Vital Trace: Protocol-Constrained Patient-State Reasoning for Longitudinal Clinical Trajectories

Researchers present Vital Trace, a protocol-constrained multi-agent AI framework designed to improve clinical risk prediction in intensive care units by tracking patient trajectories over extended periods. The system uses compact patient-state memory and structured reasoning agents rather than unbounded text histories, demonstrating better temporal consistency and interpretability on MIMIC-IV and eICU datasets.

Vital Trace addresses a fundamental limitation in applying large language models to clinical settings: the degradation of reasoning quality and stability when processing lengthy patient histories. Traditional LLM-based clinical systems struggle with context drift and mounting inference costs as patient timelines extend, forcing unreliable serialization of complex medical data. This research introduces architectural constraints—a Global Protocol with physiological state-transition rules and a dynamic patient-state representation—that guide four coordinated agents (Router, Reasoner, Auditor, Steward) through structured decision-making. The framework tracks hemodynamic, respiratory, renal, metabolic, and inflammatory markers, enabling nuanced risk assessment for vasopressor, respiratory, and renal support prediction.

The broader context reflects growing recognition that raw LLM capability alone cannot substitute for domain-specific reasoning in healthcare. Clinical AI requires both accuracy and interpretability, as physician adoption depends on understanding system logic. Vital Trace's empirical validation on real ICU datasets demonstrates that constrained reasoning actually improves performance while reducing computational burden—a crucial insight for resource-limited healthcare settings.

For healthcare technologists and AI researchers, this work validates the principle that explicit protocols and structured agent coordination outperform free-form language-based reasoning in specialized domains. The approach generalizes beyond ICU settings to other longitudinal clinical decision-making scenarios. Healthcare IT vendors and hospital systems evaluating clinical AI solutions should prioritize interpretable, protocol-grounded architectures over black-box LLM applications. Future development should focus on automating protocol generation and adapting frameworks across different patient populations and care environments.

• →Vital Trace uses structured multi-agent reasoning with explicit physiological protocols instead of unbounded text histories to improve clinical predictions
• →The framework achieves better temporal consistency, calibration, and interpretability while reducing inference costs compared to free-form LLM approaches
• →Compact patient-state memory tracking five physiological dimensions enables stable reasoning over extended ICU trajectories
• →Empirical validation on MIMIC-IV and eICU datasets demonstrates superior performance on vasopressor, respiratory, renal support, and deterioration predictions
• →Protocol-constrained design proves that explicit domain rules improve both AI reliability and physician interpretability in clinical applications