arXivPaperNeeds Review

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

DIASENTINEL combines deterministic extraction, guideline retrieval, risk prediction, and hybrid verification on-premise. It is a useful architecture reference for auditable agents handling sensitive data.

arXiv · Aug 31, 2026

Open Source Open Markdown Open JSON

Source Summary

DIASENTINEL is an **on-premise multi-agent system** for one-year diabetes-risk screening from EHRs. It combines calibrated prediction, deterministic signal extraction, **Reciprocal Rank Fusion** over ADA guidelines, and rule-plus-LLM verification.

Practical Implication

For high-stakes agent workflows, separate extraction, retrieval, generation, and verification instead of asking one model to do everything. Preserve cited recommendations, verification outcomes, and raw-input comparisons for review.

Agent-Ready Context

DIASENTINEL is an **on-premise multi-agent system** for one-year diabetes-risk screening from EHRs. It combines calibrated prediction, deterministic signal extraction, **Reciprocal Rank Fusion** over ADA guidelines, and rule-plus-LLM verification.

For high-stakes agent workflows, separate extraction, retrieval, generation, and verification instead of asking one model to do everything. Preserve cited recommendations, verification outcomes, and raw-input comparisons for review.

The material describes a demonstration but gives no accuracy, calibration, hallucination, or operational benchmarks. Its reliability and clinical usefulness therefore cannot be judged from the abstract alone.

Connected Context · Feed7 Judgment

This confirms the prior pattern that high-stakes agents need staged responsibilities, deterministic controls, provenance, and reviewable verification rather than a single fluent pass. It applies that architecture to on-premise clinical screening and adds raw-input comparison, but narrows confidence sharply: unlike candidates with measured task results, the abstract supplies no evidence that the design improves prediction, calibration, or clinical reliability.

Benchmarking and Enhancing LLMs for Rule-Intensive Review of National Standard DocumentsBoth separate rule-intensive work into specialized stages with explicit verification; the standards study supplies quantitative evidence for that pattern, while DIASENTINEL does not report outcome benchmarks.Build for the Memo, Not the Demo — Shawn Chan, China Resources HoldingsIts requirement to preserve cited recommendations, verification results, and raw-input comparisons implements the earlier claim-level provenance and auditability contract in a clinical workflow.Split the Labor: Separating Evidence Interpretation from Decision AggregationBoth isolate evidence processing from downstream judgment and reserve part of the workflow for deterministic methods, reducing reliance on unstructured multi-agent agreement.What Does Done Even Mean? Agents and Paperclip's Liveness Model - Dotta, PaperclipDIASENTINEL’s retained verification evidence supports treating completion as reviewable evidence rather than an agent-declared result, though the abstract does not define approval authority or residual risk.

Context Map

agentdata#multi-agent#retrieval#agent-reliability

Uncertainty

The material describes a demonstration but gives no accuracy, calibration, hallucination, or operational benchmarks. Its reliability and clinical usefulness therefore cannot be judged from the abstract alone.