LeJEPA + I-JEPA: A World-Model-Grounded Multi-Agent Framework for Medical AGI

Abstract

Executive Summary: LeJEPA + I-JEPA Medical AGI LEJEPA_IJEPA_MEDICAL_AGI is an open-source, multi-agent framework designed to achieve clinically validated medical image analysis. Developed by Sovereign Machine Lab (SOMALA), the system addresses the common "hallucination" issues of standard Large Language Models (LLMs) by grounding clinical reasoning in a structured World Model. FULL CODE Core Architecture The framework separates perception from reasoning through a five-pillar architecture that operates in an iterative convergence loop: Pillar 1: LeJEPA Feature Extractor (Perception): Acts as a World Model layer. It converts raw CT scan data into a discrete set of factual radiological signs, preventing the model from jumping straight to an ungrounded diagnosis. Pillar 2: Image Analysis Agent: Encapsulates the perception stage to ensure a clean interface between visual data and downstream reasoning. Pillars 3 & 4: Validation Agent (Constraint Enforcement): Uses deterministic regular expressions (regex) to check the output against hard clinical requirements. It ensures the report includes essential elements like primary diagnosis, acute interventions, and long-term therapy. Pillar 5: Prompt Engineer Agent (Adaptive Steering): Acts as the system's "executive function." It takes feedback from the Validation Agent to surgically refine the prompts until all clinical constraints are met. The Convergence Loop Rather than a single-shot response, the system uses an AGI-inspired inference process. It generates a report, validates it against expert-defined constraints, and iterates (up to five times) until the report is clinically complete. In a demonstration case involving a complex diagnosis of stercoral colitis, the system successfully converged on a comprehensive, clinically sound report in just three iterations. Clinical Significance & Ethics The framework was tested on a high-risk case of stercoral colitis in a 23-year-old patient with autism spectrum disorder (ASD). This case was chosen for its complexity and high mortality rate if misdiagnosed. The system successfully identified the hallmark triad of the condition and recommended a full management plan, including: Acute endoscopic removal. Long-term pelvic-floor rehabilitation. Surveillance for life-threatening complications like perforation. Open-Source Philosophy A primary contribution of this work is its commitment to Medical Ethics through open access: Auditability: Every step—from the World Model state to the validation rules—is human-readable and inspectable. Equity: The system is designed to run on a single consumer-grade GPU (such as through Google Colab), making it accessible to rural or low-resource clinics worldwide without proprietary fees. Reproducibility: The entire stack is available on GitHub, allowing the global research community to verify and extend the technology. Future Directions While currently a research demonstration, future goals include fine-tuning dedicated I-JEPA encoders on large-scale medical datasets, integrating longitudinal EHR data for a "whole-patient" world state, and conducting prospective clinical trials to compare the system against standard radiological workflows.

Ähnliche Arbeiten