DIASENTINEL: An Auditable Multi-Agent System for Guideline-Grounded Diabetes Risk Screening

The promise of large language models (LLMs) in clinical decision support is immense, yet their integration into high-stakes environments like healthcare remains fraught with challenges. Hallucinated facts, unsubstantiated recommendations, and opaque reasoning are unacceptable risks when patient well-being is on the line. The field demands not just powerful predictive capabilities, but verifiability and auditability. This is precisely the critical gap addressed by DIASENTINEL, a groundbreaking multi-agent system that redefines what’s possible for responsible AI in medicine.

Executive Summary: Forging Trust in Clinical AI

For years, the medical community has grappled with the double-edged sword of AI: incredible potential offset by a foundational lack of trust in its outputs. DIASENTINEL confronts this head-on, presenting a robust framework for one-year Type 2 Diabetes Mellitus (T2DM) risk screening that is not only highly accurate but demonstrably reliable. This isn’t merely another LLM application; it’s a meticulously engineered multi-agent system designed for auditable, guideline-grounded clinical decision support. By operating fully on-premise, it also sets a new standard for privacy-preserving AI in healthcare. DIASENTINEL fundamentally shifts the conversation from “can AI do it?” to “can we trust AI to do it responsibly?”

Technical Deep Dive: A Multi-Agent Framework for Audited Clinical Intelligence

The ingenuity of DIASENTINEL lies in its sophisticated multi-agent system architecture, integrating several specialized components to form a cohesive, verifiable pipeline. At its core, it tackles the inherent unreliability of standalone LLMs by embedding them within a structured ecosystem of checks and balances.

1. Calibrated Risk Prediction: The system begins with Machine Learning models trained for robust, calibrated one-year T2DM risk prediction, drawing insights from electronic health records (EHRs). This establishes a baseline of evidence-driven foresight.
2. Deterministic Clinical Signal Extraction: Before any generative AI touches the data, critical clinical signals are deterministically extracted from EHRs. This ensures foundational facts are precisely identified and not subject to LLM interpretation or hallucination.
3. Guideline-Grounded Reasoning: This is where DIASENTINEL truly distinguishes itself. Leveraging Reciprocal Rank Fusion (RRF), the system dynamically queries and integrates relevant recommendations from authoritative sources like the American Diabetes Association (ADA) guidelines. This ensures that any generated insights are explicitly linked to established clinical protocols, making the system inherently “guideline-grounded.”
4. Hybrid Verification Layer: Perhaps the most pivotal innovation is the hybrid verification layer. This isn’t just a simple factual check; it combines:
   • Rule-based checks: Hard-coded clinical rules and logic provide a deterministic guardrail against common errors or illogical recommendations.
   • LLM entailment: A separate LLM agent is tasked not with generating recommendations, but with verifying the logical consistency and support for claims made by other agents against the extracted EHR data and ADA guidelines. This creates an internal peer-review mechanism, drastically reducing the risk of unsupported statements.

The entire architecture functions like a council of expert AI agents, each with a distinct role: one predicts, another extracts facts, a third grounds recommendations in guidelines, and a final, critical agent verifies the entire output chain. This rigorous, auditable process provides a comprehensive report, complete with cited recommendations, detailed verification results, and direct comparisons to raw EHR data, offering unparalleled transparency.

Real-World Applications: Transforming Proactive Healthcare

The implications of DIASENTINEL extend far beyond academic demonstration:

• Proactive Population Health Management: Its real-time batch-screening dashboard enables healthcare providers to identify at-risk populations efficiently, facilitating early intervention strategies and resource allocation for preventative care.
• Enhanced Individualized Care: The interactive patient report interface empowers clinicians with readily accessible, thoroughly vetted information, streamlining decision-making and allowing for more personalized and effective patient counseling.
• Reducing Physician Burden: By automating the painstaking process of cross-referencing patient data with extensive clinical guidelines, DIASENTINEL frees up physician time, allowing them to focus on direct patient interaction rather than exhaustive data synthesis.
• Trust and Transparency with Patients: The ability to present patients with transparent reports detailing risk factors, guideline-backed recommendations, and verification steps fosters greater trust in AI-assisted diagnoses and treatment plans.
• Privacy-Preserving Deployments: The on-premise nature of the system addresses a paramount concern in healthcare—data privacy and security—making it a viable solution for institutions with strict compliance requirements.

Future Outlook: The Blueprint for Reliable AI in High-Stakes Domains

DIASENTINEL is more than a system for diabetes screening; it’s a blueprint for building reliable, auditable, and privacy-preserving AI solutions across all high-stakes domains. In the next 2-3 years, we can expect this paradigm to expand rapidly:

• Broader Clinical Applications: The framework could be adapted for screening and managing risk for a myriad of other chronic conditions, from cardiovascular disease to early cancer detection, transforming the landscape of preventative medicine.
• Standardization of Auditable AI: This work will likely catalyze efforts to standardize methodologies for verifiable AI, prompting regulatory bodies to define clearer guidelines for LLM and AI agents deployment in healthcare.
• “Explainable by Design” Systems: The principles of DIASENTINEL champion AI that is “explainable by design,” moving beyond post-hoc explanations to systems inherently built for transparency and verification. This will be crucial for widespread adoption and public confidence.
• Integration with Regulatory Frameworks: As AI becomes more embedded in healthcare, systems like DIASENTINEL will inform the development of regulatory frameworks, potentially becoming a foundational requirement for FDA approval of AI-driven diagnostics.

This isn’t merely about incremental improvements in prediction accuracy. It’s about fundamentally altering how we perceive and interact with AI in critical applications—moving from a leap of faith to a foundation of verifiable trust.

Key Takeaways

• DIASENTINEL directly addresses the critical challenge of LLM hallucination and unreliability in high-stakes clinical decision support.
• Its multi-agent system architecture integrates calibrated risk prediction, deterministic signal extraction, and guideline-grounded reasoning.
• The hybrid verification layer, combining rule-based checks and LLM entailment, is paramount to its auditable and trustworthy nature.
• Demonstrates a practical, privacy-preserving framework for real-time batch screening and interactive, transparent patient reporting.
• DIASENTINEL serves as a vital template for building auditable and reliable AI agents across healthcare and other sensitive domains, paving the way for responsible AI adoption.