TheraAgent: Multi-Agent Framework with Self-Evolving Memory and Evidence-Calibrated Reasoning for PET Theranostics

LLMs need specialized training for rare medical applications: TheraAgent was developed because general large language models lack sufficient knowledge about 177Lu-PSMA radioligand therapy, which was only FDA-approved in 2022, demonstrating that AI systems require domain-specific adaptation for emerging treatments rather than relying on general training data.

AI systems must integrate multiple data types to make reliable clinical predictions: The framework combines information from medical imaging (PET/CT scans), laboratory tests, and clinical notes through specialized "expert" components, highlighting that responsible medical AI cannot rely on single data sources and must handle the messy reality of how clinical information is documented.

Clinical AI predictions must be grounded in trial evidence, not AI confidence: TheraAgent includes an "Evidence-Calibrated Reasoning" component that anchors predictions in results from the VISION and TheraP clinical trials, addressing a critical concern that AI systems can produce confident-sounding but unfounded recommendations—a safeguard essential for responsible deployment in healthcare.

AI systems should learn and improve from accumulated experience: The "Self-Evolving Agentic Memory" component allows the system to identify patterns from previous cases, enabling better predictions with limited training data—a responsible approach that acknowledges both the scarcity of data for rare treatments and the potential for AI to become more reliable over time with proper oversight.

Uncertainty quantification matters in medical AI: The framework uses "confidence-weighted consensus" from multiple expert components, meaning the system explicitly measures and communicates uncertainty rather than presenting all predictions with equal confidence.