To develop a multimodal transformer, Ortho PeriFT, for real-time prediction, therapeutic recommendations, and continuous monitoring in orthopedic anesthesia.
Approach:
Key Findings:
Ortho PeriFT enhanced discrimination and precision-recall for primary outcomes compared to classical and neural baselines.
The model reduced calibration error and negative log-likelihood while maintaining narrow uncertainty bands.
Streaming analyses provided earlier warnings at matched false alarm rates across orthopedic subtypes.
Interpretation:
Limitations:
External validity across institutions is inconsistent.
Many existing systems prioritize discrimination over calibration, uncertainty, clinical utility, and fairness.
