Technical Summary: VitalAgent

Problem Statement

Current mobile health (mHealth) systems utilizing wearable devices (ECG/PPG) face significant limitations in transforming continuous physiological streams into actionable insights. Existing approaches typically rely on task-specific pipelines that map signals to predefined outputs (e.g., sleep stages, arrhythmia detection) or use Large Language Models (LLMs) operating over static, pre-computed textual summaries. These systems lack the ability to:

1. Support Temporal Reasoning: They struggle with natural language queries requiring retrieval, temporal comparison, and aggregation over longitudinal histories (e.g., "Has my resting heart rate increased recently?").
2. Maintain Persistent Context: They cannot reliably compute personalized evidence on demand or maintain a monitoring context across interactions because they lack direct access to raw signals.
3. Enable Proactive Monitoring: Prior work is largely reactive, responding only to user queries, whereas real-world systems require continuous tracking of state changes to surface relevant alerts.

Methodology: VitalAgent Framework

The authors propose VitalAgent, a unified, tool-augmented agentic framework designed for ECG- and PPG-based physiological monitoring. It integrates reactive question answering (QA) and proactive monitoring through three core components:

1. Physiological Memory

VitalAgent maintains a compact, temporal memory ($M_t$) defined as a tuple of raw signal windows ($x_{<t}$), streaming health-state trackers ($s_{<t}$), and proactive alert records ($e_{<t}$).

• Raw Signals: Stores time-indexed ECG/PPG windows.
• Derived States: Includes computed features like heart rate, beat-to-beat intervals, variability measures, signal quality scores, and rhythm labels.
• Alert History: Stores triggered alerts, rules, and justifications to support follow-up queries.
• Update Mechanism: The memory is continuously updated as new signals, statistics, and alerts arrive, allowing the system to retain context for both reactive and proactive modes.

2. Unified Reasoning Layer

A policy interface ($\Pi$) orchestrates the agent's behavior using an LLM-based planner, an evidence validation module, and a response generator.

• Reactive Mode: Triggered by user queries. The planner selects tools based on the query and current memory to retrieve signals, perform analysis, and generate answers.
• Proactive Mode: Triggered by streaming physiological updates. The system processes fixed-length (10-second) non-overlapping windows, extracts measurements, and evaluates alert criteria (extreme bradycardia, extreme/sustained tachycardia). An LLM-based judge runs every 20 windows to reduce missed alerts.
• Validation & Replanning: The system validates evidence across four dimensions: plan completeness, tool success, output field presence, and cross-tool consistency. If validation fails, the system replans by re-prompting the LLM with feedback.

3. Structured Tool Interface

VitalAgent utilizes a registry of 29 modular tools categorized into five functional groups, enabling dynamic computation over raw signals rather than relying on pre-computed statistics:

• Signal Analysis Tools (13): Deterministic computations on raw windows (e.g., analyze_heart_rate, ecg_diagnosis).
• Data-Specific Processing Tools (3): Handle heterogeneity across datasets (e.g., normalization, window alignment).
• Record Lookup Tools (6): Metadata access and record enumeration.
• Proactive Context Tools (5): Retrieval of stored alert records and explanations.
• Medical Knowledge Tools (2): Access to external biomedical knowledge (MedlinePlus, PubMed).

Key Contributions

1. VitalAgent Framework: A novel agentic mHealth system that unifies reactive QA and proactive monitoring via shared memory and dynamic tool invocation.
2. VitalBench Dataset: The first longitudinal benchmark unifying ECG and PPG data for both reactive and proactive evaluation. It comprises:
   • 1,862 QA pairs: Divided into Tier A (window-local, instantaneous) and Tier B (longitudinal, requiring temporal aggregation).
   • 90.2 hours of continuous recordings: From four datasets (Icentia11k, AF-PPG-ECG, PPG-DaLiA, WESAD) covering cardiac, activity, and stress domains.
3. Empirical Validation: Demonstrating that dynamic tool use and long-term memory significantly outperform existing baselines.

Experimental Results

Reactive Question Answering

Evaluated on VitalBench under a "leakage-free" setting (where the model must derive features from raw signals):

• Performance: VitalAgent achieved an overall accuracy of 0.860 on Tier A and 0.557 on Tier B.
• Comparison: It outperformed the strongest signal-derived baseline (LifeAgent*) by 31.5% (Tier A) and 25.7% (Tier B). It also significantly outperformed prompt-based baselines (Health-LLM) and ReAct-style agents (PHIA).
• Ablation: Removing the semantic planner caused a substantial drop in accuracy (e.g., Tier A overall dropped from 0.860 to 0.701), highlighting that question-dependent tool selection is critical. Validation and replanning modules provided marginal gains (<1.2%) but served as safeguards for edge cases.
• Error Analysis: Failures were primarily due to answer mismatches (75.4%) rather than tool selection errors. Common errors involved complex numerical reasoning (e.g., identifying peak heart rates) and distinguishing between window-local and longitudinal queries.

Proactive Monitoring

• Alert Behavior: Adding an LLM judge to rule-based detection reduced median alert latency from 220s to 105s, though it increased the false alert rate (FAR/h) from 1.81 to 2.95, illustrating a trade-off between timeliness and alert burden.
• Rhythm Classification: The pipeline achieved a balanced accuracy of 66.7% for per-window AF rhythm classification across ECG and PPG modalities, with a sensitivity of 75.5%.

Significance and Claims

The paper claims that VitalAgent represents a step toward agentic health systems that unify continuous monitoring with interactive reasoning. Its significance lies in:

• Dynamic Tool Use: Demonstrating that dynamically selecting tools for query-dependent physiological reasoning on raw signals is superior to relying on pre-computed features or static summaries.
• Long-Term Context: Highlighting the necessity of longitudinal physiological memory for both answering complex temporal questions and enabling reliable proactive alerting.
• Architectural Protocol: The framework is designed to be compatible with emerging Model Context Protocol (MCP) architectures, where raw data remains on-device, and the LLM orchestrates modular physiological operators rather than directly accessing sensor data.

The authors acknowledge limitations, including the focus on ECG/PPG modalities, reliance on predefined rules for proactive triggers, and evaluation on controlled benchmark datasets rather than noisy real-world deployments. They position VitalAgent as a decision-support tool, not a diagnostic system.