Julian Kurz 1, Ali Farnoud
1 Hms Analytical Software (Heidelberg, Germany)
Introduction/Objectives:
Exploratory data analysis (EDA) of clinical study data requires rapid iteration on visualizations and data summaries, yet many workflows still rely on specialized technical teams to implement changes. This creates a bottleneck that slows decision-making and limits the pace of data-driven insights. Prior work in collaborative analytics and computer-supported cooperative work (CSCW) highlights the value of systems designed around cooperative practices and shared analysis workflows [1], [3], while studies on decision quality under uncertainty emphasize the need to surface reliability and reduce false discovery risk during exploratory analysis [2]. We developed a multi-agent interactive EDA application where analysis code is written in R, existing R code can be reused directly, and execution occurs in an R environment, preserving continuity with established workflows. The design is aligned with emerging descriptions of agentic systems that combine multi-agent coordination and persistent memory to support complex analytical workflows [5]. Objectives were to (1) reduce turnaround time for common visualization adjustments, (2) broaden usability across technical and non-technical roles, (3) improve transparency and accountability through structured logging, (4) provide a foundation for reusable analytics across clinical programs, and (5) enable users with limited programming experience to explore clinical datasets through an AI-assisted interface that acts as a virtual data scientist alongside them.
Methods:
The system implements a modular, multi-agent workflow that routes user requests to specialized agents for visualization creation, visualization modification, data wrangling, and history retrieval. The design draws on agentic system patterns and multi-agent architectures described in the software engineering literature [4], [5], emphasizing separation of roles, intent routing, and memory-backed state. Users can interact through guided menus, natural-language chat, or an advanced scripting interface, allowing both novice and expert users to work within the same environment. The analytics layer executes in an R environment, enabling continuity with established workflows and direct reuse of previously validated analyses. The application maintains persistent state and a dual-layer memory model, including episodic logging of actions and outcomes, creating a detailed trace of analytical intent, execution, and results to support transparency and regulatory review. The architecture is designed for extensibility, allowing domain-specific adaptation without re-engineering the core workflow and supporting collaborative EDA practices described in prior hybrid analytics systems [1].
Results:
Stakeholder testing across three independent user groups with varying levels of programming expertise demonstrated that routine visualization adjustments that previously required multi-day handoffs to technical teams can be completed within minutes. Non-technical users reported improved usability and confidence in exploring study data, while expert users retained flexibility for advanced analyses. Multiple groups have expressed interest in reusing the backend, and the same data-manipulation and analysis capabilities are applicable beyond visualization. For example, to prepare datasets for downstream modeling or machine-learning workflows. This indicates broader potential utility across clinical programs. The episodic audit trail provides a more detailed record of analytical decisions than prior ad hoc workflows, supporting accountability concerns raised in agent architecture research [4]. These outcomes align with prior findings that structured EDA workflows can improve decision quality under uncertainty [2] and that collaborative systems benefit from explicit support for coordination and traceability [1], [3].
Conclusions:
A modular, multi-agent interactive EDA application that maintains R-based analytics can accelerate clinical data exploration while expanding accessibility and auditability. The combination of multi-modal interaction, persistent state, and episodic logging offers a practical model for integrating AI-assisted workflows into regulated clinical study analysis without sacrificing rigor or established practices. This is consistent with recent perspectives on agentic AI capabilities and memory-centric design [5].
References:
[1] Immersive Insights: A Hybrid Analytics System for Collaborative Exploratory Data Analysis. 25th ACM Symposium on Virtual Reality Software and Technology, 2019.
[2] Odds and Insights: Decision Quality in Exploratory Data Analysis Under Uncertainty. (Supplementary materials: https://osf.io/xtsfz/)
[3] Computer‑Supported Cooperative Work. In: Handbook of Human‑Computer Interaction.
[4] Liu Y. et al. Agent design pattern catalogue: A collection of architectural patterns for foundation model‑based agents. Journal of Systems and Software 220 (2025) 112278.
[5] Sapkota R. et al. AI Agents vs. Agentic AI: A conceptual taxonomy, applications and challenges. Information Fusion 126 (2026) 103599.
Reference: PAGE 34 (2026) Abstr 11917 [www.page-meeting.org/?abstract=11917]
Poster: Methodology – AI/Machine Learning