Technical Summary: Kauro, a Graph-Based Chatbot for High-Fidelity Information Transmission

Problem Statement

In biomedical research and clinical care, the transmission of critical information—particularly during the informed consent process—is a high-stakes activity with profound ethical, legal, and practical consequences. Current methods face significant limitations:

• Variability and Labor Intensity: Traditional consent conversations are labor-intensive, limited by staff working hours, and prone to variability in how information is transmitted across different sites and representatives.
• Limitations of Existing Digital Tools: Existing "eConsent" platforms largely reproduce static documents (PDFs) or linear multimedia, lacking reproducible conversational interaction or deterministic audit trails.
• Risks of Generative AI: While generative chatbots (LLM-based) offer flexibility, they introduce stochasticity, hallucination, and regulatory risks. Their probabilistic nature makes it difficult to guarantee that IRB-approved language is preserved or to provide verifiable audit trails, which are essential for regulated biomedical research.

There is a critical gap for a scalable, digital solution that ensures safety, precision, and trust without sacrificing reproducibility or auditability.

Methodology

The authors present Kauro, an open-source, graph-based chatbot designed to operationalize constraint rather than flexibility. The system is built on the following technical pillars:

1. Core Architecture: Deterministic Graph Traversal

Unlike probabilistic LLMs, Kauro encodes scripted conversations as version-controlled JavaScript Object Notation (JSON) structures representing a directed graph.

• Node Structure: Each conversation node contains a unique identifier, type (bot or user), parent/child links, message payloads, and rendering instructions (e.g., buttons, text fields).
• Determinism: The engine enforces a deterministic traversal of the graph. Given the same starting point and input, the path and output are predictable and identical every time. This eliminates hallucination and ensures that participants always receive the exact IRB-approved content.
• Modularity (Separation of Concerns): The system decouples three distinct modules:
  • Client: A React-based interface that dynamically renders nodes and handles varied input types (forms, inputs, informational messages).
  • Server: A Django-based backend that acts as the execution engine. It validates user requests against a schema, maps them to node identifiers, and deterministically resolves the next node.
  • Conversation Graph: Portable JSON objects that can be version-controlled, shared, and modified independently of the engine.

2. Governance and Compliance

• Human Backstops: The system includes mandatory human oversight. Humans design or approve all graph nodes. Participants can opt-out at any node to speak with a study staff member.
• Quiz-Gating: Progression through the consent process is contingent on passing knowledge checks (e.g., a 10/10 quiz score). Failure triggers a "teach-back" system or requires staff intervention.
• Audit Logging: Every interaction is logged with standardized fields (actor, node_id, payload, timestamp, graph_version, and IRB reference). This creates a complete, immutable audit trail compliant with FDA Part 11 and EMA expectations.
• Version Control: Every chat session is explicitly linked to a specific graph version and IRB approval reference, allowing investigators to reconstruct the exact experience a participant had.

3. Implementation Details

• API Design: Kauro utilizes a strict REST-aligned API with schema-driven inputs and deterministic outputs, contrasting with many public APIs that lack consistent error handling or version control.
• Deployment: As open-source software (CC-BY-4.0 license), Kauro can be deployed in various compute environments, allowing institutions to maintain control over data and auditability.
• Validation: The system employs unit tests for node structures and schema conformance, integrated into a continuous pipeline to ensure changes do not alter approved traversal behavior.

Key Contributions

• Reframing AI in Biomedicine: Kauro demonstrates that machine intelligence for high-stakes communication can be achieved through deterministic, graph-based logic rather than probabilistic models, prioritizing reproducibility and auditability over conversational novelty.
• Open-Source Portability: Unlike proprietary commercial tools, Kauro is designed for portability across institutions, with a modular architecture that allows adaptation to diverse research contexts without modifying the underlying engine.
• Full Auditability: The system provides a complete, traceable audit trail linking every user interaction to a specific IRB-approved script version, addressing a gap in existing eConsent platforms that often lack partial or no logging.
• Scalable Consent Workflow: By automating the delivery of standardized information while maintaining human oversight options, Kauro addresses the labor and time constraints of multi-site studies.

Results and Validation

The paper reports on the development and initial implementation of Kauro, supported by prior pilot data and specific use cases:

• Prior Pilot Success: A previous pilot using a similar chatbot concept (in a commercial context) showed that participants completed sessions faster than in traditional workflows, achieved high knowledge check scores, and reported no adverse events.
• Current Implementation: Kauro was implemented to support the GREGoR Consortium (a large-scale rare and Mendelian condition consortium) at the University of California, Irvine.
  • The system was used to recruit participants for exome and diagnosis-negative cases.
  • The specific implementation included a 10-question quiz with a 100% pass requirement and a teach-back mechanism for missed questions.
  • The IRB (IRB #4469) approved the process, including the JSON-based script and the ability to export PDFs for review that preserve the informatic backbone (node/edge IDs).
• Technical Validation: The system includes automated integration, regression, and unit tests to ensure that updates to the client, server, or scripts do not alter the approved traversal behavior.

Significance and Claims

The paper positions Kauro not as a replacement for human interaction in all contexts, but as a specialized tool for high-fidelity information transmission where safety, precision, and regulatory compliance are paramount.

• Safety by Design: The authors argue that "constraint" is a feature, not a bug. By limiting flexibility, Kauro prevents deviation from approved content, ensuring that the "ritual" of informed consent remains ethically sound and legally defensible.
• Regulatory Alignment: The system is explicitly designed to meet regulatory requirements for audit trails and version control (FDA Part 11, EMA), which many current digital tools fail to provide robustly.
• Generalizability: While initially designed for informed consent, the authors claim the framework is generalizable to any domain requiring standardized, accurate information transmission, such as clinical trials, disease registries, and public health initiatives.
• Transparency: By releasing the code as open-source and separating concerns, Kauro offers a transparent alternative to "black box" commercial solutions, enabling independent inspection and validation by regulators and institutions.

The paper concludes that Kauro reframes the concept of "intelligence" in consent from generative flexibility to deterministic delivery, complete audit trails, and verifiable compliance at scale.