, 322 cases of breast reduction were performed, accumulating to 629 breasts. The majority of surgeries were BBR; however, several procedures were unilateral, such as mastectomy and alloplastic reconstruction on one side, and a reduction mammoplasty on the other breast. Data analysis includes patient and breast characteristics, operative features, and postoperative follow-up descriptions of complications, as seen in Table 1.

Major complications had occurred in 47 breasts (7.4%). The most common complication was wound dehiscence (59.6% of all major complications), followed by infection, hematoma, and nipple necrosis (46.8%, 21.3%, and 21.3%, respectively). Further description of complications is seen in Table 2.

Weight, SN-N, specimen weight, and liposuction volume, when performed, were statistically significant major complication features compared with the no-complication group. When comparing severe complications to the rest of the cohort, only SN-N, specimen weight, and liposuction volume were statistically significant.

The major complication prediction model reached an average AUC-ROC score of 0.83, with an accuracy of 0.93, sensitivity of 0.66, specificity of 0.43, PPV of 0.66, and NPV of 0.95 (Figure 1). The models’ scores are considered outstanding.

Figure 1.

Major complications area under the receiver operating characteristic curve.

Open in new tabDownload slide

DISCUSSION

The incorporation of AI into clinical practice represents a paradigm shift in how we approach complex medical decision making, particularly in plastic surgery.^16,17 Our study demonstrates the potential of an AI-based model to predict complications following breast reduction surgery based on clinical and operative features, offering a promising tool for surgeons to enhance surgical planning and patient outcomes. By utilizing a GBDT model, we were able to capture nonlinear relationships between multiple variables, thereby providing a more nuanced and accurate prediction of complications compared with traditional statistical methods.

A key strength of AI in clinical settings is its ability to process and analyze vast amounts of data rapidly, identifying patterns that might be imperceptible to the human eye. Although traditional methods rely on predetermined hypotheses and may overlook complex interactions between factors, ML models like ours adapt to the data, enhancing the precision of complication predictions.

Unlike traditional risk assessment methods that often rely on fixed cutoff values (eg, BMI > 30 or age >60) to categorize patients into risk groups, AI models provide a more nuanced approach by assigning a probability value to each case based on its unique combination of features. For instance, instead of simply classifying a patient as “high risk” or “low risk” based on a single threshold, the model evaluates how factors such as specimen weight or SN-N distance interact and contribute to the likelihood of complications. This probabilistic approach offers a continuum of risk rather than rigid categories, enabling surgeons to make more personalized and informed decisions about each patient. By capturing the complexity and variability inherent in clinical scenarios, the model provides a level of precision that static thresholds cannot achieve.

Our results demonstrated that key predictive features, such as specimen weight, SN-N distance, and liposuction volume, were significantly associated with major complications. The model achieved an AUC-ROC of 0.83, indicating excellent discrimination between patients with or without major complications. The model's predictive power highlights the importance of preoperative risk stratification, allowing surgeons to identify patients at higher risk for adverse outcomes. Moreover, the model's ability to generalize well across different subsets of the data strengthens its potential as a reliable tool in clinical settings.

An additional important aspect of our model is its high NPV. The likelihood of this outcome is very high when the model predicts no complications. This has significant implications for surgical planning and resource allocation. For instance, the ability to confidently predict the absence of complications allows for optimized scheduling, such as avoiding surgeries with a higher predicted complication risk before weekends or holidays, when resources are limited. However, we acknowledge that the model's predictive utility during the preoperative phase may be constrained by the fact that specimen weight—one of the most significant predictors of complications—is typically unknown before surgery. Despite this, the tool can still assist in preoperative counseling by integrating other key clinical features, such as SN-N distance, patient weight, and liposuction volume. Future iterations of the model may incorporate predictive estimations of specimen weight based on preoperative data to enhance its applicability in surgical planning and improve its preoperative relevance.

Interestingly, based on the findings of this study, we suggest that patients categorized as current smokers, who had ceased smoking at least 4 weeks before surgery, did not experience a significant increase in complications. The authors challenge traditional perspectives on smoking cessation timelines based on this observation. Moreover, a recent study by Boccara et al,¹⁸ analyzing 1442 cases, also showed that smoking was not a risk factor.

This insight can potentially influence preoperative management strategies and warrants further exploration in future studies.

Incorporating AI into preoperative planning has the potential to significantly enhance patient care. By identifying patients at higher risk, surgeons can customize preoperative education and postoperative management strategies, delivering personalized care. This approach directly improves patient satisfaction and outcomes. For example, the model's ability to visualize the likelihood of complications based on clinical features, such as resected specimen weight, enables surgeons to have more informed discussions with patients and make decisions about limiting the amount of tissue resected during surgery. Additionally, in cases where the resected tissue weight indicates a high probability of complications, the surgeon can proactively extend the patient's hospitalization for closer monitoring.

Furthermore, this tool plays a crucial role in helping patients understand their individual risk profiles, fostering more transparent communication, and enabling better-informed decisions about their surgery.

Furthermore, from an economic standpoint, reducing postoperative complications could lead to more efficient resource utilization and reduced healthcare costs. Complications after surgery often result in extended hospital stays, additional treatments, and, in some cases, further surgical interventions. By using AI to predict and possibly prevent complications, healthcare systems can minimize these costly outcomes. This benefit extends to patients, reducing the financial burden associated with unexpected postoperative issues.

An additional advantage of AI is its adaptability. Unlike traditional models, AI systems can be retrained as new data becomes available, allowing for continuous improvement in predictive accuracy. This ensures that the model remains relevant as clinical practices and patient populations evolve, positioning AI as a flexible, long-term asset in surgical planning and risk assessment. This is in line with our future research involving additional centers with large datasets.

However, as with any new technology, challenges remain. One major concern is the interpretability of AI models, often described as “black-box” algorithms. Clinicians may hesitate to trust a model whose decision-making process is not fully transparent, particularly when critical surgical decisions are influenced by outputs that may not be easily understood. This concern stems from the inherent difficulty in relying on a system that processes complex interactions between numerous variables, which may seem abstract compared with traditional, more intuitive decision-making approaches. To address this, our study emphasizes the importance of providing tools that enable clinicians to visualize and understand the impact of individual features on the model's predictions. Because the primary dependable variable in the procedure is the reduction weight, we have plotted the complication probability as a function of specimen weight. The plot with the patient’s preoperative features allows a clear model interpretation, as seen in Figure 2. Providing this level of interpretability is crucial to fostering trust in AI-based tools and ensuring that clinicians feel empowered, rather than replaced, by the model's insights.

Figure 2.

Probability of major complication as a function of specimen weight reduction for a specific case. The plot aids surgeons in identifying ranges where the risk is lower and potentially adjusting surgical goals accordingly.

Open in new tabDownload slide

Beyond concerns about interpretability, ethical issues may emerge when AI-generated recommendations are employed by third parties, such as health insurance companies, to allocate resources based on cost-effectiveness rather than medical necessity. For instance, an AI model may identify candidates for breast reduction surgery who are likely to have favorable outcomes at lower costs, potentially deprioritizing patients who may derive significant benefits but carry a higher risk of complications. This raises critical questions about fairness and equity in healthcare decisions. To prevent such biases, the use of predictive models must be guided by ethical principles that emphasize patient well-being and clinical need over financial considerations. It is crucial to ensure that clinicians remain at the forefront of decision making, with AI tools serving to support rather than replace their judgment, thereby promoting transparency and equitable access to care.

Although our study yielded promising results, several limitations must be addressed before presenting a ready-to-use product for plastic surgeons. First, our model was developed using retrospective data from a single center, which may restrict its applicability to broader populations or institutions with varying patient demographics and surgical practices. To enhance the model's robustness and reliability, external validation involving data from multiple centers with diverse settings is essential. Additionally, our current analysis focuses on predicting short-term complications. Future research should include a comprehensive assessment of the model's performance in predicting long-term outcomes. We also aim to refine the model's interpretability, develop a user-friendly interface, and conduct prospective studies to confirm its real-time clinical utility. These steps are crucial to ensure the tool is both accurate and practical for integration into routine surgical workflows.

We would also like to address our approach to defining complications in this study. We focused on developing the classifier model specifically for major complications requiring inpatient care, as these significantly impact healthcare resource allocation. Although we acknowledge that minor complications can affect patient well-being, they generally do not necessitate changes to the surgical plan. By concentrating on major complications, our model targets the outcomes most likely to influence clinical decision making and the allocation of postoperative resources.

The integration of AI into the clinical workflow for breast reduction surgery represents a significant advancement in personalized medicine. Our AI-driven model provides a means for surgeons to better predict complications, optimize patient care, and reduce healthcare costs. Because AI technology continues to evolve and its applications broaden, its role in supporting clinical decisions will undoubtedly expand, offering new opportunities to advance the field of plastic surgery and improve patient outcomes.

CONCLUSIONS

To our knowledge, this study is the first to apply AI to predict postoperative complications in breast reduction surgery up to 30 days following surgery, marking a significant advancement in personalized patient care. We have built an ML-based classifier model with outstanding performance, achieving an AUC-ROC of 0.83 with an NPV of 0.95. In addition, we developed an interpretability tool that allows surgeons to visualize the probability of complications based on clinical features as a function of specimen weight, enhancing preoperative counseling and patient education. This AI-driven approach offers a powerful tool for improving surgical planning and outcomes, although further validation in diverse populations is needed.

Disclosures

Dr Barnea participates in educational lectures for Mentor Worldwide LLC (Irvine, CA), a Johnson & Johnson (New Brunswick, NJ) company. The remaining authors declared no potential conflicts of interest with respect to the research, authorship, and publication of this article.

Funding

The authors received no financial support for the research, authorship, and publication of this article.

REFERENCES

Author notes