CGM Device Guide
Roche Accu-Chek SmartGuide
9.32pm, midweek. The SmartGuide app lights up before bed: raised probability of glucose under 3.9 between now and the small hours. A small carbohydrate top-up, an alarm set with a wider safety net, and a settled night. The same app gave a 30-minute heads-up earlier in the day, and a two-hour forecast curve before an evening meal that did not quite behave. The SmartGuide is the only CGM in this cluster carrying three prediction layers as a labelled feature, and the overnight one is the layer most people name first.
Ask Grace
Want to ask how the three prediction layers work in practice, or what the overnight risk score does and does not see? Ask Grace.
Framework status
Data sufficiency: Met. Mader et al. 2024, Journal of Diabetes Science and Technology, n=48, 139 sensors analysed across three German and Austrian sites. 48 participants sits at the lower bound of the threshold; the three-sensor-per-participant design supports the analysis.
Device specifications
The headline numbers sit in the framework card above. The full device profile is here for anyone who wants the detail.
Full device profile
Manufacturer: Roche Diabetes Care GmbH, Mannheim. Family: Accu-Chek SmartGuide CGM device, SmartGuide app, SmartGuide Predict app, AC Care HCP platform. CE marked: July 2024, MDR class IIa. UKCA non-adjunctive: MHRA, May 2025, 18 years and older. Wear duration: up to 14 days. Application site: back of the upper arm. Sensor profile: 5.9 mm height with adhesive, 5 g, 33.3 mm diameter on skin (Glatzer 2024). Reading interval: every 5 minutes via Bluetooth Low Energy. Run-in: one hour. Calibration: two paired finger-prick calibrations on day one (around 12 and 14 hours after insertion); factory-style operation thereafter. Display: Accu-Chek SmartGuide app on iOS and Android (no dedicated reader). Indication: non-adjunctive (no confirmatory finger-prick required for insulin dosing once calibration is complete). Age indication: 18 years and older. Hospital use: not intended.
The Predict app, separately. A second app, the Accu-Chek SmartGuide Predict app, is CE-marked MDR class IIa medical software in its own right. It carries the three prediction layers (30-minute, two-hour, overnight) and runs as cloud-supported inference on iOS and Android. The CGM app and the Predict app are distinct downloads with shared data flow.
Accuracy
Mader and colleagues (2024, Journal of Diabetes Science and Technology) reported the SmartGuide pivotal in-clinic accuracy: 90.5% within ±20/20, 99.7% within Parkes consensus error grid zones A and B, and 1.2% outside the ±40/40 safety band. The hypoglycaemia-range agreement (94.3% within ±20/20 below 3.9 mmol/L) is the strongest sub-range performance and a meaningful differentiator, since many CGM systems perform worst in the low range. The reported positive bias of approximately 11% below 3.0 mmol/L is worth knowing: a sensor reading of 4.5 may correspond to a slightly lower true value, so hypoglycaemia thresholds should not be relaxed on the assumption that the sensor over-reads. Cross-CGM agreement comparisons need a comparator-equivalence note: the SmartGuide pivotal used capillary blood (Accu-Chek Guide), Dexcom and Libre pivotals used venous YSI; comparator choice can shift apparent agreement by several percentage points. The full thesis lives on the accuracy page.
Three prediction layers
Three independent machine-learning models packaged into the Predict app (Herrero et al. 2024, Journal of Diabetes Science and Technology): a 30-minute hypo classifier, a two-hour glucose curve, and an overnight risk score. The framing matters. None of these prevents an event; each gives an earlier window to act.
Predicts the night, no AID partnership. The overnight risk score is the only one of its kind in this cluster. The trade-off the page sits on is between proactive overnight protection and access to algorithm-driven insulin delivery: SmartGuide does not yet pair with Omnipod 5, Tandem Control-IQ, CamAPS FX, or the MiniMed 780G. The prediction layers do not see exercise, alcohol, illness, or future meals or insulin doses; they are reactive to the data the user logs.
What the layers do not do. They do not see exercise, alcohol, or illness. They do not anticipate meals or insulin doses the user has not entered. The Night Low Predict model misses approximately 45% of overnight hypoglycaemia events at the published threshold, which is why the 30-minute Low Glucose Predict alarm is the active safety net during the night, not a substitute. Honest framing, in line with Roche’s own published descriptions: these layers reduce the likelihood of a low and give a window to act before one. They do not prevent.
AID system compatibility
The SmartGuide does not currently pair with any AID system. For people on or considering Omnipod 5, Tandem Control-IQ, CamAPS FX, or the MiniMed 780G, the device choice is currently driven by the AID system first. The CGM Guide hub carries the AID compatibility overview across the cluster.
What SmartGuide brings beyond accuracy
Strong hypoglycaemia-range accuracy
The 94.3% within ±20/20 below 3.9 mmol/L (Mader 2024) is the strongest sub-range performance of any device in the cluster, and is unusual: many CGMs perform worst in hypoglycaemia. This sits well with the prediction layers, the device whose accuracy is best in the low range is also the one labelling overnight low risk before bed.
Mandatory day-one calibration
Two paired finger-prick calibrations on day one, taken around 12 and 14 hours after insertion. This is the structural difference from Dexcom and Libre, which are factory-calibrated end-to-end. The trade-off is real-world: an extra-step at start-of-wear in exchange for the day-one calibration anchor that supports the accuracy profile. After the first day, no further finger-pricks are needed.
AC Care, into the clinical record
The AC Care platform produces AGP-standard outputs (Time in Range, Glucose Management Indicator, coefficient of variation) for clinical review. As CGM grows in primary care for insulin-treated type 2 diabetes, having the data in a recognisable AGP format inside the diabetes review matters for audit and the way reviews are run.
End-of-wear stability holds
Mader 2024 reports 85.9% within ±20/20 on days 13 to 14, compared to 92.8% on day 2. A roughly seven-percentage-point drop end-of-wear, comparable to other 14-day sensors, supports the labelled wear duration without surprise.
Adults only, by labelling
The CE mark and UKCA non-adjunctive indication cover adults 18 years and older. Paediatric labelling has not been pursued in the published evidence. This is a structural gap in the family / paediatric / adolescent space that the CGMs higher up the cluster (Dexcom, Libre) do not have.
Survive and Thrive, SmartGuide
A one-page A4 resource for the first two weeks on the SmartGuide is planned to land alongside the upcoming Roche podcast episode. Sensor placement, day-one calibration, the three prediction layers, and what to do when the overnight risk score lights up.
Device 3 of 5
Roche Accu-Chek SmartGuide