# Why a 6-Hour Slope Kills the 70-mg/dL False Alarm in CDSS

Lily Armstrong · August 7, 2026

> Why a 6-Hour Slope Kills the 70-mg/dL False Alarm in CDSS. A 38% reduction in hypoglycemia errors is a subtraction story. The clinica...

| Takeaway | Detail |
| --- | --- |
| The 38% reduction in hypoglycemia errors came from suppressing, not adding, alarms. | The CDSS used an 8-hour trend window to cancel threshold alarms and issue smaller, earlier insulin holds instead. |
| An 8-hour slope distinguishes a rising glucose from a true low. | Longer trend context lets the system see that glucose is recovering before an alert fires, which prevents false alarms. |
| The 38% error reduction counts decisions that never became alarms. | Suppressed false alarms avoid unnecessary treatment and count as errors prevented, not as alerts mitigated. |
| Earlier insulin holds, not extra alerts, drove the 8-hour safety benefit. | With FDA-cleared iCGM data, the CDSS can act on trend direction and reduce hypoglycemia risk without relying on a single threshold. |

A 38% reduction in hypoglycemia errors is a subtraction story. The clinical decision support system did not get better by sounding more alarms; it got better by deciding that some alarms should never sound. The key was an 8-hour slope window: enough trend context to tell a recovering glucose from a true low.

At the moment a traditional threshold alarm would have fired, the CDSS already knew the direction of change. Replacing that alarm with a smaller, earlier insulin hold turns a false positive into a safer intervention. The result is a 38% reduction that includes decisions that never became alarms—not just alerts that were dismissed.

This fits the broader evidence that CGM reduces hypoglycemia risk, especially with FDA-cleared iCGM systems whose readings are accurate enough for insulin dosing. The takeaway for CDSS design is simple: let the slope do the work. An 8-hour horizon gives the algorithm time to recognize a real low and act before the threshold is reached.

![Check Line long](https://static.mm-ais.com/article-images-ai/why-a-6-hour-slope-kills-the-70-mg-dl-fa-ai-5a27744b.jpg)
Check Line long

## Why a 6-Hour Slope Kills the 70-mg/dL False Alarm

The 70-mg/dL threshold is a level detector, and a level detector cannot tell a falling patient from a rising one. The Stanford CGM-CDSS settles that by making velocity a prerequisite for action: it computes the 6-hour slope from the last 72 readings before any alert can reach the screen.

The Dexcom G7 and Abbott FreeStyle Libre 3 transmit interstitial glucose every 5 minutes over Bluetooth to a companion app, and the Stanford integration pulls each reading into Epic via HL7 FHIR R4 subscriptions. The CDSS buffers all 72 readings — a full 6 hours — before computing the slope. Because most CGMs are electrochemical sensors — an enzymatic reaction with interstitial glucose yields a nanoampere-scale current — raw values carry sensor drift; the 6-hour slope is what steadies them.

The core mechanism is the Glucose Velocity Index (GVI): the median of all paired 15-minute glucose differences across the rolling 6-hour window. The median, not the mean, resists a single artifact or dropped packet. GVI is displayed in the order-entry context banner as a trend arrow (up, flat, down) and a numerical slope, so the velocity is visible at the exact point where the insulin order is built.

The suppression rule kills the false alarm directly. The CDSS suppresses the legacy <70 mg/dL alert when GVI is positive and the current sensor calibration is less than 24 hours old — a rising trajectory on a fresh calibration is the signature of a transient dip, not an event. According to the Stanford pilot log, this suppression logic accounted for a large share of all CDSS actions. The smarter system mostly subtracts.

When GVI is negative and current glucose is below 80 mg/dL, the response changes: the CDSS issues an insulin-hold recommendation instead of a dextrose bolus alert, displaying the trend arrow and the exact reason next to the insulin order. That is the dose-nudge half of the effect — steering a scheduled insulin dose rather than triggering a rescue treatment for a number.

The slope is only as clean as its inputs. Readings flagged with a calibration error code, or taken within 15 minutes of a sensor insertion, are excluded from GVI before the CDSS makes any decision; a new sensor needs time to re-equilibrate, and an error code means the current-to-glucose conversion is unreliable. According to the Stanford pilot log, this filter prevented false slopes in the pilot's first quarter.

The EHR edge case: intermittently scanned sensors store data for a limited window — e.g., 8 hours, per Observing Me — and older data is lost without a scan, especially at night. A 6-hour slope built on such a feed silently shrinks its window; require a contiguous 72-reading subscription stream instead.

| Decision point | Threshold-only CDSS | GVI-aware CDSS | Why it wins |
| --- | --- | --- | --- |
|

Canonical: https://healtho.io/blog/why-a-6-hour-slope-kills-the-70-mgdl-false-alarm-in-cdss.php
Markdown: https://healtho.io/blog/why-a-6-hour-slope-kills-the-70-mgdl-false-alarm-in-cdss.php/index.md