CGM Accuracy Interpretation Tool

CGM Accuracy Interpretation Tool

animation
healthcare
medication
Try Change CGM MARD (%) while keeping the example reference value unchanged. Then press Calculate percentage arithmetic.
Observe After pressing Calculate percentage arithmetic, watch percentage arithmetic only as CGM MARD (%) changes.
Explain Explain why CGM MARD (%) does not provide a per-reading bound after you press Calculate percentage arithmetic.

Technical boundaries. Multiplication illustrates scale; it does not estimate error at the selected value. Aggregate MARD cannot establish a reading range, alert buffer, or dosing rule.

CGM Accuracy Interpretation Tool

Scenario: A diabetes management app integrates data from continuous glucose monitors (CGMs). Interpret the published aggregate accuracy metric before specifying any alert or dosing behaviour.

Given:

  • CGM sensor: Abbott FreeStyle Libre 3 (14-day wear, factory calibrated)
  • For the FreeStyle Libre 3 study, report 7.8% aggregate MARD against a venous plasma reference in participants aged at least 6 years; do not interpret it as a per-reading bound. (study)
  • Patient population: Type 1 diabetics with 4-6 insulin injections daily
  • Intended function: Insulin-dosing decision support
  • Safety evidence needed: Device labeling, stratified accuracy data, time-series alert validation, and clinical review

Steps:

  1. Interpret the accuracy metric correctly:
    • MARD is the mean of the absolute relative differences across matched CGM and reference measurements.
    • The arithmetic 7.8% of 150 mg/dL equals 11.7 mg/dL, but the aggregate MARD does not establish that as the expected error at 150 mg/dL.
    • It does not imply that every reading lies between 138 and 162 mg/dL, and it cannot establish a dosing-error bound.
  2. Identify conditions that need separate evidence:
    • Performance can vary by glucose range, rate of change, wear period, insertion state, sensor placement, and interfering substances.
    • A population-wide average can hide clinically important tails and subgroup differences.
    • Review the approved labeling and stratified performance tables for the exact sensor version and intended population.
  3. Translate labeling into system states:
    • Preserve device status, alarm state, trend information, and data-quality flags with every reading.
    • Prompt for a blood-glucose check when the approved labeling requires it, including when symptoms or expectations do not match the sensor reading.
    • Do not invent one confirmation threshold for every CGM; indications and warnings differ by device and version.
  4. Validate alert and dosing behaviour:
    • Treat candidate thresholds as design parameters, not outputs of MARD arithmetic.
    • Evaluate sensitivity, specificity, time-to-alert, false-alarm burden, and missed-event severity on representative time-series data.
    • Obtain clinical and regulatory review for the intended use, particularly if the app recommends or automates insulin delivery.
  5. Implement decision support guardrails:
    • Apply clinician-configured dose limits and approved device restrictions.
    • Track active insulin on board and the timing of previous corrections.
    • Fail safely when readings, device status, or required context are missing or contradictory.

Result: The design produces a validation plan rather than an invented universal threshold. Release evidence must show how each alert and dosing rule performs across clinically relevant ranges and failure modes, and must trace confirmation prompts to the approved labeling for the selected device.

Key Insight: Healthcare IoT systems must design around sensor limitations without over-interpreting aggregate metrics. MARD is useful for comparing average performance under stated study conditions; it is not a worst-case error, a confidence interval, or a recipe for clinical thresholds.

Back to the chapter