Artificial intelligence is moving rapidly into the operating theater and the wards that surround it, but a group of anesthesiology researchers argues that the technology will only improve patient care if it is designed around clinicians rather than around raw predictive performance. In an editorial published in the journal Anesthesiology and Perioperative Science, Ruihao Zhou, Tao Zhu and colleagues from the Anesthesiology, Intelligence Research Collaboration Group lay out both the promise and the pitfalls of digitally empowered perioperative medicine, and they call for a deliberate shift toward what they describe as human-centered AI.
The authors’ starting point is a striking paradox. Digital health technologies and AI are profoundly transforming perioperative medicine, with applications now spanning preoperative risk stratification, intraoperative clinical decision support, and postoperative remote monitoring. Yet an increasingly evident translational gap persists: many high-precision technologies demonstrate excellent performance in data-driven studies, while their adoption and impact in real-world clinical practice remain substantially below expectations. The editorial contends that the advancement of perioperative intelligence should not merely pursue the accumulation of AI technologies but should also focus on developing human-centered systems with the potential to alleviate clinicians’ cognitive load, augment rather than replace clinical judgment, and better accommodate the inherent complexity of perioperative care.
The most concrete near-term opportunity, according to the authors, lies in anticipation. AI algorithms trained on large-scale datasets can identify patterns that precede perioperative adverse events such as intraoperative hypotension, blood product requirements, postoperative ICU admission, mortality, and surgical site infections. This enables early prediction, giving healthcare teams valuable time for intervention. From a patient safety perspective, the authors frame this as a significant advancement: clinicians are no longer limited to reactive responses but can move toward prophylactic care, reducing delays in the recognition and management of adverse events. In other words, the operating room of the near future may warn its team minutes before a patient’s blood pressure begins to slide, rather than waiting for a human to notice the trend on a crowded monitor.
Algorithmic approaches have evolved from conventional statistical modeling to encompass machine learning, deep learning, and multimodal fusion, driving continuous enhancement in prediction performance. But the editorial is candid about the unresolved problems. Most studies are based on retrospective, single-center datasets and lack prospective, multicenter validation, raising concerns about generalizability and real-world applicability. Many models remain correlational rather than causal, identifying associations but rarely elucidating truly modifiable contributors of outcomes. Predictive performance may also deteriorate over time because of data drift and concept drift, potentially compromising model reliability in evolving clinical environments. Moreover, the authors note, most AI research teams have not yet reported the clinical deployment of their published models, and AI-based tools are unlikely to improve patient care unless they are effectively integrated into existing electronic health record systems and clinical workflows.
Beyond risk prediction and decision support, the editorial examines AI-based closed-loop anesthesia systems, which are expected to improve responsiveness and reduce the cognitive workload of anesthesiologists. These technologies have the potential to automate repetitive tasks, including anesthetic titration and administrative workflows, thereby improving efficiency and reducing clinician workload. Although AI-driven closed-loop anesthesia systems can offer stability, the authors are firm that they cannot replace human dimensions or competencies such as ethical responsibility, clinical intuition, situational awareness, and accountability. A machine may hold a patient’s anesthetic depth steady, the argument goes, but responsibility for that patient cannot be delegated to an algorithm.
The editorial also highlights generative AI, specifically large language models, as a potential facilitator of perioperative intelligent anesthesia management. Owing to their reasoning capabilities, multimodal information processing, and long-context understanding, LLMs can support a wide range of perioperative functions, from risk prediction and clinical decision support to precision interventions and workflow optimization. Yet the authors temper this enthusiasm with a sobering observation: despite the continuous development of AI in the medical field, rigorous patient-centered empirical evidence demonstrating its clinical benefits remains scarce. They call for prospective randomized clinical trials to study the effects of human-AI collaboration before clinical adoption, a notable demand in a field where most published work still stops at retrospective model validation.
Looking ahead, the authors sketch a pathway for responsible implementation. Future research should define a clear pathway for clinical implementation from the earliest stages of development, foster close collaboration between clinicians and AI engineers through co-design approaches, and place greater emphasis on implementation science evaluations of real-world deployment rather than pursuing technical complexity for its own sake. Intelligent tools must be rigorously validated across diverse surgical populations and clinical scenarios to ensure fairness and reliability. Crucially, these technologies should augment, rather than replace, the clinical judgment of experienced teams while remaining human-centered. Nevertheless, the authors see the trajectory as clear: AI is transforming perioperative care from an intuition-based decision model to a data-driven one, thereby enhancing safety, consistency, and clinical outcomes.
None of this will work without better data, and here the editorial identifies a structural weakness. Perioperative intelligence generates massive multi-modal data characterized by the classical four Vs of volume, variety, velocity, and value, and this data is multi-dimensional and heterogeneous. However, compared with many databases in the critical care field, such as MIMIC-IV and the eICU database, high-quality perioperative databases and cohorts are still scarce. The authors point to efforts by researchers in China to build high-quality perioperative patient-specific cohorts, including the China surgery and anesthesia cohort, known as CSAC, and the PeRiOperative sTress risk assEssment and Clinical decision cohorT, known as PROTECT. Future research, they argue, should strengthen the collection and processing of perioperative information, including preoperative and intraoperative time-series data as well as postoperative follow-up outcomes, and real-world data development will require robust, high-quality multicenter datasets, effective data sharing frameworks, and standardized data formats.
The ethical dimension receives equal weight. The authors argue that future research in intelligent perioperative medicine should be guided by the World Health Organization’s guidance, Ethics and Governance of Artificial Intelligence for Health, which provides a comprehensive framework centered around six consensus principles: protecting human autonomy; promoting human well-being, safety, and public interest; ensuring transparency, explainability, and intelligibility; fostering responsibility and accountability; ensuring inclusiveness and equity; and promoting responsive and sustainable AI systems. For AI aimed at improving patients’ perioperative safety, the editorial insists, these principles are not merely aspirational ideals but practical requirements for implementation. AI should serve as a tool to facilitate clinical decision-making, and the human-centered approach will not only improve clinical outcomes but also reframe the ultimate goal of healthcare delivery within the complex perioperative environment.
The editorial closes with a human factors agenda and a workforce warning. The actual effectiveness of AI depends not only on its technical performance but also on the quality of human-AI interaction and the appropriate interpretation of its outputs. The probabilistic nature of AI outputs, limited interpretability, adaptive behavior, risk of misclassification, and capacity to generate large volumes of data may introduce unique human factors challenges in perioperative care. Drawing on human factors analysis, the authors list seven considerations for building human-centered AI: defining intended users and excluded populations in advance, designing outputs that are clinically trustworthy and interpretable, prioritizing integration into clinical workflows, providing training to prevent skill degradation, ensuring safe backup pathways, monitoring real-world deployment scenarios, and clearly communicating updates. Incorporating AI-specific human factors considerations into routine regulatory and health technology assessment processes, they argue, can reduce preventable harm while supporting innovation. Future progress will depend on strong interdisciplinary collaboration among clinicians, computer scientists, data scientists, and software engineers, and educational reform is equally critical: future anesthesiologists must be proficient not only in physiology and pharmacology but also in data literacy, algorithmic reasoning, and human-AI collaboration. Societal acceptance of AI-assisted perioperative care, the authors conclude, depends on transparency, accountability, and demonstrable patient benefit.
Subject of Research: Human-centered artificial intelligence for digitally empowered perioperative anesthesia care
Article Title: Digitally and intelligently empowered perioperative care: a shift toward perioperative human-centered AI
Article References: Zhou, R., Anwar, S., Luo, Z., Peng, X., Wang, Y., Li, K., Chen, G., Hao, X., Zhu, T., the Anesthesiology, Intelligence Research Collaboration Group (AIR-CG), Shu, H., Wang, C., Kui, W., Li, P., Ning, S., Yu, J., Duan, J., Cai, J., Zhu, Y., … Gu, Y. (2026). Digitally and intelligently empowered perioperative care: a shift toward perioperative human-centered AI. Anesthesiology and Perioperative Science, 4(1), Article 35. https://doi.org/10.1007/s44254-026-00183-0
Image Credits: AI Generated
DOI: 10.1007/s44254-026-00183-0
Keywords: perioperative medicine, artificial intelligence, anesthesiology, machine learning, closed-loop anesthesia, large language models, clinical decision support, patient safety, human factors, data drift, WHO ethics guidance, health informatics
News Source: Ophelia Keating. (November 4, 2026). Perioperative AI needs to put humans first, anesthesiologists argue. Scienmag.
