CGM Sensor Performance in the Real World

How reliably does each CGM model produce clean, accurate readings? PurpleSensor analyzes uploaded CGM exports for anomalous reading clusters -- patterns that can trigger false alarms or, worse, mask real glucose events. The charts below aggregate findings across real uploads from real users. For methodology, see the About page.

Loading sensor data...

Sensor Tier Distribution

Each sensor is rated Green, Yellow, Red, or Purple based on its contamination score. Bar height shows the percentage of sensors in that tier for each sensor type.

Green (score <1)

Yellow (1-4)

Red (4-8)

Purple (>8)

Time in Anomalous Readings

What fraction of total recorded sensor-time shows noise clusters? Elevated are medium-confidence anomalies; Critical are high-confidence.

Elevated

Critical

Post-Cluster Hypoglycemia Rate

How often do anomalous sensor reading clusters precede hypoglycemic events? A sensor masking a real glucose low -- or triggering a phantom correction -- can cause dangerous hypoglycemia. Bars show post-cluster lows per 100 sensor-days: a 30% bar means roughly one low per 3 sensor-days following a noise cluster. Total/day = all post-cluster lows. Nocturnal/day = midnight to 6 am only -- the higher-risk window when patients cannot self-rescue. Red bars exceed clinical warning thresholds.

Median Contamination Score

Each sensor scores 0-100 based on anomalous reading time, weighted for severity. Lower is better. Thresholds: Green <1 * Yellow 1-4 * Red 4-8 * Purple >8

Sensor Consistency

Each dot is one sensor session. A wide spread means the sensor behaves unpredictably -- you can't know in advance whether a given session will be clean or noisy. Tight clustering means predictable behavior, even if the median score is poor. Predictability matters: a consistently noisy sensor at least fails the same way every time. Score 0 = perfectly clean * Score 100 = fully compromised.

0 (clean)255075100 (compromised)

Sensor Session Duration

How many days did each sensor session last? Sessions ending before the expected wear period may indicate early failures, compression lows causing user restarts, or sensor errors. G7 is rated for 10 days; a 15-day extended-wear variant launched in late 2024, but the two cannot be distinguished from export data alone -- both appear as standard G7 sessions. Early terminations are flagged relative to the 10-day baseline, which means 15-day users will show inflated apparent failure rates. This is a known data limitation.