Edge-Decimated Biometric Ingestion and Deterministic AI Clinical Orchestration: A Three-Tier Architecture for Safe Remote Patient Monitoring

Abstract

Remote patient monitoring (RPM) systems face two critical unsolved challenges: (1) Bluetooth Low Energy (BLE) data continuity failure when progressive web applications (PWAs) enter background execution, causing silent, irrecoverable loss of patient biometric data; and (2) large language model (LLM) hallucination during clinical alert generation, where AI fabricates vital sign values at rates of 15-40% in clinical interpretation tasks. This paper presents PulseView, a three-tier architecture that addresses both challenges simultaneously through a novel combination of edge-decimated biometric ingestion and deterministic AI clinical orchestration. Tier 1 introduces a native micro-buffer daemon that maintains BLE GATT subscriptions independently of the web-view lifecycle, achieving zero data loss with 24:1 time-series compression. Tier 2 implements a stateful temporal aggregation engine that computes per-parameter Z-scores against patient-specific 72-hour rolling baselines and a Multi-Variate Correlation Index (MVCI) using simplified Mahalanobis distance. Tier 3 introduces the paper's primary contribution: a rule-based AI validation layer that constrains LLM operations, intercepts every output for four-category hallucination detection with ±5% tolerance cross-referencing against verified ground truth, and substitutes deterministic fallback templates when hallucination rates exceed configurable thresholds. A working prototype has been developed and is available upon request. The system is described in USPTO Provisional Patent Application #64/005,956, filed March 15, 2026.