Clinical decision support systems face a fundamental challenge: they must reason over heterogeneous and longitudinal patient information, not just answer isolated medical questions. According to a preprint on arXiv by Aueawatthanaphisut and Aueaphum, current medical large language models (LLMs) and retrieval-augmented generation (RAG) systems often rely on single-step prompting or retrieval, which is fragile when clinical evidence is distributed across long electronic health records (EHRs), medical images, sensor streams, guidelines, and referral constraints.
To address this, the authors propose MedRLM, a Recursive Multimodal Health Intelligence framework for long-context clinical reasoning, sensor-guided screening, evidence-grounded decision support, and community-to-tertiary referral support. Instead of compressing all patient information into one prompt, MedRLM treats the patient case as an external clinical environment that can be recursively inspected, decomposed, retrieved, verified, and synthesized.
Framework Architecture
MedRLM coordinates specialized agents for different data modalities and tasks:
| Agent | Function |
|---|---|
| Clinical Text | Processes clinical notes and reports |
| Longitudinal EHR | Handles temporal electronic health records |
| Medical Imaging | Analyzes radiology images |
| Physiological Sensor Signals | Interprets sensor streams (e.g., ECG) |
| Guideline Retrieval | Fetches relevant clinical guidelines |
| Uncertainty Auditing | Assesses confidence in predictions |
| Referral Planning | Optimizes community-to-tertiary referrals |
The framework introduces a Clinical Evidence Graph Memory to connect patient-specific observations with retrieved evidence, standardized definitions, sensor-derived biomarkers, and referral criteria. This graph-based memory enables the system to maintain context across multiple reasoning steps and information sources.
Recursive and Sensor-Guided Reasoning
A key innovation is the sensor-guided recursive triggering mechanism, which activates deeper reasoning when abnormal physiological or behavioral patterns are detected. This allows the system to dynamically allocate computational resources to complex cases while maintaining efficiency for routine ones. Additionally, uncertainty-gated refinement supports clinician review for high-risk or low-confidence cases, ensuring that the human expert remains in the loop when the AI is uncertain.
Evaluation Design
The authors outline a real-data evaluation design using public and credentialed clinical datasets spanning EHR, radiology, ECG, ICU time series, and referral-proxy outcomes. This multi-dataset approach aims to validate MedRLM across diverse clinical scenarios, from primary care screening to tertiary hospital referrals.
Implications for Enterprise AI
While MedRLM is specifically designed for healthcare, its architecture – recursive multi-agent coordination, graph-based memory, and sensor-guided triggering – offers a template for complex decision support in other domains. Enterprise technology leaders evaluating AI for high-stakes applications may find parallels in supply chain anomaly detection, trade compliance auditing, or logistics troubleshooting, where heterogeneous data sources and long-context reasoning are equally critical. The framework's emphasis on auditability and uncertainty quantification aligns with growing requirements for explainable AI in regulated industries.
Moving Beyond Static QA
MedRLM aims to move medical AI from static question answering toward auditable, multimodal, and workflow-aware clinical decision support, according to the authors. By combining recursive reasoning with evidence grounding and sensor integration, the framework represents a step toward AI systems that can handle the complexity of real-world clinical workflows.