Reading the Whole Glucose Curve Turning Continuous Monitor Data Into a Coaching Plan

This application covers a system that takes the raw output of a continuous glucose monitor, reduces it to a small number of states, and generates a personalised pathway of food, activity and medication adjustments to move the user from a poor state to a good one. The applicant is Welldoc, the Maryland digital health company behind BlueStar. The claimed system is software and sensor together, with a machine learning model doing the interpretation.

The Problem

Diabetes care has historically been built on sparse data. A person checks their blood sugar a handful of times a day with a fingerstick, and clinical decisions about diet, exercise and medication are made on the basis of those few isolated readings, plus a laboratory average taken every few months. The specification is blunt about the consequence: any medical, dietary or lifestyle change recommended from a given reading is limited because the reading itself carries so little information.

Continuous monitors inverted that problem rather than solving it. A CGM sensor worn under the skin returns a reading every few minutes, which is hundreds of data points per day and tens of thousands over a quarter. That is far more data than a clinician can read in a consultation, and the standard summaries used to compress it, principally the three month average captured by glycated haemoglobin, throw away exactly the detail the sensor was bought to capture.

The information that gets lost is the shape of the day. Two people can post the same average while living completely different lives: one whose glucose sits quietly in the target band, another who swings between highs after meals and lows overnight and averages out in the middle. The second person is at considerably greater risk, and the average cannot tell them apart.

The commercial context makes this worse. Rising costs have squeezed access to care and increased clinician workloads, so the amount of professional attention available per patient is falling at exactly the moment the volume of data per patient is rising. Something has to do the interpreting.

What This Invention Does

The system compresses the glucose trace into two measures and then acts on them. The first is time in range, the proportion of a base period, typically 24 hours, that the user’s glucose stays inside a threshold band of roughly 70 to 180 mg/dL. The second is glucose variability, expressed either as a standard deviation or as a coefficient of variation, which is the standard deviation divided by the mean. Each is classified as good or bad, and the pair of classifications defines a starting state. Three of the four combinations are non-ideal, and the system’s job is to move the user out of them.

To do that it builds what the specification calls an optimised pathway. The system holds a set of account vectors for each user, describing self management behaviour across food, activity and medication use, and the pathway is a set of specific adjustments to those vectors: an increase in state improving habits, a decrease in state worsening ones. Which adjustments apply is determined by a machine learning model that has learned optimisation profiles mapping account vectors and user attributes onto likely state changes, so the recommendation is drawn from what has worked for comparable users rather than from a fixed rule table.

The second half of the claim deals with events rather than averages. The system detects individual excursions in the trace and characterises them, assigning a severity score based on the height of the excursion and how long the trace stays above the target value, along with the glucose level at the beginning and end of the event and the shape of the trace through it. A three hour post meal spike and a slow overnight drift produce very different scores, and the pathway is generated with those scores as inputs.

Claim 1 as filed covers the sensor, the memory, the processor and the interface as one system, and it closes the loop: after the pathway is delivered through a graphical interface, an updated trace is captured over a second period, and an updated pathway is generated from it, this time including insulin intake information. It also brings in a habit index score derived from a cohort of users sharing attributes with the individual, which is what lets the model reason about a new user from population behaviour.

Key Features

• Two state classification. Time in range and glucose variability are each reduced to a good or bad state, and the combination defines the user’s starting position.
• Account vectors for behaviour. Food, activity and medication use are held as structured vectors that the system can propose specific adjustments to.
• Severity scored events. Individual glucose excursions are detected and scored on the height of the trace and the time spent above the target value, not just counted.
• Machine learned optimisation profiles. A model trained on account vectors and user attributes generates the pathway, rather than a fixed clinical rule set.
• Cohort derived habit index. Recommendations are informed by a habit index score built from users sharing medical, metabolic and demographic attributes with the individual.
• Closed loop reassessment. A second monitoring period produces an updated trace and an updated pathway, with insulin intake information included.

Who Is Behind It

Welldoc, Inc. is a digital health company based in Columbia, Maryland, best known for BlueStar, a prescription platform for diabetes and broader cardiometabolic care that was among the first software products cleared by the US Food and Drug Administration as a digital therapeutic. The company’s business is precisely the problem this application describes: connecting device data to coaching that reaches the user between appointments.

The eight named inventors reflect that mix. Mansur Shomali is Welldoc’s chief medical officer and a practising endocrinologist, and Anand Iyer has led the company’s artificial intelligence and data strategy; the remainder of the list combines product and data science staff with researchers from a health informatics background. It is a filing written by people who have to make the recommendations defensible to clinicians as well as usable in an app.

The application is a divisional of Australian application 2024278555, which is the national phase entry of PCT international application PCT/US2021/023226, published as WO 2021/188942. That PCT claims priority from three United States provisional applications: two filed on 20 March 2020 and one on 11 January 2021.

Why It Matters

Continuous glucose monitoring has moved well beyond its original population. Sensors that were once reserved for people with type 1 diabetes on intensive insulin therapy are now sold over the counter in several markets and marketed to people with type 2 diabetes and to people with no diagnosis at all. Every one of those users generates a trace that means very little without interpretation, which puts a large and growing market behind software that does the interpreting.

The technical trend is a shift in where the value sits, from measurement to inference. Sensor accuracy is now good enough that the differentiating work happens downstream, in deciding what a pattern means and what to do about it. That is also where regulation bites, because software that recommends a medication adjustment is a regulated medical device in most jurisdictions, and a claim covering the specific method of generating the recommendation is worth more in that setting than it would be in an unregulated one.

The filing strategy reads as portfolio maintenance around a core product. Priority dates in 2020 and 2021, a PCT, an Australian national phase and now a divisional filed in 2026 indicate a company keeping claim scope open in a market where the sensor makers themselves are expanding into analytics. Divisional practice in Australia allows exactly that: a second bite at claim breadth once the shape of the competitive landscape is clearer than it was at filing.

Related Concepts

• Continuous glucose monitor – the sensor that supplies the data stream this system interprets.
• Digital therapeutics – the regulated software category Welldoc’s BlueStar helped establish.
• Coefficient of variation – the variability measure used to classify the user’s glucose state.
• Diabetes management – the clinical practice the optimised pathway is designed to support.
• Machine learning – the technique that generates the optimisation profiles from cohort data.
• Welldoc – the applicant and one of the longest running companies in prescription digital health.

---

AU 2026202087 was published in the Australian Official Journal of Patents on 9 April 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.