Leveraging transformer-based artificial intelligence for enhanced anesthetic decision-making in orthopedic surgery

Orthopedic anesthesia necessitates the real-time integration of rapidly changing physiological states, specific interventions, and contextual narratives to avert hypotension, postoperative nausea and vomiting, and uncontrolled pain. We present Ortho PeriFT, a multimodal transformer that integrates perioperative prediction, therapeutic recommendations, and continuous monitoring with a calibrated uncertainty. The encoder was designed to align with clinical time scales by processing second-level waveform patches and minute-level numerical data, alongside medication and event tokens and extended clinical text, with optional prompts from preoperative imaging. Self-supervised pre-training on extensive clinical time series was followed by multitask fine-tuning for the primary endpoints. A constrained Decision Transformer suggests titrations of fluids, vasopressors, and anesthetic depth under guideline-aware action masks, with all outputs accompanied by conformal risk intervals to allow abstention when confidence is low. Across both internal and external cohorts, Ortho PeriFT enhanced discrimination and precision-recall for all primary outcomes compared to robust classical and neural baselines, reduced calibration error and negative log-likelihood, and maintained narrow uncertainty bands. Off-policy estimators indicate higher counterfactual utility than clinician behavior and behavior cloning while ensuring zero-guardrail violations. Streaming analyses demonstrated earlier warnings at matched false alarm rates, and performance generalized across orthopedic subtypes with stable calibration across demographic strata. Attribution maps and prototype trajectory retrieval offer case-based rationales that are aligned with clinical reasoning. These findings illustrate that a hierarchical, safety-aware, and interpretable transformer can provide accurate risk estimates, actionable therapeutic suggestions, and timely alerts for orthopedic anesthesia within a unified framework.