Precision perioperative AI: from signals, images, and records to applications in anesthesia-a narrative mini-review...

Artificial intelligence (AI) is increasingly positioned as an assistive and decision-support layer across the perioperative pathway, transforming heterogeneous clinical data into patient-specific risk estimates and management recommendations. Yet perioperative AI remains conceptually fragmented: terms such as "precision," "real-time," and "multimodal integration" are frequently invoked without operational definition, and many studies still prioritize retrospective discrimination over calibration, workflow integration, and prospective clinical impact. This mini-review provides operational definitions for precision perioperative AI, real-time inference, and multimodal integration within the specific constraints of perioperative care, then synthesizes representative applications across the preoperative, intraoperative, and postoperative phases, emphasizing perioperative-specific evidence and implementation maturity. Preoperatively, machine-learning models trained on electronic health records and multimodal data can improve individualized risk stratification, supporting triage, shared decision-making, and tailored prehabilitation or monitoring strategies. Intraoperatively, waveform-based early-warning systems can reduce the duration and severity of hypotension when embedded in treatment protocols; reinforcement-learning approaches and closed-loop controllers are being explored for anesthetic depth and hemodynamic control. Computer vision applications include support for ultrasound-guided regional anesthesia and operating-room scene analysis. Postoperatively, AI-enhanced surveillance combines continuous monitoring with perioperative risk profiles to detect deterioration and forecast complications such as sepsis, acute kidney injury, and delirium. We argue that perioperative AI must be evaluated as a clinical intervention rather than a static classifier. Deployment-grade requirements include robust calibration, external validation, decision-curve analysis, human-in-the-loop design, drift detection, and structured lifecycle oversight ("algorithmovigilance"), aligned with emerging regulatory expectations and real-world perioperative workflows.