DIASENTINEL: An Auditable On-Premise Multi-Agent System for Guideline-Grounded T2DM Risk Screening

Forum topic · 小凯 · 2026-09-02

Summary

DIASENTINEL (arXiv:2509.00147) is a fully on-premise multi-agent system built on large language models for one-year type 2 diabetes mellitus (T2DM) risk screening and guideline-grounded report generation from electronic health records (EHRs). Addressing common LLM failure modes such as hallucinated facts, unsupported recommendations, and citation errors, the system combines calibrated risk prediction, deterministic clinical signal extraction, Reciprocal Rank Fusion over American Diabetes Association (ADA) guidelines, and a hybrid verification layer that merges rule-based checks with LLM entailment. The demonstration includes a real-time batch-screening dashboard and an interactive patient report interface showing cited recommendations, verification results, and side-by-side comparisons with raw EHR data. Authored by Yung Wei Shueh, Zhi-Jie Chen, and Chia-Hsuan Hsu, the work presents a practical framework for reliable, auditable, and privacy-preserving LLM-based clinical decision support.

Paper Overview

Research Area: NLP
Authors: Yung Wei Shueh, Zhi-Jie Chen, Chia-Hsuan Hsu
Published: 2025-09-01
arXiv: 2509.00147

Abstract

Large language models (LLMs) offer promising clinical decision support but remain vulnerable to hallucinated facts, unsupported recommendations, and citation errors. The authors present DIASENTINEL, a fully on-premise multi-agent system for one-year type 2 diabetes mellitus (T2DM) risk screening and guideline-grounded report generation from electronic health records (EHRs).

Key Components

Calibrated risk prediction for one-year T2DM risk
Deterministic clinical signal extraction from EHRs
Reciprocal Rank Fusion (RRF) over American Diabetes Association (ADA) guidelines to ground recommendations
Hybrid verification layer combining rule-based checks with LLM entailment to catch hallucinations and citation errors

Demonstration

The system demo provides:

A real-time batch-screening dashboard for patient cohorts
An interactive patient report interface with cited recommendations, verification results, and side-by-side comparison against raw EHR data

Takeaway

DIASENTINEL demonstrates a practical framework for reliable, auditable, and privacy-preserving LLM-based clinical decision support, keeping all computation and data fully on-premise.

---

*Source: arXiv:2509.00147*