Artificial intelligence in onco-anaesthesia: current applications, challenges, and future directions

The review frames onco-anaesthesia as shifting from reactive physiological management toward predictive and precision-based care, and organizes existing evidence along the perioperative cancer pathway. Rather than discussing technologies or anaesthetic stages in isolation, it integrates preoperative, intraoperative, and postoperative AI applications into a single continuous pathway for cancer patients. This is a narrative-review level synthesis, expressed in the text as outlining "current AI applications across the perioperative cancer pathway," without quantitative pooling.

Preoperatively, machine learning and deep learning models are used for automated frailty assessment, electronic health record phenotyping, and prediction of cancer-specific outcomes, thereby enhancing risk stratification. It extends the information sources for risk stratification beyond conventional scoring toward automated frailty assessment and electronic health record phenotyping, aimed at cancer-specific outcome prediction. Stated in the text as models that "enhance risk stratification through automated frailty assessment, electronic health record phenotyping, and prediction of cancer-specific outcomes," summarizing existing application directions.

Intraoperatively, closed-loop anaesthesia delivery systems, predictive haemodynamic monitoring, and automated depth-of-anaesthesia control are used to optimize drug dosing and reduce physiological stress, and may help preserve perioperative immune function, with potential implications for long-term oncologic outcomes. It extends the value of intraoperative AI from drug delivery and physiological stability toward perioperative immune preservation and longer-term oncologic outcomes. The text uses hedged language such as "may help preserve" and "potential implications," indicating these links are framed as possibilities rather than established conclusions.

Postoperatively, AI-driven integration of multimodal data (genomics, radiomics, wearable biosignals, high-resolution physiological waveforms) facilitates early detection of complications such as delirium, persistent pain, acute kidney injury, and anastomotic leakage; the review also discusses AI's role in evaluating the "onco-anaesthesia hypothesis." It places complication detection alongside the mechanistic question of anaesthetic technique, inflammation, and cancer recurrence, discussing postoperative monitoring and oncologic mechanism research within one framework. Expressed in the text as "facilitates early detection" and as examining "the role of AI in evaluating the 'onco-anaesthesia hypothesis'," a directional synthesis without reported effect sizes.

This review is aimed at researchers and clinical teams in onco-anaesthesia and perioperative medicine, and is suited to understanding where AI sits across preoperative risk stratification, intraoperative delivery and monitoring, and postoperative complication detection, as well as the framing of the mechanistic "onco-anaesthesia hypothesis." The future directions it proposes—explainable AI, federated learning, real-time clinical decision-support systems, and large prospective validation—define the conditions and settings under which these applications would move from current exploration toward routine clinical use.

Readers may still wish to watch: the specific study designs, populations, and validation approaches behind each application mentioned; the links between closed-loop delivery, predictive haemodynamic monitoring, and automated depth-of-anaesthesia control and long-term oncologic outcomes, which are framed as "may" and "potential" and whose direction and magnitude await prospective confirmation; how the challenges of data heterogeneity, limited generalisability, algorithmic opacity, regulatory uncertainty, and equity are elaborated in the original; and how complete the evidence around the "onco-anaesthesia hypothesis" is. Because this is a fast abstract-level parse without figures or reference details, these questions cannot be further verified within the present text.