Intraoperative complications contribute significantly to morbidity and mortality, and reducing their risk is a primary objective for all operating room’s healthcare professionals. Many of these complications are predictable and could be anticipated by the surgeon or anaesthesiologist. Various clinical scores were developed to assess cardiovascular risk, acute kidney injury or acute respiratory failure preoperatively. However, these scores require time for calculation and are not designed to be adjusted in real time during surgery, based on physiological signals and new intraoperative events. Besides, some events remain unpredictable because they are multifactorial.

In recent decades, Artificial Intelligence (AI)-based algorithms have been tested for the real-time prediction of intraoperative complications. These algorithms have the potential to continuously analyse patient data and provide early warnings, enabling professionals to intervene more effectively.

The aim of this review is to address the question: ‘What is the performance of AI models in predicting intraoperative complications during surgery using baseline and real-time data?’.

The review will follow the Transparent Reporting of multivariable prediction models for Individual Prognosis or Diagnosis: Checklist for Systematic Reviews and Meta-Analyses and BMJ guidelines. MEDLINE, Embase, CENTRAL (Cochrane), IEEE Xplore and Google Scholar databases will be explored for peer-reviewed papers up to 25 March 2025. First, two reviewers will independently screen titles, abstracts and full texts based on the inclusion and exclusion criteria. A third reviewer will resolve any disagreements. Eligibility criteria include AI models that predict or forecast intraoperative complications or immediate postoperative complications (up to the stay in the Post-Anaesthesia Care Unit) involving any patient undergoing surgery or interventional procedures with general or locoregional anaesthesia. The primary target is the algorithm’s performance, depending on the choice of the authors. Key items from the CHARMS 2014 checklist will be extracted using a standardised form. Risk of bias assessment will be performed using the PROBAST+AI tool. If possible, meta-analysis will be conducted by implementing a random effects meta-analysis model.

Ethical approval is not required. The results will be published in a peer-reviewed journal and presented at national and international conferences.

PROSPERO registration number: CRD420250599920. Any future amendments will be updated in the PROSPERO record.