A Multi-Analyst LLM Pipeline for Auditable Rule Discovery Across 68 Public Physiological Corpora

Technical Summary: A Multi-Analyst LLM Pipeline for Auditable Rule Discovery

Problem Statement
Open physiological corpora are inherently heterogeneous, utilizing diverse sensors, labels, sampling rates, and clinical endpoints. While these datasets support detector design, they do not directly specify engineering rules for new contactless monitoring platforms. Specifically, a rule performing well on one dataset may fail due to sensor changes, label shifts, or deployment context differences. The authors address the challenge of constructing a detector-rule library for a contactless nocturnal-monitoring platform (utilizing millimeter-wave radar, WiFi CSI, thermal sensing, acoustic features, and ambient sensors) where no single public corpus matches the target hardware, and no existing corpus directly specifies an engineering rule library.

Methodology
The paper proposes a controlled, four-analyst Large Language Model (LLM) workflow designed to convert documentation from 68 screened public physiological corpora into an auditable library of candidate rule shapes. The process is structured as a constrained engineering cascade:

1. Corpus Selection: 68 public corpora were selected based on commercial-use compatibility, documented sensor modalities/sampling rates, and relevance to nocturnal monitoring (e.g., heart rate, respiration, seizure events, bed-exit behavior).
2. Four-Analyst Extraction: Four independent commercial LLM families acted as parallel extractors using a controlled prompt. They were not treated as authorities; rather, their outputs served as candidate markers.
3. Normalization and Deduplication: The analysts produced 695 "top-markers." These were normalized (canonicalizing signal/modality and time-windows) and deduplicated, retaining 649 rule records.
4. Threshold-Bounds Audit: A separate audit flagged 51 sanity violations regarding thresholds, routing them for clamping or curator review to prevent unsafe literature-derived thresholds from proceeding.
5. Cross-Corpus Consolidation: The retained records were consolidated into 436 unique rule shapes.
6. Gate-Tagging and Invariants: Rules were tagged for build-readiness based on two hard invariants:
   • Native Sensor Channel: The rule must map to a sensor channel existing natively on the target hardware.
   • No Multi-Night Personalization: The rule cannot require per-patient baseline learning or individualized user profiles.
7. Curator Review: A human curator reviewed the 436 unique shapes against a fixed rubric (semantic distinctness, literature support, native-channel availability, etc.) to perform engineering triage.

Key Results
The pipeline processed the input through a specific reduction cascade:

• Input: 68 public corpora.
• Extraction: 695 analyst top-markers.
• Retention: 649 rule records (after deduplication).
• Consolidation: 436 unique rule shapes.
• Output: 94 "build-now" detector components.

These 94 components were distributed across four detector-family buckets:

• Autonomic surge with failed recovery (62 components).
• Postictal respiratory compromise (24 components).
• Bed-exit after high-movement event (6 components).
• Postictal recovery risk (2 components).

Inter-Analyst Disagreement Analysis
Contrary to the assumption that agreement implies correctness, the study found that agreement did not predict survival in the final library.

• Agreement Distribution: 67.2% of top-markers had 3/4 or 4/4 agreement.
• Survival Rate: The retention rate was uniform across agreement levels (approx. 93–95% for 4/4 down to 1/4 agreement; chi-square p=0.91).
• Disagreement Utility: Disagreement served as a triage signal. Seizure markers showed significantly higher disagreement (29.4% single-AI) compared to cardiac markers (12.3% single-AI), indicating that contested extractions required more curator attention.

Significance and Claims
The paper explicitly states that its contribution is not a validated clinical detector, nor a benchmark of LLM vendors. It does not report sensitivity, specificity, false-alarm rates, or latency, as these require prospective hardware validation against medical-grade ground truth.

Instead, the significance lies in the methodology:

1. Auditable Engineering Cascade: It demonstrates a workflow where analyst disagreement, threshold checks, and invariant enforcement route literature-derived rules toward prospective validation without making unverified clinical claims.
2. Disagreement as a Signal: It establishes that analyst disagreement is a useful triage mechanism to expose contested extractions, rather than a measure of truth.
3. Proprietary Disclosure Pattern: It offers a pattern for reporting aggregate provenance (counts, methods, audit stages) while withholding proprietary rule definitions, thresholds, and gate rationales to protect product logic.

The authors conclude that the 94 "build-now" components are scoping-stage candidates ready for prospective hardware validation, not deployable clinical alerts. The pipeline successfully converts heterogeneous open-corpus evidence into a structured, auditable set of engineering candidates.