Diagnóstico multimodal da doença de parkinson com arquitetura agentica colaborativa baseada em internet de modelos de linguagem médica

Abstract

This thesis proposes an intelligent platform to support the diagnosis of Parkinson’s Disease (PD), integrating advanced technologies such as the Internet of Medical Things (IoMT), Multimodal Retrieval-Augmented Generation (RAG), efficient

model compression, and medical language model operations (LLMOps and VLMOps). Unlike previous approaches that rely on isolated models or single data modalities, this work defines a clinical agentive ecosystem composed of autonomous and collaborative agents powered by Large or Small Language Models (LLMs/SLMs), with reinforcement learning (RL). The system architecture is designed to be adaptable, explainable, and patient-centered, with an emphasis on reliability, fast response times, and practical deployment in real clinical environments. Heterogeneous data modalities, including clinical text, medical imaging, sensor signals, and voice recordings, are fused through advanced RAG variants such as Multimodal RAG, GraphRAG, and Agentic RAG, enabling contextualized inference without the need for extensive fine-tuning. The methodology includes evaluating computational cost and operational efficiency in simulated production environments, using metrics such as GPU utilization, LLM call cost, and infrastructure scalability. Compression techniques such as 4-bit quantization and QLoRA are employed to enable robust models to run on edge devices. Finally, intelligent agents refine their inference strategies through post-training with reinforcement learning, using rewards derived from clinical analyses to improve both the accuracy and practical applicability of the system. Experimental results showed competitive performance across classifiers, with Gradient Boosting achieving 83.85% accuracy and an F1-score of 88.74%, and HistGradient Boosting reaching 82.92% accuracy, an F1-score of 88.12%, and an AUC-ROC of 0.909. While models such as SVM and KNN demonstrated higher sensitivity in some cases, they also showed a greater rate of false positives, highlighting the superior balance of the proposed approach. This work stands out for its capacity to integrate with real clinical workflows and holds strong potential to serve as a reference model for neurointelligent platforms in connected digital healthcare.