Large language models (LLMs) have demonstrated remarkable capabilities across different domains, yet two critical challenges limit their reliability: opaquely generated responses and role drift in extended interactions. Moreover, the possibility of a multi-layered dialogue, shaping LLM behavior both asynchronously across sessions and, optionally, in real time, remains largely unexplored. This paper presents XBot, a conversational agent designed to address these challenges. Built on the GPT-4o API, XBot requires no access to internal weights or activation space, making it portable across different LLM-based systems. It analyzes the user message and decomposes it into chunks, each annotated with topic and sentiment. It then selects a domain validated strategy from a curated expert-defined set, based on the assigned role, spanning multiple levels of granularity from general response types down to topic- and sentiment-specific guidance. Each response is therefore accompanied by explicit justifications accessible to non-technical users, supporting a continuous three-way dialogue among users, domain experts and XBot for ongoing validation and iterative refinement. Domain experts can inspect, modify and extend the strategy set at any time, ensuring alignment with professional standards and ethical requirements. Experimental comparisons with GPT-4o vanilla across three roles, evaluated through an ablation study and a multi-evaluator panel combining LLM-based and human judges, consistently rank XBot as the best performing system across all dimensions, demonstrating superior empathy, role stability and conversational depth, while GPT-4o vanilla exhibits pervasive persona drift across all experimental scenarios.
Luciano Caroprese, Ester Zumpano, M. Aracne et al.· Discover Artificial Intellig...· 0 citations
Healthcare digitalization demands platforms capable of integrating heterogeneous biomedical data and delivering AI-powered clinical decision support at scale [1, 2]. The SupportAI project, funded by the Italian National Recovery and Resilience Plan (PNRR) through the Tech4You program (grant ECS 00000009, Spoke 6, Line B, Project 2.1), addresses this challenge through a comprehensive modular platform that unifies real-time clinical data integration, advanced 3D imaging, generative AI assistance, and collaborative tools within a standards-compliant architecture. The platform’s core architectural innovation is the embedding of Retrieval-Augmented Generation (RAG) directly within live HL7 FHIR R4 and DICOM infrastructure. Unlike offline RAG systems that operate on pre-exported or periodically updated datasets, SupportAI queries the Microsoft FHIR Server and Orthanc PACS at inference time, ensuring AI-generated responses always reflect the patient’s current multimodal record with full traceability to FHIR resource IDs and DICOM study UIDs. Additional contributions include Fourier frequency-domain imaging for anomaly detection beyond the spatial domain, an AI-assisted 3D bioprinting pipeline converting DICOM studies into patient-specific STL anatomical models, and a Jitsi-based telecollaboration infrastructure for multidisciplinary remote consultation. The platform comprises seven interconnected microservices: a Multimodal Data Source Module with automated ETL pipelines and GDPR pseudonymization; a FHIR Server organizing clinical information via RESTful APIs; an Orthanc PACS for DICOMweb-compatible image access; an Advanced 3D Imaging Module with GPU-accelerated Cornerstone3D rendering; a 3D Bioprinting Module supporting AI-assisted MONAI segmentation and manual VTK workflows for STL generation; an AI Module implementing LangChain-orchestrated generative assistance grounded in live clinical records; and a Collaboration Module providing Jitsi-based video consultation in virtual 3D laboratory environments. Security is enforced through TLS 1.3, OAuth 2.0/OpenID Connect, and role-based access control. Technical validation at Technology Readiness Level 6 was conducted using Apache JMeter with 10 concurrent virtual users. Results demonstrate platform robustness across all modules. 3D imaging visualization, AI segmentation, and FHIR patient reads achieved sub-15 ms mean response times; the 3D bioprinting pipeline completed end-to-end within minutes even for complex studies; and the generative chat pipeline averaged 23,629 ms, consistent with GPT-4-class deployments under equivalent context complexity. All eight tested configurations achieved a 0.00% error rate under concurrent load. Static code analysis (SonarQube) on the full 1,214k-line codebase achieved Quality Gate: Passed with zero open security issues. Functional correctness was confirmed through complete User Acceptance Testing across all modules and four representative clinical scenarios: 3D imaging and bioprinting, generative AI clinical assistance, AI-assisted image segmentation, and telecollaboration. Physician evaluators reported substantially increased trust in AI outputs due to transparent source attribution, and post-segmentation manual correction requirements were markedly reduced compared to purely manual workflows. These results confirm the technical feasibility of unified multimodal biomedical data integration within a standards-compliant clinical AI platform and establish SupportAI as a solid foundation for AI-powered decision support targeting personalized, accessible healthcare.
Francesco De Rose, Francesco Dattola, Catia Grazioso et al.· ACM International Conference...· 0 citations