José Luís Oliveira

Objective: Clinical AI tools for medical imaging are increasingly capable but remain largely isolated from one another, requiring custom integration work that slows adoption and creates fragmented, difficult-to-maintain deployments. We present MedGov-AI, an agentic system that addresses this through the Model Context Protocol (MCP) as a standardized interoperability layer, with a large language model (LLM) acting as an autonomous orchestrator that discovers and invokes clinical tools at runtime.Methods: Six MCP servers expose tasks spanning DICOM parsing, computed tomography (CT) segmentation, radiology report generation, whole slide image (WSI) cell segmentation, and digital pathology. Domain-specific workflow governance is provided through skill files that encode clinical sequences and guard conditions without model retraining. The system supports both interactive and event-driven operation, in which file ingestion automatically triggers autonomous agent execution without user input. Three end-to-end use cases were evaluated, comparing cloud-hosted against local LLMs and skill-guided against unguided execution.Results: Cloud-based LLMs completed multi-step clinical workflows reliably within a bounded iteration budget, whereas locally hosted 7–8B-parameter models did not, revealing a tension between privacy and reliability in clinical deployment. Skill-guided execution prevented clinically incorrect tool selection at ambiguous decision points, with pathology reporting succeeding in 10/10 runs with the skill versus 0/10 without. Six MCP servers were integrated in this open architecture without any agent-side connector code.Conclusion: MCP combined with skill-based governance can reduce the integration overhead currently limiting clinical AI adoption, providing a practical, extensible layer for autonomous clinical-imaging workflows. The system’s code and documentation are open-source at https://github.com/BMDSoftware/MedGov-AI-MCP, and a demo is publicly available at https://medgovai.bmd-software.com/.

The continuous generation of data in clinical practice and research settings holds enormous potential for improving diagnoses and treatments. Medical data science encompasses a wide range of modalities and methodologies, each with distinct workflow requirements and specialized processing tools across diverse data domains such as audio, video, and text. However, these tools are often developed in isolation, with limited interoperability, and are not designed to function together within a unified workflow. As a result, building analysis pipelines becomes difficult and time-consuming. Conflicting installation requirements further complicate pipeline management, limiting reproducibility and hindering the sharing of workflows among researchers. To address these challenges and automate the development of processing pipelines, this work proposes a modern cloud-native platform, built entirely on open-source components, for constructing data analysis workflows in the medical sciences. It integrates the analysis tools already used by scientists into reproducible environments while offering straightforward mechanisms for incorporating new tools through a user-friendly interface that requires no advanced technical expertise. This work is part of a broader initiative to streamline research on Alzheimer's disease. The platform has been deployed in production, serving over 300 users across more than 250 workspaces, and currently integrates 60 processing modules spanning imaging, audio, text, and genomics.

The secondary use of clinical data has been widely studied, addressing many challenges in conducting observational studies. However, the complexity of dataset structures and detailed data requirements has led researchers to develop user-friendly query builders, enabling medical researchers to define cohorts more efficiently across the datasets. Although these tools simplify the workflow, their learning curve can sometimes be steep. Motivated by improving the usability to conducted observational studies, we propose a modular conversational assistant framework that would address these limitations. It can be integrated in any web application as an javascript component, improving the system usability. Additionally, the proposed framework would employ deterministic algorithms to reduce computational overhead. The system would enable integration into existing medical information systems through configuration files rather than code modifications. Validation within the OHDSI ecosystem would demonstrate practical applicability for real-world observational research scenarios.

Clinical research studies face substantial operational and methodological challenges that compromise quality and reproducibility. Traditional risk assessment requires specialized expertise and significant time investment, creating barriers for research teams. With the advent of new agentic AI models, there are emerging opportunities to leverage these processes more effectively. Therefore, this paper proposes an integrated web-based assistant that democratizes risk assessment in observational clinical studies through automated guideline import, structured question-answer workflows, and automated report generation. By combining document-oriented database architecture with workflow orchestration, the proposed tool addresses gaps between commercial quality management systems and practical needs of smaller research teams when conducting observational studies.

Background: Genomics analyses often rely on command-line tools executed via remote servers, imposing usability barriers for non-technical users and raising privacy concerns. WebAssembly (WASM) enables native-code execution directly in web browsers, eliminating installations and data transfers. Results: We introduce BioChef, a client-side genomic workflow platform that uses WASM. BioChef compiles a genomics toolkit into browser-executable modules accessible via an intuitive drag-and-drop GUI. The system provides real-time validation, flexible input methods (form-based and JSON), intermediate step inspections, and reproducible workflows exportable as bash scripts or configuration files. Performance benchmarks across major browsers (Chromium, Gecko, WebKit) demonstrate rapid initialization (LCP 0.583 s), responsive interactivity (INP 30.5 ms), minimal layout shifts (CLS 0.01), and acceptable overhead (average 181.5 ms initial WASM module load). Although browser execution introduced performance penalties (~130x slower than native), BioChef workflows still significantly outperformed traditional web services such as Galaxy by avoiding network delays and server-side queueing (11.3x faster in a standard pipeline benchmark). Conclusions: BioChef shows how WebAssembly on the client side can democratize genomic data processing, ensuring privacy, reproducibility and ease of use without external dependencies. To our knowledge, this is the first fully client-side, graphical genomic workflow environment powered by WASM. Availability: Online demo: https://ieeta-pt.github.io/Biochef. Source code, WebAssembly binaries, and reproducibility artifacts: https://github.com/ieeta-pt/Biochef.

Clinical research often requires integrating data from diverse sources, which differ not only in structure but also in semantics and language. Traditional extract-transform-load (ETL) pipelines struggle to handle semantic variability and lack built-in support for multilingual or ontology-driven harmonisation. This fragmentation limits the interoperability and reuse of clinical datasets in large-scale analyses. In this paper, we propose an integrated framework that combines an embedding-based concept mapping engine with an automated ETL pipeline using Apache Airflow. The mapping engine uses transformer-based embeddings to align clinical terms with standard concepts, producing outputs in White Rabbit and Usagi-compatible formats to ensure backward interoperability. We validated the system using multilingual real-world datasets demonstrating its ability to handle heterogeneous inputs and maintain end-to-end reproducibility.