A genuinely mixed and, in places, quite skeptical picture
Several studies have measured diabetic alert dog accuracy against a continuous glucose monitor or blood glucose readings, and the results vary enormously. A 2016 exploratory study involving eight people with type 1 diabetes who used trained alert dogs found that the dogs alerted to only 36 percent of recorded hypoglycemic events. Because inappropriate alerts were common, the positive predictive value of a dog alert, meaning the proportion of alerts that actually corresponded to a confirmed low, was just 12 percent. When a dog and a continuous glucose monitor both alerted for the same event, the monitor generally alerted first, doing so roughly 22 minutes sooner in 73 percent of those cases. The study's authors concluded that a dog alert on its own is unlikely to reliably help someone distinguish a genuine low from a genuine high or normal reading.
A separate 2017 study analysed daily diaries from 18 people with type 1 diabetes, or parents of children with type 1 diabetes, who used alert dogs. Individual dogs' sensitivity for detecting low blood glucose ranged from 33.3 percent to 100 percent, with an average across all the dogs of around 59 percent, illustrating how differently individual dogs performed within that study. A separate 2019 narrative review pulling together the available studies found sensitivity and specificity figures ranging from roughly 29 to 80 percent and 49 to 96 percent respectively across different studies, though these figures come from studies that measured things differently, so they are not perfectly comparable to one another. A 2013 study found that trained dogs performed poorly when asked to identify skin-swab samples collected during hypoglycemia, with performance not clearly better than chance. A separate 2017 study found that most of the dogs tested did not demonstrate accurate detection of extreme low or high blood glucose levels, although performance varied and a small number of dogs performed better than chance.
Los et al., Reliability of Trained Dogs to Alert to Hypoglycemia, 2016 · Gonder-Frederick et al., Diary Study, 2017
High owner satisfaction, modest measured accuracy
One of the more interesting, and slightly uncomfortable, findings across this research is the gap between how families feel about their diabetic alert dog and what the measured accuracy actually shows. In the 2016 exploratory study, participants rated their satisfaction with their dogs at 8.9 out of 10 and their confidence in the dogs' ability to detect hypoglycemia at 7.9 out of 10. Yet the same dogs alerted to only 36 percent of recorded hypoglycemic events, and many of their alerts were not confirmed by an actual glucose measurement. This does not, by itself, establish that owners are being misled. A dog can provide real comfort, companionship and a sense of partnership in managing a difficult condition, independent of whether its detection accuracy holds up to close measurement. But it is an important reason to separate the emotional experience of having a diabetic alert dog from the specific, measurable claim that the dog reliably detects dangerous blood sugar events.
Why some dogs seem to do far better than others
Researchers looking into this variability have pointed to a few likely contributors, rather than one single explanation:
- Training differences. Published studies describe substantial variation in training procedures across programmes and individual trainers, and the field lacks a universally adopted system for independently validating diabetic alert accuracy.
- The complexity of the target scent itself. The target odours associated with glycemic changes may be complex and may vary between people or between episodes in the same person. That possibility is one reason researchers have not treated diabetic alert performance as a simple, standardised scent detection task.
- Individual dog variation. Just as with any working dog role, individual dogs genuinely differ in ability, motivation and consistency, which is part of why accuracy in the research ranges so widely from one dog to the next.
A more recent scoping review of this field explicitly names training protocol inconsistency and a lack of standardised outcome reporting as key gaps still holding the research back.
Scoping Review, Role of Diabetic Alert Dogs in Glycemic Monitoring
A companion to monitoring, not a replacement for it
Given this evidence, the responsible way to think about a diabetic alert dog is as a potential addition alongside proper medical glucose monitoring, not a substitute for it. Continuous glucose monitors are regulated medical devices with defined performance requirements and known limitations of their own, and the research so far does not support relying on a dog alone to catch a dangerous blood sugar event. Unlike a dog, a continuous glucose monitor provides a continuously measured glucose estimate that can be checked against established performance standards. Some studies did find real benefits reported by families, including confidence and quality of life improvements, alongside the accuracy concerns, so this is not a case of the whole idea being without value. It is a case of the idea being genuinely useful for some families in ways that go beyond pure detection accuracy, while the specific claim of reliable medical detection remains only partly supported by the evidence that actually exists.
Why we are being especially careful here
Diabetic alert work is not currently part of our own training programme, and this research is a large part of why. Unlike some of the other assistance roles we have written about, where the evidence, while still developing, points in a reasonably encouraging direction, the research on diabetic alert accuracy specifically is genuinely mixed, with credible studies finding quite low detection rates under real world conditions.
If we ever explore this direction in the future, we would want to see it built on individually verified accuracy testing for each dog, rather than an assumption that training alone guarantees reliable detection. Until the evidence here is stronger and more consistent, we think the honest position is to say plainly that this is a promising but genuinely unresolved area of work, not a settled one.
