Performance
→→

8 independent studies. 3,000+ participants.
70M+ real-world glucose values.

Can complications be predicted before they happen? Eight independent evaluations asked. The answer is yes. We started at 60 minutes, peer-reviewed, and have been extending the horizon since, with the accuracy to act on it. Predict. Prioritize. Prevent.

Peer-reviewed · JMIR
What continuous monitoring delivers · peer-reviewed studies & CMS-modeled income 49% fewer acute diabetes hospitalizations RELIEF · Diabetes Care 2021 +3.6 hrs in range daily MOBILE · JAMA 2021 250% per-patient income modeled · CMS fee schedule
Sharpest overnight. Above 90% out to hour 11.
Prediction accuracy by forecast horizon · tested on stable Type 1 patients · “accurate” = within 10% of the real glucose value · measured at every hour, not extrapolated · internal validation record, full metrics in the investor brief, NDA required
Optimal · nocturnal event window (~1–9 h, during sleep)
Algorithm viability · uncertainty widens
90% accuracy floor · error is tightest (± 1.7%) overnight and widens the further out we forecast Other methods lose accuracy past ~30 minutes, off the left edge of this chart
Hours 1–9 are where it counts: the overnight window, tightest predictions, severe events flagged during sleep. Past hour 11 the curve slips below the 90% floor, shown honestly out to 13 hours, where accuracy falls to 78%.
86% · 97%
Sensitivity · Specificity
85.7% sensitivity · 96.6% specificity across 7.7M real-world forecasts.
Real-world forecasts
Lows, overnight
Seen before they happen
Severe overnight events flagged during sleep, caught before morning.
Optimal window · during sleep
86.9%
Parkes Zone A at 30 minutes
Clinical-grade point accuracy, peer-reviewed. 70.6% at 60 minutes.
JMIR Diabetes 2020
800+ patients
Tested, incl. public benchmarks
From blind pediatric testing to the OhioT1DM benchmark and T1D Exchange registries. Every cohort independent.
Multi-study · public benchmarks · ages 7–80
The evidence
In-silico simulation→Blind testing→Real-world evidence
The same engine, scored independently at every step. Cohorts overlap across studies; each row is an independent evaluation.
Model accuracy independently verified by the National Research Council of Canada under a Data Analytics Centre research partnership. The full verification report ships with the evidence brief.
01Prediction testing
What is being predicted: future glucose values, hours ahead, scored against what the sensor actually recorded. The final row scores event prediction: will this patient cross into a danger range, yes or no.
StudyPatientsData pointsResult
Real-world, live clinical14210.4MFull results in the investor brief · NDA required
Real-world 1-hour forecasts vs ADA ranges1007.7M94.4% accuracy · 85.7 / 96.6 sens / spec
OhioT1DM public benchmark12401,91693.6% accuracy · 83.6 / 96.1 sens / spec
Performance at scale (Metatron)41233.7MAccuracy held at population scale
External registries · Weinstock + Chase40011.2M~90% accuracy held on the noisiest outside cohorts, ages 7 to 80
BC Children's Hospital · blind80.17MFull results in the investor brief · NDA required
In-silico (Epsilon Group)30015.5M~98% in Parkes A+B, the clinically safe zones
Latest production model · binary event prediction, ADA ranges14510.5M93% sensitivity · 98.4% specificity (most recent evaluation)
Presented at ADA Scientific Sessions · posters, 78th · 79th · 80th
02Auto-triage & prioritization testingThe workflow product: every patient reviewed, the urgent ones surfaced first
30 min → 3 min
CGM chart review per patient, a 10× efficiency gain in live clinical use
99.6%
Forecasts clinically acceptable (Parkes A+B) · JMIR peer-reviewed · 280 users · 6.9M values
Zero missed nights
The whole panel is re-read continuously. Urgent patients surface themselves, instead of waiting for someone to call
Give your CDCES team superpowers
Not a clinician replaced by an algorithm, or the other way round. A third capability: your specialists' judgment, with the reach to apply it to every patient at once. Endobits triages the full 1,500-patient panel in seconds so the team's hours land where they change outcomes. That reach is what drives the per-patient economics.
CDCES team + Endobits
~ 1.5 seconds
CDCES team alone
~ 5.5 months
How it works in your clinic →
03Glucotyping & stratification testingSorts patients by their glucose curves, not the average: know who is drifting toward complications, and why
90% sens · 84% acc
Dysglycemia detection: flagging the patients whose patterns need a clinician's eye
4 hours ahead
Nocturnal hypoglycemia warning at 93.5% specificity, before the patient is asleep, not after the ER. Nightly risk classification holds 99% sensitivity across 2,300+ nights, on both Dexcom and Libre data
4 glucotypes
Four recurring glucose curves: stable responders · post-meal spikers · variable/brittle · sustained-elevated, read from everyday CGM data, no glucose-tolerance test needed
2 granted patents
Glucotype classification and intervention-pathway IP. Plus 1 pending: Glucose Atlas
Corroborated in the field
Lancet D&E · staging callPrediabetes and T2D should be staged by CGM-derived metrics, the exact data layer Endobits reads and classifies
Penn Medicine · 19 RCTsGLP-1 agents differ by metabolic profile: precision prescribing needs phenotyping, not a 90-day average
Univ. of Liverpool · n=357,883Higher post-meal glucose → 69% higher Alzheimer's risk, the signal HbA1c can't see
Compared to today's standard of care
CapabilityHuman alone · standard of careHuman + Endobits · real-world testing
Forecast horizon~30 minutes, pump algorithms & apps use linear models that break down beyond that60 minutes published and peer-reviewed, longer horizons in ongoing validation
Glucose signalFingersticks: ≤12 readings / month · HbA1c: one lagging 90-day average288 CGM readings / day, analyzed continuously
Overnight lowsUnwitnessed, discovered the next morning, or in the ERFlagged up to 4 hours in advance · 93.5% specificity
Panel review~30 minutes of chart review per patient~3 minutes with automated triage, a 10× gain
Who gets care firstReactive, whoever calls or shows upA predictive priority list: who's heading for trouble, and when
30-minute comparator: pump algorithms (Medtronic 770G, Tandem Basal-IQ, linear-regression forecasting) · workflow timings from live US clinical deployment
Also exercised against ShanghaiT1DM, ABC4D, D1NAMO, Tidepool Loop, OpenAPS and simulator suites · full record available under NDA
Request the full evidence brief → Study protocols, verification reports and the complete metric record. NDA required.

We surface the signal.
The call stays yours.