Can AI Predict Surgical Complications Before Operations?

By Rasit Dinc

Introduction

Surgery, while often life-saving, carries inherent risks. Postoperative complications are a significant cause of morbidity and mortality worldwide, with an estimated 4.2 million people dying within 30 days of surgery each year [1]. For healthcare professionals, the ability to accurately predict which patients are at the highest risk is a cornerstone of effective perioperative care. Traditional risk assessment tools, such as the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) Surgical Risk Calculator, have been invaluable. However, these models often rely on static, population-level data and may not fully capture the nuanced risk profile of an individual patient [2].

In recent years, the convergence of big data in healthcare and sophisticated computational power has ushered in a new era of predictive medicine, driven by artificial intelligence (AI) and machine learning (ML). These technologies offer the potential to move beyond generalized risk scores to a more personalized, precise, and proactive approach to patient safety. This article explores the growing body of evidence on the use of AI to predict surgical complications before an operation even begins, examining the methodologies, their effectiveness, and the implications for the future of surgical care.

The Shortcomings of Traditional Risk Stratification

For decades, clinicians have relied on scoring systems to estimate a patient's risk of adverse postoperative outcomes. These tools typically use a limited set of preoperative variables—such as age, comorbidities, and the type of surgery—to generate a risk percentage. While useful for general guidance, their predictive accuracy can be limited when applied to diverse patient populations and unique surgical scenarios. Studies have shown that the performance of these calculators can decrease when used at different institutions or for specific, complex procedures, highlighting a gap between population-based estimates and individual patient reality [2].

These models are often static, meaning they are not designed to be dynamically updated with new patient information as it becomes available. The complexity of human physiology and the multitude of factors that can influence surgical outcomes often exceed the capacity of these linear, rule-based systems. This is the challenge that AI is uniquely positioned to address.

How AI Models Are Revolutionizing Preoperative Prediction

Unlike traditional calculators, AI and machine learning models can analyze vast and complex datasets from electronic health records (EHRs) to identify subtle patterns that are invisible to human observers. A typical ML model for surgical risk prediction might be trained on hundreds of thousands of patient encounters, incorporating a wide array of preoperative data points, including:

• Patient Demographics: Age, sex, body mass index (BMI).
• Comorbidities: History of congestive heart failure, diabetes, renal disease, etc.
• Laboratory Results: Values such as brain natriuretic peptide (BNP), aspartate aminotransferase (AST), and red cell distribution width (RDW).
• Procedural Data: The planned surgical procedure and its complexity.

Using algorithms like gradient-boosted decision trees, these models learn the intricate, non-linear relationships between these variables and the likelihood of specific adverse events, such as mortality, major adverse cardiac and cerebrovascular events (MACCEs), or acute kidney injury [2]. The result is a highly personalized risk score that reflects a patient's specific clinical profile.

The Evidence: AI's Superior Predictive Power

A growing body of research demonstrates that AI-powered models can outperform traditional risk calculators. A large-scale prognostic study published in JAMA Network Open developed and validated an ML model on over 1.4 million patients. The model demonstrated exceptional accuracy in predicting 30-day mortality, achieving an area under the receiver operating characteristic curve (AUROC) of 0.956 in a prospective validation set. For the combined outcome of mortality and MACCE, the AUROC was 0.899 [2].

Crucially, when compared directly with the widely used NSQIP calculator, the machine learning model showed superior performance, with a significantly higher AUROC (0.945 vs. 0.897) and better specificity. This indicates that the AI model was not only more accurate overall but also better at correctly identifying patients who were not at high risk, potentially reducing unnecessary interventions and patient anxiety.

Other studies have corroborated these findings. Research from institutions like Johns Hopkins University and Washington University has highlighted the development of AI models that can predict a range of complications, from pneumonia and blood clots to sepsis, often using unstructured data like clinical notes to further enhance their predictive power [3, 4].

Implications for Clinical Practice and Future Directions

The ability to accurately identify high-risk patients before they enter the operating room has profound implications for clinical practice. It allows for the implementation of a "precision perioperative" approach, where care can be tailored to an individual's predicted risk. This could involve:

• Preoperative Optimization: Initiating targeted interventions, such as nutritional support, "prehabilitation" exercise programs, or more intensive management of comorbidities.
• Informed Decision-Making: Facilitating more detailed and personalized conversations with patients and their families about the risks and benefits of surgery.
• Resource Allocation: Ensuring that high-risk patients receive heightened monitoring and are scheduled for appropriate levels of postoperative care, such as an intensive care unit (ICU) bed.

While the promise is immense, challenges remain. The "black box" nature of some complex algorithms requires further development of interpretable AI (XAI) methods, like the Shapley additive explanations (SHAP) used in the JAMA study, to help clinicians understand why a model has flagged a patient as high-risk [2]. Furthermore, models must be continuously monitored and validated to ensure they remain accurate as clinical practices and patient populations evolve.

In conclusion, the evidence strongly suggests that AI can indeed predict surgical complications with a degree of accuracy that surpasses traditional methods. As these tools become more integrated into clinical workflows, they stand to become an indispensable part of modern surgical care, empowering healthcare professionals to anticipate and mitigate risk, ultimately leading to safer surgery and better patient outcomes.

References

[1] Weiser, T. G., Haynes, A. B., & Molina, G. (2016). Estimate of the global volume of surgery in 2012: an assessment supporting improved health outcomes. The Lancet, 388(10040), 11.

[2] Mahajan, A., Esper, S., Oo, T. H., et al. (2023). Development and Validation of a Machine Learning Model to Identify Patients Before Surgery at High Risk for Postoperative Adverse Events. JAMA Network Open, 6(7), e2322285.

[3] Washington University School of Medicine. (2025). Foundation AI model predicts postoperative risks from clinical notes. Source. Retrieved from https://source.washu.edu/2025/03/foundation-ai-model-predicts-postoperative-risks-from-clinical-notes/

[4] Bronnert, R., Besch, G., Hild, O., Lihoreau, T., Chaussy, Y., & Ferreira, D. (2025). Performance of artificial intelligence models for predicting intraoperative complications during surgery in real time: a systematic review and meta-analysis protocol. BMJ Open, 15(10), e106204.