If Only My CGM Could Speak: A Privacy-Preserving Agent for Question Answering over Continuous Glucose Data

1. Problem Statement

Continuous Glucose Monitors (CGMs) generate dense, longitudinal personal health data that could significantly improve diabetes self-management. However, current patient platforms (e.g., Dexcom Clarity, LibreView) offer only static, pre-defined summaries, preventing users from exploring their data through natural language queries.

While Large Language Models (LLMs) could enable conversational interfaces, deploying them for health data presents two critical challenges:

1. Privacy: Cloud-based LLMs require transmitting sensitive raw health data to external servers, creating significant security and regulatory risks.
2. Accuracy: LLMs are prone to numerical errors and hallucinations when processing time-series data, making them unreliable for direct calculation of clinical metrics.

2. Methodology: CGM-Agent Architecture

The authors propose CGM-Agent, a modular, privacy-preserving framework that decouples reasoning from computation. The system operates entirely on the user's local device, ensuring raw CGM data never leaves the device.

The architecture consists of three distinct layers:

• Layer 1: Input Processor (Query Refiner)

  • Function: Acts as a gatekeeper. It validates query feasibility (determining if the question can be answered with CGM data alone), resolves temporal ambiguities (e.g., converting "last weekend" to specific dates), and maps vague user intents to precise system features.
  • Handling Unanswerable Queries: If a query requires external data (e.g., insulin logs, food intake) not present in the CGM file, the processor generates a graceful refusal with an explanation rather than attempting a hallucinated answer.
  • Proxy Instantiation: For questions about unobserved behaviors (e.g., "How does exercise affect my glucose?"), the system maps them to "proxy" time-window analyses (e.g., comparing glucose during typical exercise hours vs. other times).

• Layer 2: Analytical Agent (Tool-Use Engine)

  • Function: The LLM acts solely as a reasoning engine to decompose the refined query into a sequence of deterministic function calls.
  • Execution: These calls are executed in a local sandbox using a custom Python toolkit. The LLM never sees raw time-series data; it only receives aggregated numerical results (e.g., "TIR = 72%").
  • Toolkit: The toolkit implements clinically validated metrics (Time in Range, Mean Glucose, Glucose Variability) and handles data quality checks (e.g., flagging results if CGM wear time is <70%).

• Layer 3: Response Generator

  • Function: Synthesizes the numerical results into a natural language response.
  • Output: It provides direct answers with specific numbers, explicitly states the time period analyzed to ensure alignment with user intent, and offers brief clinical context (e.g., comparing results to ADA targets).

3. Key Contributions

• CGM-Agent Framework: A novel architecture that enables free-form QA over personal health data while strictly adhering to privacy constraints by keeping raw data local and exposing only aggregated metrics to the LLM.
• CGM-QA Benchmark: The construction of a comprehensive dataset containing 4,180 queries:
  • 2,470 Synthetic Questions: Generated from clinical guidelines (ADA, International Consensus) to test functional correctness.
  • 1,710 Real User Questions: Derived from interviews with CGM users, capturing authentic, often ambiguous, information needs.
  • Ground Truth: Unlike many benchmarks relying on human annotation, ground truth here is derived from deterministic program execution, ensuring mathematical precision and reproducibility.
• Empirical Analysis: A rigorous evaluation of 6 leading LLMs (proprietary and open-weight) across the three layers of the pipeline.

4. Experimental Results

The authors evaluated the system using 6 models (including GPT-5, Gemini 3.0, Llama-4, and Nemotron-Nano).

• Performance on Synthetic Queries (Controlled):

  • Top proprietary models achieved 94% Value Accuracy and high F1 scores (>0.80) on tool selection.
  • Lightweight models (e.g., GPT-5-Mini, Gemini Flash) performed competitively with larger models, suggesting cost-effective deployment is feasible.
  • Open-weight models (Llama-4-17B) lagged slightly (75% Value Accuracy) but remained viable.

• Performance on Real-World Queries (Ambiguous):

  • Performance dropped compared to synthetic data (F1 dropped from ~0.80 to ~0.65), primarily due to intent disambiguation and temporal ambiguity rather than computational failures.
  • Value Accuracy remained high (>88% for top models) once the correct intent was identified, confirming that the tool-use paradigm is robust for calculation.
  • Ablation Study: Removing the Input Processor (Layer 1) caused a significant drop in F1 score (-0.19), proving that semantic refinement is critical for handling real-world ambiguity.

• Readability: Generated responses were concise (avg. 108 words) and readable at a 10th-grade level, appropriate for general users.

• Extensibility: A pilot study integrating insulin and carbohydrate logs showed the framework could generalize to multi-modal data with near-ceiling performance (F1 0.987), requiring only the registration of new local executors.

5. Significance and Implications

• Privacy-Utility Trade-off Resolved: The paper demonstrates that high-quality, personalized health insights can be delivered without compromising patient privacy. By confining raw data to a local sandbox and using the LLM only for reasoning, the system mitigates the risks of cloud-based data transmission.
• Trustworthy Health AI: The use of deterministic tool execution eliminates the "hallucination" problem common in LLMs for numerical tasks, providing a reliable foundation for clinical decision support.
• Future Directions: The framework is designed to be modular, allowing for the integration of additional data modalities (sleep, activity) and interactive clarification loops to further resolve user ambiguity.
• Open Science: The authors released the code and the 4,180-question benchmark to foster future research in trustworthy, privacy-preserving health agents.

In conclusion, CGM-Agent represents a significant step forward in making personal health data actionable through natural language, balancing the need for advanced AI capabilities with the strict privacy and accuracy requirements of medical applications.