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

Area of Science:

• Artificial Intelligence in Medicine
• Anesthesiology
• Machine Learning for Healthcare

Background:

• Orthopedic anesthesia requires real-time integration of complex physiological data, interventions, and patient context.
• Existing methods struggle to proactively manage risks like hypotension, nausea, vomiting, and pain.

Purpose of the Study:

• To introduce Ortho PeriFT, a multimodal transformer model for enhanced orthopedic anesthesia.
• To integrate perioperative prediction, therapeutic recommendations, and continuous monitoring with calibrated uncertainty.

Main Methods:

• Utilized a hierarchical, safety-aware, and interpretable transformer architecture.
• Processed diverse data including second-level waveforms, minute-level numerical data, medication/event tokens, clinical text, and imaging prompts.
• Employed self-supervised pre-training and multitask fine-tuning, with a constrained Decision Transformer for treatment suggestions.

Main Results:

• Ortho PeriFT outperformed classical and neural baselines in discrimination and precision-recall for key outcomes.
• Demonstrated reduced calibration error, negative log-likelihood, and maintained narrow uncertainty bands.
• Showed earlier warnings in streaming analyses and generalized performance across orthopedic subtypes and demographics.

Conclusions:

• A hierarchical, safety-aware, and interpretable transformer framework can provide accurate risk estimates and actionable therapeutic suggestions in orthopedic anesthesia.
• Ortho PeriFT offers timely alerts and case-based rationales aligned with clinical reasoning, enhancing patient safety.