Sensor Calibration Algorithms and Data Fusion

Calibration is the step that turns a CGM sensor's raw electrical signal into a glucose number. As of October 2026, the Dexcom G7 and the FreeStyle Libre 3 Plus arrive calibrated from the factory [1][2], while the implanted Eversense 365 still asks for regular fingersticks [3]. Data fusion comes in wherever the software combines that signal with something else: a fingerstick value, a temperature reading, or prior statistics about how sensors typically behave. Filtering and lag are covered in Signal Processing Algorithms in CGM.

What calibration actually calculates

Most CGM sensors produce a current that rises with the glucose level, plus a small background current [4]. Calibration is the rule that converts one into the other. The simplest rule, and the one most manufacturers have used, is a straight line with two settings: sensitivity (how much current each unit of glucose produces) and baseline, the current that flows anyway [5]. Pair a single sensor reading with a time-matched blood glucose value and, if the baseline is negligible, you can calculate the sensitivity. That number then converts every reading from that moment on [4].

The trouble is that the line doesn't hold still. Sensitivity drifts, most of all on the first day, as the sensor's membrane meets the body's immune response [5]. The two numbers being matched also come from different places: the fingerstick measures blood, the sensor measures interstitial fluid (the fluid between cells), and published estimates of the time constant linking them range from 6 to 15 minutes [5]. Early systems handled both problems by brute force. The first MiniMed system needed at least four calibrations a day [4], and later commercial sensors still recalibrated every 12 to 24 hours [5].

How factory calibration works, step by step

Factory calibration moves the job to the production line. Two Abbott engineers laid out the process in 2017 [4]:

• Build sensor lots with very little sensor-to-sensor variation. The parts that matter most are the enzyme layer and the membrane that limits how fast glucose reaches it, on a sensing area smaller than 1 square millimeter.
• Pull a sample of sensors from each lot and measure how they respond to known glucose concentrations in the lab.
• Turn the lot's average sensitivity into a code. The code may include a correction for the difference between lab and body, worked out from clinical trials.
• Program the code into the sensor's electronics so the user never types it.
• Show that sensitivity doesn't change while the sensor sits on a shelf.

That only sets the starting point. To skip recalibration entirely, sensitivity also has to stay steady during wear, and the blood-to-tissue glucose relationship has to be similar from person to person [4]. Here's what surprised us: in Abbott's 72-person study, sensitivity dipped on day 1 and then held flat from day 2 through day 14, and across three lots, 80% to 92% of sensors fell within 10% of their lot's median sensitivity [4]. Abbott's original FreeStyle Libre systems were the first factory-calibrated CGMs sold [4], and their successors kept the approach through the Libre 3 Plus and Libre 2 Plus sold as of October 2026 [2]. FDA's 2017 summary for the 10-day FreeStyle Libre says it "does not require or accept any user-entered calibration" [6].

The Dexcom G6 took a middle path. Each sensor came with a code; once it was entered, no fingersticks were needed for the 10-day wear [7]. Dexcom no longer manufactures the G6 after July 1, 2026, and its transition page points G6 users to the G7 or G7 15 Day [8]. When researchers reprocessed raw G6 data from 62 participants using the assigned codes, mean absolute relative difference (MARD, the average percentage gap from a lab reference) was 9.0%, and accuracy held across all 10 days [9].

How a fingerstick calibration updates the sensor

A user calibration is data fusion in its plainest form. The algorithm pairs the meter value with the sensor current at the same moment, refits its parameters, and applies the new conversion going forward [5]. On the G6, users who skipped the code had to enter two fingersticks at startup and one every 24 hours after; those who entered it could still add manual calibrations, and FDA's calibration study found results similar to the factory codes [7].

MiniMed's 780G system surprised us. With the Guardian 4 or Simplera Sync sensor, calibration "is not required," yet the system "is designed to use every BG meter reading either entered manually or received from a linked glucose meter to calibrate the sensor" [10].

Research algorithms fuse more. Kalman filters, which weigh a model's prediction against each new measurement, have been used to estimate sensor sensitivity in real time from sparse fingersticks [5]. Bayesian calibration goes a step further: it starts from prior statistics on typical calibration settings and lets each person's fingersticks adjust them. In one study, that approach cut calibrations from two a day to one while lowering MARD by 1.2%; a later version used only two calibrations across a 7-day sensor and still beat the manufacturer's calibration [5].

Temperature is another input, and one of the factors that differ between lab testing and the body [4]. The original FreeStyle Libre's electronics measured temperature, and its components were tested for temperature sensitivity [6]. Neither source says how the algorithm uses that reading.

Which current sensors allow or require calibration

As of October 2026, FDA summaries describe these rules:

In the Eversense 365 summary's calibration-stability table, the share of readings within 20 mg/dL or 20% of the lab reference stayed between 90.5% and 96.1% in every 12-hour window out to a week after the last calibration [3]. When and how to calibrate any of these devices is set by its labeling and your care team.

What studies show about calibration error

A user calibration is only as good as the fingerstick behind it. Unwashed hands, a mistyped number or a late entry can all throw it off [4]. Worse, if the sensor signal happens to be temporarily too high or too low at the moment of calibration, the readings that follow stay off in the same direction [4]. The Abbott authors call calibration "one of the more dominant sources of sensor error" [4], and FDA credited the G6's factory calibration with avoiding "potentially erroneous SMBG calibrations," meaning faulty fingerstick entries [7].

Timing matters more than count. In a 2006 DirecNet study, 50 young people aged 10 to 18 with type 1 diabetes were monitored in a research center while an early Medtronic monitor, the CGMS, was recalibrated after the fact using different schedules [13]. Going from three calibrations to seven improved the median error only from 14% to 13%. Glucose stability mattered more: on visits where glucose changed by less than 0.5 mg/dL per minute at calibration times, median error was 13%, against 19% at 1.5 or faster [13]. Overnight accuracy improved, from 15% to 12% median error, when only the 9 p.m. and 6 a.m. calibrations were kept [13].

What we couldn't find is the math inside today's commercial algorithms. MiniMed's summary names a "zero-calibration" version of its Athena Plus algorithm [10], but the FDA summaries we read say whether calibration is allowed, not how the conversion is computed. The most detailed public account of factory calibration is still the 2017 Abbott paper [4]. For how these inputs feed dosing decisions, see Sensor Fusion in Automated Insulin Delivery (AID).