Evaluating the Application of Deep Learning-Based Artificial Intelligence Models in Predicting Hemodynamic and Respiratory Complications During Surgery: A Systematic Review

Background: Artificial intelligence (AI) models based on deep learning are progressively utilized in surgical environments to anticipate intraoperative problems, including hemodynamic instability and respiratory distress. These models examine extensive datasets, encompassing physiological factors, to deliver real-time forecasts and improve therapeutic decision-making. This systematic study seeks to assess the efficacy and therapeutic relevance of deep learning models in forecasting hemodynamic and respiratory issues during surgery, emphasizing their precision, dependability, and influence on patient outcomes. Materials and Methods: The review complied with PRISMA guidelines. A systematic search was performed across PubMed, Scopus, Web of Science, and Embase for studies published without time limitation until ۲۰۲۴. This study examines the role of artificial intelligence and deep learning in predicting hemodynamic and respiratory complications during intraoperative procedures. Inclusion criteria consisted of peer-reviewed studies that reported the application of deep learning models for intraoperative monitoring, with an emphasis on hemodynamic or respiratory outcomes. Exclusion criteria comprised studies not conducted in English, animal research, and studies lacking validation datasets. The PROBAST tool was utilized to assess the risk of bias in prediction model studies. Results were synthesized both qualitatively and quantitatively, with pooled accuracy, sensitivity, and specificity calculated when applicable. Results: Of the ۷۴۳ studies identified, ۲۵ fulfilled the inclusion criteria, involving a total of ۳,۸۲۰ surgical cases. Many studies employed convolutional neural networks (CNNs) or recurrent neural networks (RNNs) for predictive modeling. The models demonstrated pooled accuracy between ۸۵% and ۹۵%, with sensitivity and specificity surpassing ۸۰% in the prediction of intraoperative complications, including hypotension and hypoxemia. Models exhibited lower false positive rates in comparison to conventional risk stratification tools (risk ratio: ۰.۷۵, ۹۵% CI: ۰.۶۵-۰.۸۵). The assessment of bias revealed a low risk in ۱۵ studies, a moderate risk in ۸ studies, and a high risk in ۲ studies. Conclusion: Deep learning-based AI models show high potential for accurately predicting hemodynamic and respiratory complications during surgery, offering opportunities to enhance patient safety and optimize decision-making. However, challenges remain regarding model generalizability, integration into clinical workflows, and ethical considerations. Future studies should focus on external validation across diverse surgical populations and development of user-friendly interfaces for real-time application.

کلیدواژه ها:

Artificial intelligence ، deep learning ، hemodynamic complications ، respiratory complications ، intraoperative prediction