TECHNOLOGY EXPLAINER
AI and Machine Learning for Sensor Data
— the calculations that turn a signal into a reading, and how they are checked
The output of a biosensor is not yet a measurement. A current or an optical signal becomes a displayed value only after calibration, correction, outlier handling and smoothing. The FDA does not regulate AI as such; it regulates medical devices that contain AI, according to intended use and risk. What matters about the AI is less the name of the model than which data it was built on, how the data were split, and what it was compared with.
- What AI and machine learning analysis means here (the short version)
- From sensor signal to measured value — six processing steps
- A materials engineer's view (1): compensating for material variation by calculation
- Our calculation: how far does the displayed value drift as sensitivity slowly falls?
- Outliers, smoothing and prediction — how they work, and the trade-offs, on synthetic data
- How to validate — data splits, external validation, sample size
- A materials engineer's view (2): the unit of independence is the person and the lot, not the sample
- How the FDA and Japan treat it
- Cautions about overstated claims
- What is still open, and what we could not confirm
- Glossary / References / Claim-to-source audit
This article explains signal processing, AI and regulation for sensors. It does not evaluate or recommend any product or company, and it is not medical advice on how to read a measured value or on diagnosis, treatment or medication. Many of the numbers in the figures are calculations based on assumptions we chose to illustrate a mechanism; they are not the performance of any real product. Please consult a health professional about decisions concerning your health or your readings.
Sourced = stated in published material (link given)
Our calculation = a figure this article derived from stated assumptions (including calculations on synthetic data)
Not yet confirmed = a draft or plan that is not final
Readings of the processing steps and of the materials side are marked separately as Commentary. Every performance figure is given together with what it was compared with, in whom, and under what conditions.
1. What AI and machine learning analysis means here (the short version)
- What the calculation does: it turns the raw signal from a sensor (current, light intensity, impedance and so on) into a measured value or a prediction ahead in time. Here "AI and machine learning" means the part of that calculation whose coefficients or model are learned from data. Simple calibration formulas and moving averages sit in the same flow, and where "AI" begins differs from article to article and product to product (our summary).
- How it is regulated: the FDA states on its own page that it does not regulate AI as such but regulates medical devices, including AI-enabled onesSourced. A device is reviewed through 510(k), De Novo or PMA, according to its intended use and the risk that follows from its technological characteristics.
- The core of checking it: keep the test data apart from the training data, split by person (or lot), and where possible test at another site and another time. FDA guidance asks that test data be independent of the training and tuning data and generally come from multiple sites different from those used for training and tuningSourced.
"AI-powered" is not evidence of performance. The FDA's list of AI-enabled devices states that the listed devices met the applicable premarket requirementsSourced, yet a study of 692 AI/ML-enabled devices authorized from 1995 to 2023 found that only 3.6% of approvals reported race or ethnicity and only 46.1% gave comprehensive, detailed results of the performance studiesSourced. The thing to check is not whether the product is "AI" but what the figure was compared with, in whom, where, and how the data were split.
2. From sensor signal to measured value — six processing steps
Take a continuous glucose monitor (CGM) as the example. A review by Facchinetti (2016) describes a CGM as three parts: a needle-type sensor inserted under the skin, a transmitter, and a portable device that displays the value, and says the final value is obtained by converting the raw electrical signal into a concentration through a calibration procedureSourced. The same review presents a design that stacks three software modules to improve CGM output: denoising to reduce noise, enhancement to reduce systematic error, and prediction of future valuesSourced.
| Step | Purpose | Basis in the literature and regulatory documents |
|---|---|---|
| 2. Quality check, outliers | Spot transient non-physiological drops caused by pressure, and faults in the sensor and its companion devices | Facchinetti (2016) cites compression artifacts, in which external pressure causes a rapid non-physiological drop, and failures on the insulin pump side as things to detect from the continuous signal, and says such detection algorithms are mostly still prototypesSourced |
| 3. Correction | Reduce systematic error (bias) and change of sensitivity over time | The review lists as one pillar of accuracy the reduction of systematic differences and time drift caused by imperfect calibration or time variation of sensor sensitivitySourced. For blood glucose test strips, the FDA asks for an explanation of how the system compensates for differences between test strip lots (how the meter is calibrated or coded for each lot)Sourced |
| 4. Calibration | Convert the raw signal (such as current) into concentration units | One approach uses a reference measurement such as self-monitoring of blood glucoseSourced. FDA product codes include a separate category for continuous glucose monitoring systems that are "Factory Calibrated"Sourced |
| 5. Smoothing | Reduce random noise | The review's denoising module uses real-time digital filters (an example is a Kalman-filter-based smoother) to reduce the uncertainty due to measurement noiseSourced |
| 6. Prediction | Estimate values ahead of time and use them for alerts | The review gives a "simple but effective" example: a first-order autoregressive model updated by recursive least squares. More elaborate predictors can use other signals (insulin amount, exercise)Sourced |
The division into steps 2 to 6 is our summary, not the classification used in the papers or regulatory documents. Steps 3 (first half), 5 and 6 follow Facchinetti (2016) [Ref. 8]; step 3 (second half) follows the FDA guidance [Ref. 10]; step 4 (second half) follows the FDA product codes [Ref. 19].
A figure, with its conditions
The same review describes cases in which signal processing alone changed accuracySourced. Premise: this is a report from the authors' own research group in a 2016 review, about devices of that generation. It is not the accuracy of any current product.
- In data from 23 people with type 1 diabetes who wore one generation of CGM alongside frequent reference blood glucose measurements, adding signal processing alone reduced the MARD (mean absolute relative deviation from the reference) from 15.1% to 10.3% (reference: frequent reference blood glucose values).
- For another generation of device, rewriting the signal-processing code alone is described as lowering the MARD from 13% to 9% (the review cites three other papers for this; we did not check the original conditions).
MARD is an average and hides the tail of large errors. Lining up devices by MARD from different studies does not rank them, because the participants, references and methods differ.
3. A materials engineer's view (1): compensating for material variation by calculation
Disposable electrode chips and test strips differ slightly from lot to lot and unit to unit: enzyme activity, membrane thickness, electrode surface condition, print bleed. One road is to hold these differences down by materials and process alone. The other is to measure the difference and compensate for it by calculation. FDA documents describe both side by side.
- Compensating: the FDA guidance for over-the-counter (OTC) blood glucose test systems (29 September 2020) asks a 510(k) submission to explain "how the system compensates for differences between test strip lots (that is, how the meter is calibrated or coded for each lot)"Sourced.
- Holding down: the same guidance requires lot release criteria and says they should typically be tighter than the criteria used to evaluate total error in the performance studiesSourced.
- On the evaluation side: it recommends performance studies that use at least three manufacturing lots (for the precision study, at least 500 strips from at least 10 vials)Sourced, and strips that have undergone typical shipping and handlingSourced.
What we read from this (commentary): a compensation formula works only on quantities that can be measured (a lot coefficient, temperature and so on). It cannot correct factors that are not in the formula, such as interfering substances in the sample, skin condition or degradation in storage. Compensation therefore works only together with the materials and process work of keeping the variation small and measurable. When the FDA says lot release criteria should be tighter, it can also be read as asking the materials side for the small within-lot variation that the compensation assumes.
For the software of blood glucose test systems, the same guidance lists, in a table of sources of error to consider (Appendix 1), software failures such as "incorrect mathematical algorithm" and "undetected or unrecognized signal errors"Sourced. The compensation and calibration algorithm is the receptacle for material variation, and at the same time it is itself something to be verified.
4. Our calculation: how far does the displayed value drift as sensitivity slowly falls?
A hypothetical example turns the meaning of calibration and drift correction into numbers. Premises (ours): the sensor is calibrated on day 0 and the displayed value then equals the true value. Its sensitivity falls by a constant fraction every day (exponentially) and no correction is applied. The true concentration is held constant and we look at the ratio of displayed to true valueOur calculation.
If sensitivity falls at a rate k per day, exponentially, the displayed/true ratio on day t is exp(−k·t)Our calculation.
- k = 2% per day: day 7, exp(−0.14) = about 86.9%; day 14, exp(−0.28) = about 75.6%
- k = 1% per day: day 7, exp(−0.07) = about 93.2%; day 14, exp(−0.14) = about 86.9%
- With the decline rate estimated and corrected: if the true k is 2% per day and the estimate is off by 0.2% per day, the remaining error is exp(−0.002·t). On day 14, exp(−0.028) = about 97.2%
Premises and limits: there is no guarantee that real sensitivity changes exponentially, and the values of k are examples. They are not numbers for any real product. The calculation shows only the structure: the error piles up in proportion to the number of days, and the value of correction is set by the estimation error times the days. How large the estimation error is also depends on the unit, the wear site and the state of the wearer.
The special controls for integrated continuous glucose monitoring systems (iCGM) in 21 CFR 862.1355 require clinical data to show "consistent analytical and clinical performance throughout the sensor wear period"Sourced. A device whose accuracy deteriorates only at the end of the wear period cannot meet this condition. Drift correction is a process that connects directly to this requirement (commentary).
5. Outliers, smoothing and prediction — how they work, and the trade-offs, on synthetic data
We show how outlier handling, smoothing and forward prediction work, and where they hit limits, using randomly generated synthetic data. These are not measurements, and the parameters were chosen for illustration. Premises: 6 hours at 5-minute intervals; the true value is a smooth curve that rises and falls around a baseline of 1.0; measurement noise is 6% relative (normal distribution); six points are lowered by 28% to mimic about 30 minutes of pressure. The outlier rule is "detect a drop of more than 15% in 5 minutes and hold until the value returns"; smoothing is exponential with weight 0.4; prediction is a straight-line extrapolation of the last four points, 30 minutes aheadOur calculation.
- During the pressure interval: the raw readings have an error of about 0.24 (about 24%). If not removed, it goes straight into the displayed value. After holding the value and smoothing, the error is about 0.12, about half, but not zero (during the hold the value cannot follow the true movement).
- Outside the pressure interval: the raw error is about 0.058 and the smoothed error about 0.056. Noise falls, but the smoothed value lags behind the true movement, so the gain is small. Changing the weight (the strength of smoothing) swaps less noise for more lag.
- Error of the 30-minute-ahead prediction: straight-line extrapolation from raw readings gives about 0.19, and from smoothed values about 0.15. The simple prediction that "30 minutes from now it will be the same as now" (using the latest value as is) gives about 0.14 from raw readings and about 0.15 from smoothed values.
Premises and limits: the 6% noise, the shape of the curve and the rule thresholds are all assumptions. With, for example, 3% noise and a weight of 0.6, the ranking changes (we tried it). The point is not the numbers themselves but these two things.
- Smoothing trades noise for lag. The harder you apply it, the quieter the signal becomes and the later it follows changes.
- A prediction does not necessarily beat "doing nothing." When noise is large, straight-line extrapolation amplifies it. A statement about prediction performance cannot be interpreted unless it includes a comparison with a simple baseline (for example, using the latest value as is).
A review of machine learning for glucose prediction (Woldaregay et al., 2019) says that, because blood glucose dynamics are complex, "it remains difficult to achieve a universal model that produces an accurate prediction in every circumstance", and that carbohydrate estimation, quantifying the effects of exercise, stress and infection, and model portability that captures variability between and within patients are lackingSourced (the review finally analyzed 55 papers).
Research that uses machine learning to compensate for the drift of sensor elements over time is not limited to biosensors. A well-known example is a dataset in which 16 chemical sensors were used continuously for 36 months (January 2008 to February 2011) and 13,910 measurements were recorded for six gases. It was made public to support research on sensor drift compensationSourced. The data are organized in 10 batches, so the design allows a time-based split in which you train on older data and test on newer data (commentary).
This example concerns gas sensors and is not a result for biosensors. It is offered only as a reference for how to think about material aging, its correction and validation (split by time).
6. How to validate — data splits, external validation, sample size
(1) Three kinds of data: training, tuning, test
The FDA's PCCP guidance defines three kinds of data used for an AI-enabled software functionSourced.
- Training data: data used to set the model's weights and so on. They should typically be representative of the intended use population (race, ethnicity, disease severity, sex, age and so on) and of the intended environments.
- Tuning data: data used to compare a small number of trained models and to choose architecture and hyperparameters. This is part of the training process and comes before testing. The AI and ML communities sometimes call this phase "validation", but the guidance states that the FDA does not use the word "validation" in this context.
- Test data: data used to characterize performance. They are never shown to the algorithm during training and are used to estimate performance after training. They should be independent of the training and tuning data and should generally come from multiple sites different from those that produced the training and tuning data.
The word "validation" is confusing for materials engineers. "Verification and validation" of a material or device is not the same thing as the "validation set" in machine learning. The FDA avoids the clash by calling the machine-learning side "tuning" (commentary).
(2) How to split: by record or by person
Saeb et al. (2017) compared two kinds of cross-validation for studies that use smartphone or wearable sensor data to make clinical predictionsSourced. With a record-wise split, records from the same person appear in both training and testing. With a subject-wise split, people in the test set never appear in training. Using a public dataset and a simulation, they reported that record-wise cross-validation often massively overestimates prediction accuracy. A systematic review of the literature found that almost half of the studies that used accelerometers, wearable sensors or smartphones to predict clinical outcomes had used this record-wise methodSourced.
(3) Internal and external validation
Steyerberg and Harrell (2016) is a commentary arguing that prediction models need appropriate internal validation, internal-external validation and external validationSourced (taken from the title; this article does not cover the methods in detail). The FDA's draft guidance (January 2025) says reviewers need to understand whether the test data came from "completely different clinical sites" from the training data and were sequestered from the model developersSourced (draft). If development and test data are not properly separated, data leakage occurs and the true performance cannot be evaluated.
(4) What to compare with: the reference standard
The FDA draft defines a reference standard as "the best available representative truth" and asks for an explanation of how the reference standard was established and of the uncertainty inherent in itSourced (draft). In the continuous glucose monitoring rule (21 CFR 862.1355), the comparison is with values measured on "an FDA-accepted laboratory-based glucose measurement method that is precise and accurate, and that is traceable to a higher order" in specimens collected in parallelSourced.
(5) Reading the numbers: agreement rates and the lower confidence bound
The special controls in 21 CFR 862.1355 require, for adults, that agreement with the reference meet the conditions below. Each item requires that the lower one-sided 95% confidence bound, not the observed fraction, exceed the thresholdSourced. The table is an excerpt.
| Reference range | Condition of agreement with reference | Fraction the lower confidence bound must exceed |
|---|---|---|
| Below 70 mg/dL | Within ±15 mg/dL | 85% |
| 70 to 180 mg/dL | Within ±15% | 70% |
| Above 180 mg/dL | Within ±15% | 80% |
| Below 70 mg/dL | Within ±40 mg/dL | 98% |
| 70 to 180 mg/dL, and above 180 mg/dL | Within ±40% | 99% |
| Whole measuring range | Within ±20% | 87% |
All Sourced (21 CFR 862.1355(b)(1)(v), adults; excerpted by us) [Ref. 9]. This table gives the content of the regulatory special controls, not the record of any specific product. The conditions also cover rate-of-change errors, children, interfering substances and data gaps, among others.
To see what requiring a lower bound means, we put numbers on it. Premises (ours): comparison pairs are independent and agreement follows a binomial distribution; the one-sided 95% lower bound is found with the exact binomial (Clopper-Pearson) method. In a real study many pairs come from the same person, so the effective sample size is smaller than thisOur calculation.
| Number of comparison pairs n | Observed fraction needed to satisfy "within ±15%, lower bound above 70%" | Observed fraction needed to satisfy "within ±20%, lower bound above 87%" |
|---|---|---|
| 100 | 78.0% (78 pairs) | 93.0% (93 pairs) |
| 300 | 74.7% (224 pairs) | 90.3% (271 pairs) |
| 1,000 | 72.5% (725 pairs) | 88.8% (888 pairs) |
| 3,000 | 71.4% (2,142 pairs) | 88.0% (2,641 pairs) |
Our calculation. Values assuming independent pairs and the exact binomial method; not the statistical method the FDA requires.
What this shows (commentary): an observed fraction of 78% satisfies the 70% threshold with 100 pairs, but with 1,000 pairs it does not (72.5% or more is needed). A fraction given without n and a confidence interval cannot be compared. An average index such as the MARD mentioned earlier is a different view from this agreement rate by range.
(6) How to report: TRIPOD+AI
For reports of prediction-model studies, TRIPOD+AI (BMJ) was published in 2024. It updates TRIPOD 2015 as a 27-item checklist that applies whether regression or machine learning is used, and is described as superseding the 2015 checklistSourced. For a reader of a paper, it is a handle for checking whether "how the data were split" and "whether it was tested externally" are written down (commentary).
7. A materials engineer's view (2): the unit of independence is the person and the lot, not the sample
In machine learning for materials, one chip, one wire or one measurement looks like "one sample". In validating a sensor, however, what is independent is not the sample but the unit that produced it. If training and test data are mixed across the same unit, the model learns the quirks of that unit, and the score looks "good" in a way that will not carry over to a new unit (commentary).
| Unit of independence | What happens if it is mixed across training and test (commentary) | Related public documents and literature |
|---|---|---|
| Person (participant) | The model learns individual differences, skin condition and habits; a score that will not carry over to new people comes out high | Subject-wise splitting (Saeb et al., 2017)Sourced |
| Manufacturing lot | The model learns the lot-specific sensitivity or coefficient and does not fit another lot | The FDA asks for compensation between lots and evaluation with three or more lotsSourced |
| Individual sensor (each wear) | The model learns the adhesion or insertion state of each unit | Commentary (our summary). The iCGM rule requires consistency throughout the wear periodSourced |
| Site, instrument | The model learns quirks of the reference method and the environment | Test data from multiple different sites are expected (FDA PCCP guidance)Sourced |
| Time | The model learns the direction of drift over time and does not fit future data | An example of data that can be split by time (the gas sensor drift dataset)Sourced |
The first two columns and parts of the last are our summary (commentary) and not views stated by the agencies or authors. The Sourced marks in the last column show that the publication contains the corresponding statement.
The familiar practice in materials evaluation of testing across several lots to see lot-to-lot differences carries over to AI validation unchanged. If a study says it used the lot as the unit of splitting and tested on lots different from those used for training, it can be read as a check of robustness to material variation. A validation that only split samples at random says nothing about that robustness (commentary).
8. How the FDA and Japan treat it
(1) The FDA regulates the device, not "AI as such"
The FDA page "Artificial Intelligence-Enabled Medical Devices" (as displayed on 22 September 2026) states that "the FDA does not regulate AI as such; it regulates medical devices, including AI-enabled medical devices", explains that devices are reviewed through 510(k) (premarket notification), De Novo or PMA (premarket approval), and that the FDA also reviews modifications to software functions that could significantly affect safety or effectivenessSourced. The page says the FDA has authorized over 1,600 AI-enabled medical devices for marketing in the United States as of September 2026Sourced. The differences between 510(k) clearance, De Novo authorization and PMA approval are covered in our explainer on medical device approval categories and home use.
(2) What is in the list, and what is not
The FDA's "List of Artificial Intelligence-Enabled Medical Devices" had 1,614 entries as displayed on 4 September 2026Sourced. The FDA itself says the list is "not a comprehensive resource" and that devices were identified primarily from the use of AI-related terms in the summary descriptions of marketing authorization documents and/or the device classificationSourced. We tallied the list's CSV file and found the followingOur calculation.
- Pathway: 1,553 entries are 510(k), 40 are De Novo and 21 are PMA. The great majority are 510(k).
- Lead review panel: Radiology has 1,230 entries, about 76% of the list, and Cardiovascular 154, about 10%. Clinical Chemistry has 9 (about 0.6%).
- Glucose-related: of the 9 Clinical Chemistry entries, 2 fall under blood glucose test systems (product code NBW, over-the-counter blood glucose test systems) and 2 are one insulin pump therapy adjustment calculator for healthcare professionals (product code QCC; a De Novo and a later 510(k) for the same product). The rest concern urine testing and kidney disease.
- Continuous glucose monitoring systems are not in the list: the CSV contains no entry with an iCGM product code (QBJ, QLG, QDK, SAF, SBH, QDL, QII).
Not being in the list does not mean a device does not use machine learning. The list is a non-comprehensive list that mainly picks up devices whose summaries use AI-related terms, and the public summaries often do not reveal how the signal processing inside a sensor was built (commentary). Conversely, being in the list does not mean the device performs well.
(3) Updating the algorithm after marketing: the PCCP
The FDA's PCCP guidance was first issued on 4 December 2024, and the page now carries the version dated 18 August 2025Sourced. A PCCP is a document describing what modifications will be made to a device and how they will be assessed; it lets changes that would otherwise need a new premarket notification or a PMA supplement be made within a plan that has already been authorized. It has three components: the Description of Modifications, the Modification Protocol and the Impact Assessment. The Modification Protocol covers data management, re-training, performance evaluation and update proceduresSourced.
It applies to AI functions that the manufacturer intends to modify over time, whether the model is updated automatically by software ("continuous learning") or manually through human input and decisionsSourced. The guidance says a PCCP can be used in submissions through any of the 510(k), De Novo or PMA pathwaysSourced. In our tally, two entries in the FDA list carry "with PCCP" in the device nameOur calculation.
(4) The draft, the principles and transparency
- Draft guidance (January 2025): "Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations" is a draft marked "not for implementation". It sets out a comprehensive approach to risk across the product life cycle, with sections including data management, model description and development, validation and device performance monitoring. As far as we could confirm on the official page in October 2026, no final version has been issuedNot yet confirmed. The draft points out that AI-enabled devices can be sensitive to differences in input data (data drift)Sourced (draft).
- Good Machine Learning Practice (GMLP): in 2021 the FDA, Health Canada and the UK MHRA set out 10 principles, and in January 2025 the IMDRF published a final document (IMDRF/AIML WG/N88 FINAL:2025) with 10 principlesSourced. The titles in the 2021 version include "Training Data Sets Are Independent of Test Sets" (principle 4), "Selected Reference Datasets Are Based Upon Best Available Methods" (principle 5) and "Deployed Models Are Monitored for Performance and Re-training Risks Are Managed" (principle 10)Sourced.
- Transparency principles: the FDA, Health Canada and the MHRA have also published guiding principles for the transparency of machine learning-enabled devices, building on GMLP principle 7 (focus on the performance of the human-AI team) and principle 9 (users are provided clear, essential information)Sourced.
(5) Japan: confirmation of change plans (IDATEN)
Japan has had, since 2020, a system for prior confirmation of change plans according to the characteristics of the medical device (known as IDATEN). PMDA material describes its purpose as approving the improvement plan itself for devices in which future improvement is expected, and says that if changes made under a notified plan give the planned results, a partial-change approval application is not needed each time, and that it can also apply to software as a medical device (program medical devices)Sourced. The basis is the notification "Handling of applications for confirmation of change plans for medical devices" (Yakusei Kishin-hatsu 0831 No. 14, 31 August 2020), and the PMDA page also lists the Q&A on change plans for medical devices using AI-related technology and for program medical devices, and the examples of how to fill in the forms, both issued on 22 December 2023Sourced. The concept is close to the FDA's PCCP, but we have not compared the procedures or the scope of application (we could not confirm them).
9. Cautions about overstated claims
The more numbers an AI claim carries, the more conditions there are to check. The table below sorts out what to check when reading public material. It does not judge any product and is not advice to any individual reader.
| Typical wording | What to check (how to read the material) | Public documents and literature that help |
|---|---|---|
| "High accuracy thanks to AI" | Accuracy compared with what (reference method)? In whom (number, age, disease)? Where (number of sites)? Under what conditions (wear site, period)? | The iCGM rule requires the reference method, performance throughout the wear period and labeling by siteSourced |
| "FDA-cleared AI" | Which is it: 510(k) clearance, De Novo authorization or PMA approval? What is the intended use? | The FDA list is not comprehensive and being listed does not rank performanceSourced |
| "XX% accuracy" | Definition of the index (MARD, agreement rate, sensitivity/specificity) and the denominator. Is there a confidence interval? Is it an average or per range? | The rule requires a lower confidence bound (Section 6). MARD is an average indexSourced |
| "It gets smarter the more it is used" | Is it updated automatically in use? Were the permitted range of updates and their verification decided in advance? | A PCCP reviews the content, protocol and impact assessment of changes in advanceSourced |
| "Predicts the value ahead" | How many minutes ahead? Is there a comparison with a simple baseline (using the latest value, for example)? Are the situations where it tends to miss described? | A review concluding that a universal prediction model is hard (Woldaregay et al., 2019)Sourced |
| "Same accuracy in other populations" | Was external validation (another site, time or population) done? Which populations were included? | Reporting gaps in approval documents (Muralidharan et al., 2024)Sourced |
Muralidharan et al. (2024) examined the public documents of 692 AI-enabled medical devices approved or cleared by the FDA from 1995 to 2023 and reported that 3.6% of approvals reported race or ethnicity, 99.1% provided no socioeconomic data and 81.6% did not report the age of study subjects, that 46.1% provided comprehensive detailed results of performance studies, 1.9% included a link to a scientific publication and 9.0% contained a prospective study for post-market surveillanceSourced. The paper discusses this as missing information that raises the risk of algorithmic bias and health disparity. The finding concerns what the public approval documents say and does not directly show how each device was actually validated (our caution).
- A displayed value is the result of a chain of processing on the raw signal. Outlier handling, correction, calibration, smoothing and prediction offset variation and aging, and are themselves subject to verificationSourced
- Material variation is handled both by "holding it down" and by "measuring and compensating". The FDA asks for an explanation of lot-to-lot compensation and for lot release criteriaSourced
- If sensitivity falls 2% per day, the displayed value is about 75.6% on day 14. A calculation based on assumptionsOur calculation
- Smoothing trades noise for lag, and prediction needs comparison with a simple baseline. Illustrated on synthetic dataOur calculation
- Validation should split by person (lot) and test at another site. Record-wise splitting can overstate accuracySourced
- The FDA regulates the device, not AI as such. Its list had 1,614 entries (as displayed on 4 September 2026) and is not comprehensiveSourced
- Updates after marketing are bounded in advance by a PCCP. The draft life-cycle guidance is not finalNot yet confirmed
10. What is still open, and what we could not confirm
(1) Titles of the 10 principles in the final IMDRF document
We could not open the body of the IMDRF final document (N88, January 2025). The titles of the principles shown in Section 8 are therefore those of the 2021 version shown on the Health Canada page. We could not confirm whether the 2025 titles are the same.
(2) The scope of "AI", and the algorithms of individual products
The examples in this article (CGM signal processing, lot compensation of test strips) are general descriptions based on published literature and regulatory documents. The internal algorithms of individual products are often not public, and we could not confirm which processing uses machine learning.
(3) Real examples of continual learning, and a US-Japan comparison
Whether any biosensor that updates itself in use (continual learning) has been authorized under a PCCP or IDATEN is something we could not confirm (we did not investigate it). The differences in procedures and scope between Japan's IDATEN and the US PCCP are also not compared.
(4) Finalization of the draft guidance
The guidance on lifecycle management of AI-enabled device software functions is, as far as we could confirm on the official page in October 2026, still a draftNot yet confirmed. We could not confirm the content or timing of a final version.
11. Glossary
- Calibration
- The procedure that fixes the relation between the sensor's raw signal (such as current) and units like concentration.
- Drift
- A slow change of a sensor's output or sensitivity over time.
- Outlier
- A value far from the other measurements. It can arise from a transient abnormality of the sensor.
- Smoothing
- Averaging neighbouring values to reduce measurement noise. Applied strongly, it delays the response.
- Extrapolation
- Extending the trend so far to estimate a future value.
- MARD
- Mean absolute relative deviation: the average of the relative differences from the reference. The tail of large errors is hidden.
- Training, tuning and test data
- Data used to build the model, to choose it and to measure its performance. Test data must be independent of the training data.
- Data leakage
- Test information getting mixed into training so that performance looks better than it is.
- External validation
- Checking performance on data from another site, time or population than the model was built on.
- PCCP
- Predetermined change control plan. An FDA framework that reviews in advance the scope and verification procedure for updates to an AI function.
- GMLP
- Good Machine Learning Practice: ten principles for AI development in medical devices.
- iCGM
- Integrated continuous glucose monitoring system: a category intended for use with digitally connected devices such as automated insulin dosing systems.
- 510(k) / De Novo / PMA
- The FDA's three premarket pathways, described as clearance, authorization and approval respectively.
- Lower confidence bound
- The lower end of the confidence interval of an estimate: "at least this much", allowing for variability.
- IDATEN
- The common name of Japan's system for prior confirmation of change plans according to the characteristics of the medical device.
12. References
- FDA "List of Artificial Intelligence-Enabled Medical Devices" (as displayed on 4 September 2026; CSV tallied by us) — fda.gov
- FDA "Artificial Intelligence-Enabled Medical Devices" (as displayed on 22 September 2026) — fda.gov
- FDA "Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions", version of 18 August 2025 (first issued 4 December 2024), guidance page — fda.gov
- FDA "Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations", January 2025 (draft), guidance page — fda.gov
- Health Canada "Good machine learning practice for medical device development: Guiding principles" (the 2021 ten principles of the FDA, Health Canada and MHRA; page modified 29 July 2024) — canada.ca
- IMDRF "Good machine learning practice for medical device development: Guiding principles", IMDRF/AIML WG/N88 FINAL:2025 (published 29 January 2025) — imdrf.org
- FDA "Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles" — fda.gov
- Facchinetti A. "Continuous Glucose Monitoring Sensors: Past, Present and Future Algorithmic Challenges", Sensors 2016;16(12):2093 (open access) — pmc.ncbi.nlm.nih.gov
- eCFR "21 CFR 862.1355 Integrated continuous glucose monitoring system" — ecfr.gov
- FDA "Self-Monitoring Blood Glucose Test Systems for Over-the-Counter Use", guidance of 29 September 2020 (PDF) — fda.gov
- Saeb S et al. "The need to approximate the use-case in clinical machine learning", GigaScience 2017;6(5):gix019 — doi.org
- Steyerberg EW, Harrell FE "Prediction models need appropriate internal, internal-external, and external validation", J Clin Epidemiol 2016;69:245-247 — doi.org
- Collins GS et al. "TRIPOD+AI statement", BMJ 2024;385:e078378 — doi.org
- Muralidharan V et al. "A scoping review of reporting gaps in FDA-approved AI medical devices", npj Digital Medicine 2024 — doi.org
- Woldaregay AZ et al. "Data-driven modeling and prediction of blood glucose dynamics: Machine learning applications in type 1 diabetes", Artif Intell Med 2019 — doi.org
- UCI Machine Learning Repository "Gas Sensor Array Drift at Different Concentrations Dataset" (corresponding to Vergara et al., Sens Actuators B 2012;166-167:320-329) — archive.ics.uci.edu
- PMDA "Approval systems specific to medical devices" (in Japanese; includes a description of the IDATEN system) — pmda.go.jp
- PMDA Page of notifications related to the IDATEN change-plan confirmation system for medical devices (in Japanese) — pmda.go.jp
- openFDA "Device Classification" (product code QBJ as an example; other product codes are queried the same way through the API) — api.fda.gov
13. Claim-to-source audit
| Claim in the text | Basis | Label |
|---|---|---|
| "The FDA does not regulate AI as such; it regulates medical devices, including AI-enabled medical devices." The review pathways are 510(k), De Novo and PMA. The FDA says it has authorized over 1,600 AI-enabled devices as of September 2026. The page was current as of 22 September 2026 | FDA "Artificial Intelligence-Enabled Medical Devices"[Reference 2] https://www.fda.gov/medical-devices/digital-health-center-excellence/artificial-intelligence-enabled-medical-devices | Sourced |
| That the FDA list is not comprehensive and identifies devices mainly from AI-related terms in summaries or classification; that listed devices met the applicable premarket requirements; that the list had 1,614 entries as displayed on 4 September 2026 with the latest final decision on 29 June 2026 | FDA "List of Artificial Intelligence-Enabled Medical Devices"[Reference 1] https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/list-artificial-intelligence-enabled-medical-devices | Sourced |
| Counts by pathway (510(k) 1,553, De Novo 40, PMA 21), by review panel (Radiology 1,230, Cardiovascular 154, Clinical Chemistry 9 and so on) and by year. The breakdown of the 9 Clinical Chemistry entries (NBW 2, QCC 2). That the CSV has no entry with any of the seven iCGM product codes. That two entries carry "with PCCP" in the name | Our tally of the CSV in Reference 1 (classified by the submission-number prefixes K, DEN and P). The meaning of product codes was checked in the openFDA classification[Reference 19] https://api.fda.gov/device/classification.json?search=product_code:QBJ&limit=1 | Our calculation |
| That the PCCP guidance was first issued on 4 December 2024 and the current version is dated 18 August 2025; the three PCCP components and the four items of the Modification Protocol (data management, re-training, performance evaluation, update procedures); that it can be used with 510(k), De Novo and PMA; that both continuous learning and manual updates are covered; the definitions of training, tuning and test data; that the FDA does not use "validation" in this context; that test data should be independent of training and tuning data and generally from multiple different sites | FDA PCCP guidance (PDF text checked) and the guidance page[Reference 3] https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence | Sourced |
| That the lifecycle guidance for AI-enabled device software functions is a draft of January 2025 marked not for implementation; the mention of data drift; that test data need to come from different clinical sites and be sequestered from developers; the mention of data leakage; the definition of a reference standard and the request to explain it | FDA draft guidance (PDF text and page checked)[Reference 4] https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing | Sourced |
| Whether the draft has been finalized, and the content and timing of a final version | As far as we could confirm on the official page in October 2026 it remains a draft and no final version was found. A draft, so an outlook that is not settled[Reference 4] https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing | Not yet confirmed |
| The titles of the 2021 version of the GMLP ten principles (principles 4, 5, 7, 9 and 10); that the FDA, Health Canada and the MHRA published them in 2021 | Health Canada page[Reference 5] https://www.canada.ca/en/health-canada/services/drugs-health-products/medical-devices/good-machine-learning-practice-medical-device-development.html | Sourced |
| That the IMDRF issued a final document in January 2025 (IMDRF/AIML WG/N88 FINAL:2025, published 29 January 2025) | IMDRF document page. The PDF body could not be opened; only title, code and date were confirmed[Reference 6] https://www.imdrf.org/documents/good-machine-learning-practice-medical-device-development-guiding-principles | Sourced |
| That the transparency principles build on GMLP principles 7 and 9 | FDA transparency principles page[Reference 7] https://www.fda.gov/medical-devices/software-medical-device-samd/transparency-machine-learning-enabled-medical-devices-guiding-principles | Sourced |
| The three structural elements of a CGM (sensor, transmitter, portable device); that the raw signal is converted to concentration by calibration; the three software modules (denoising, enhancement, prediction); the statements on noise, drift and compression artifacts; MARD 15.1% to 10.3% (23 people, frequent reference blood glucose) and 13% to 9%; the example of an extrapolating predictor; that fault-detection algorithms are still prototypes | Facchinetti (2016)[Reference 8] https://pmc.ncbi.nlm.nih.gov/articles/PMC5191073/ | Sourced |
| The iCGM special controls (agreement conditions with the reference and the lower one-sided 95% confidence bound, the reference method, performance throughout the wear period, labeling by site, manufacturing controls) | 21 CFR 862.1355[Reference 9] https://www.ecfr.gov/current/title-21/chapter-I/subchapter-H/part-862/subpart-B/section-862.1355 | Sourced |
| That the FDA asks for an explanation of how lot-to-lot differences of test strips are compensated; that lot release criteria should be tighter than the total-error criteria; studies using three or more lots, at least 500 strips and at least 10 vials; use of strips that have undergone shipping and handling; examples of software failures (incorrect mathematical algorithm, undetected signal errors); issued 29 September 2020 | FDA guidance for OTC blood glucose test systems (PDF text checked)[Reference 10] https://www.fda.gov/media/87721/download | Sourced |
| The comparison of record-wise and subject-wise cross-validation; that record-wise validation often massively overestimates accuracy; that almost half of the wearable studies used the record-wise method | Abstract of Saeb et al. (2017) (PubMed)[Reference 11] https://doi.org/10.1093/gigascience/gix019 | Sourced |
| That prediction models need internal, internal-external and external validation (from the title) | Title of Steyerberg and Harrell (2016)[Reference 12] https://doi.org/10.1016/j.jclinepi.2015.04.005 | Sourced |
| That TRIPOD+AI is a 27-item checklist and supersedes the 2015 version | Abstract of TRIPOD+AI (BMJ 2024)[Reference 13] https://doi.org/10.1136/bmj-2023-078378 | Sourced |
| In a study of 692 approval documents: race or ethnicity reported 3.6%, no socioeconomic data 99.1%, no age 81.6%, comprehensive performance results 46.1%, link to a publication 1.9%, prospective post-market study 9.0% | Abstract of Muralidharan et al. (2024)[Reference 14] https://doi.org/10.1038/s41746-024-01270-x | Sourced |
| The review of machine learning for glucose prediction: that a universal prediction model remains difficult and what is lacking (carbohydrate estimation, exercise, stress and infection, model portability); that 55 papers were finally analyzed | Abstract of Woldaregay et al. (2019)[Reference 15] https://doi.org/10.1016/j.artmed.2019.07.007 | Sourced |
| The gas sensor drift dataset (16 chemical sensors, six gases, 36 months, 13,910 measurements, 10 batches, aimed at drift compensation research) | UCI Machine Learning Repository[Reference 16] https://archive.ics.uci.edu/dataset/270/gas+sensor+array+drift+dataset | Sourced |
| The IDATEN system: start in 2020, purpose, procedure, applicability to program medical devices; the notification number (Yakusei Kishin-hatsu 0831 No. 14, 31 August 2020); the Q&A and examples on AI-related and program medical devices of 22 December 2023 | PMDA material and PMDA notification page[Reference 17] https://www.pmda.go.jp/files/000272419.pdf[Reference 18] https://www.pmda.go.jp/review-services/drug-reviews/about-reviews/devices/0039.html | Sourced |
| Displayed/true ratios for sensitivity falling 2% (or 1%) per day, exponentially (day 7 and day 14) and the remaining error with a 0.2% per day estimation error. Root-mean-square errors on synthetic data (pressure interval 0.24 to 0.12, normal interval 0.058 to 0.056, 30-minute prediction 0.19 and 0.15, simple baseline 0.14 and 0.15). The observed fraction needed to meet the iCGM criteria (n = 100 to 3,000) | Our calculation. Premises: exponential sensitivity decline with example k values; 6% noise, 28% pressure drop, exponential smoothing 0.4 and straight-line extrapolation (synthetic data, fixed random seed); independent pairs and the exact binomial (Clopper-Pearson) method. Not the performance of any real product | Our calculation |
| The ordering of the six processing steps. The reading that compensation works only on measurable quantities. The framing that smoothing trades noise for lag and prediction needs a simple baseline. The table of units of independence. The reading of material lots as the unit of splitting. The table for checking overstated claims | Commentary and organization by this article, based on the published material. Not views stated by the companies or agencies | Commentary |
| That Figs. 1, 3, 4, 5, 6 and 7 are explanatory drawings, and that the hero image and Fig. 2 are AI-generated concepts | Note by this article | Commentary |
Last updated October 2026. Sources are limited to primary material (publications of the FDA, Health Canada, IMDRF and PMDA, 21 CFR, and peer-reviewed papers). Because the article includes readings of the processing steps and of the materials side, those are marked as commentary and kept separate from sourced fact. The titles of the ten principles in the 2025 IMDRF final document, the algorithms of individual products, authorized examples of continually learning devices, a comparison of IDATEN and PCCP procedures, and a final version of the draft guidance could not be confirmed in published primary sources and are therefore not stated. The numbers in Figs. 3 and 4 come from assumed premises and Fig. 6 is our tally; none of them shows the performance of a real product.