Tuesday, November 26, 2013

Do Dogs Detect Hypoglycemia? Two Studies Say Yes, but a Third Raises Serious Doubts

Three UK scientists, one of which is the CEO of a dog training organization that specializes in medical detection dogs, have published a study (Rooney et al. 2013) that “points to the potential value of alert dogs, for increasing glycaemic control, client independence and consequent quality of life and even reducing the costs of long term health care.” Another preliminary study (Gonder-Frederick et al., 2013), also using self-reporting, largely confirmed the results of the first study.  A third study (Dehlinger et al., 2013), however, found that trained hypoglycemia alert dogs, when faced with swabs taken from patients during hypoglycemic periods and swabs taken during normal glycemic level periods, could not distinguish them at a level greater than chance.  It will be important for future research, and for the credibility of this kind of service dog, to determine how such disparate results can be reconciled. 

The UK Study

Summarizing prior research, Rooney et al. noted that “studies relied on owners accurately recalling past events,” meaning that “the frequency with which dogs respond may be over-reported.”  They also note that previous studies of dogs have concentrated on hypoglycemic episodes exclusively, whereas their study also examined hyperglycemia.  The subjects in this study had Type 1 diabetes.

Medical Detection Dogs, the charity of one of the authors, has trained over 20 dogs in hypoglycemia alerting. This and other charities have used trained alerting behaviors that “include licking, pawing, jumping, staring, vocalizing and even fetching a blood testing kit” when an owner’s blood sugar level falls outside a target range, usually 5-15 nm/l. (Alerting by fetching a blood testing kit is a specific set of actions that is not likely to be accidental, but it must be wondered if owners might sometimes confuse some of the other actions with non-alerting activities of a dog.  It is to be hoped that the authors will follow with an article about the training methods used in this study.)  This study sought to assess claims that using such dogs facilitates tightened glycemic control, reduces hypoglycemic episodes, nocturnal lows, and calls for paramedic assistance.   

The 16 subjects completing the study were clients of Medical Detection Dogs with trained and certified or “advanced trainee dogs,” the latter being deemed to alert sufficiently accurately to participate in the study despite a lack of certification.   Subjects had lived with a hypoglycemia alert dog as short as four months and as long as seven years.  The subjects provided detailed information about how having an alert dog had changed their lives, including the estimated frequencies of low blood sugar pre-dog and with the dog, of episodes of losing consciousness, and of paramedic calls.  All subjects reported a decrease in at least one of these categories after obtaining a trained dog.  Eight people who reported having episodes of unconsciousness before getting a dog said they did not have such episodes after getting one.  Three people reported having made paramedic calls before getting a dog but not after.  Almost all subjects (15) trusted their dogs to alert to low blood sugar levels, while 13 trusted them to alert to high blood sugar levels. The scientists explain this discrepancy from the fact that alerting to high blood sugar “is a secondary task, trained subsequent to a strong alert to low blood sugar.” 

In the second phase of the study, subjects were asked to record their dog’s alerting behavior and to provide blood test results.  The study found that blood tests for eight of ten subjects showed that a sample taken after a dog’s alert was significantly more likely to be out of target range than was a routine sample.  One dog was apparently alerting at random.  The study states that for “the best performing dog, the odds of an alert being when bloods were out of range were 10,000 times higher than that of routine tests.”  Eight subjects who recorded nocturnal lows pre-dog had fewer nocturnal lows post-dog, though two had an increase post-dog.

The authors of the study state their results with some hedging on the fact that much of their data depends on self-reporting by subjects:

“The population, overall, reported reduced unconscious episodes and paramedic call outs, which if accurate, is of great importance since not only does it represent increase health and safety of the client, but also potentially significant reduced costs in health care.”    They also note that “for 80% of the clients providing sufficient data, when their dog was recorded to perform an alerting behaviour their blood was significantly more likely to be out of target range than it was during routine samples. In addition, comparison of owner’s routine test records from before and after obtaining their dog, showed highly significant overall change: all but one client being more likely to be within target range post-dog; five out of nine clients experienced a significantly reduced incidence of low blood sugars, and three of the remaining four showed a significant reduction in high blood sugars, suggesting improved glycaemic control in most clients. The two clients who showed no significant increase in percentage within target (1 and 5), had dogs which were unqualified and the clients reported to be experiencing training problems, which were subsequently resolved.” 

As to what the dogs are actually alerting to, these scientists argue that odor cues are the most plausible explanation, particularly given that when this occurs when owners are asleep, behavioral cues would presumably be few, “although changes in breathing rate may occur.”  Also, some owners report the dogs alerting when they are in another room.  Thus, it “is likely that dogs detect changes in the chemical composition of their owners’ sweat, or breath (including products of ketosis)….” 

The Virginia Survey
A short paper appearing in Diabetes Care, Gonder-Frederick et al. also gathered data from persons with Type 1 diabetes.  The patients had received diabetic alert dogs from Service Dogs by Warren Retrievers, Inc., located in Culpeper, Virginia.  The survey asked how often respondents experienced hypoglycemia with no corresponding alert from their service dog.  More than a third (36.1%) reported no such occurrences, 27.8% reported fewer than one event per week, and 36.1% reported more than one per week. Respondents reported significant decreases of severe and moderate hypoglycemia since getting a dog, though detailed statistics were not included in the one-page summary.  Subjects reported decreased worry about hypo- and hyperglycemia, and increased participation in physical activities.  The authors of the study conclude that their preliminary results justify additional research. 

The Oregon Study

In a study that did not use self-reporting, Dehlinger et al., subjects rubbed sterile cotton swab samples by rubbing them on the skin of both arms during hypoglycemic and normal glycemic periods.  (The UK study also indicated that the dogs were “increasingly” trained “using remote odour samples collected from clients during times of hypoglycaemia” before the dog and owner were introduced.) The three dogs used in the study had been trained to press a bell after sniffing a container with a hypoglycemic swab.  The dogs were trained by Dogs Assisting Diabetics Foundation of Forest Grove, Oregon.  The owners of the dogs, and their trainer, believed the dogs were consistently able to detect hypoglycemia.  Each dog was tested with 24 samples by being presented with a sample from 30 to 45 seconds.  The overall results for each dog, and combined, are contained in the following table.

Dog 1
Dog 2
Dog 3
Percent correct, each
Percent correct, all
Sensitivity, each
Sensitivity, all
Specificity, each
Specificity, all

The researchers did not provide an explanation as to why dogs trained to alert to hypoglycemic swabs could not do so in their experimental setting.  The researchers concluded that “trained dogs were largely unable to identify skin swabs obtained from hypoglycemic T1D subjects.” They indicate that future studies should factor in behavioral effects, and should perhaps look at swabs taken only from the usual human companions of the dogs. 

I am particularly concerned about this study because a researcher with whom I occasionally work told me that he met a detection dog trainer at a diabetes conference.  Concerned that her description of how the dogs worked allowed for a Clever Hans effect, he asked to use three dogs for a quick double-blind study.  The dogs did not perform better than chance. 

How Can the Inconsistencies Be Explained? 

There are a number of ways that the inconsistencies of the studies might be explained. Chemical changes might be complex, and vary in the sweat of different owners.  Dogs trained from sweat swabs from the skin of multiple owners might, once deployed with a single owner, display different levels of recognition of hypoglycemia.  If dogs primarily or even partially recognize low or high blood sugar from behavioral changes in their owners, this could also explain some inconsistencies.  If it is ultimately verified that dogs recognize changes in blood sugar changes when their owners are asleep, behavioral recognition might be reduced to detection of changes in breathing patterns.  It is also possible that if dogs recognize changes in blood sugar from a mixture of chemical and behavioral changes, a study only looking only at chemical changes, as is true of the third paper discussed above, will not produce good results. 

There were many more dogs in the first two studies than in the third, which had only three.  It could be that a larger number of dogs in the third study would have produced more positive results.  As noted above, the researchers did not explain why dogs trained to recognize hypoglycemic swabs did not do so effectively in their study.  The research used double-blind investigators.  Were the dogs trained in an environment where cueing was possible so that in an experimental environment where it was not the results were not significant? Was there a Clever Hans effect because the dogs were responding to the patients’ behavior?  Were other controls missing from the training environment, meaning the dogs were not really trained at the level the researchers supposed?  Were the methods of collecting sweat in training and in the experiment really identical?    

Self-reporting may be more flawed than is acknowledged in the first two studies. Patients may be seeing what they want to see, or the companionship of the dogs may make patients more attentive to their own care.  The blood samples collected by subjects in the first study after their dogs alerted were significantly more likely to be out of the target range than a routine sample.  If, however, the owners recognize the change in themselves, could they be unconsciously cueing their dogs to alert, then taking the samples.  If this turns out to be the case, are the dogs, presently recognized as service dogs, really emotional support animals?  Since they are trained to perform behaviors related to the disability of their owners, they would arguably still be service dogs, but if those behaviors do not in fact correlate with the aspect of the disability the dogs are supposedly trained to react to, is the training still related to the disability?  That is a legal question that, my guess is, courts will not look forward to considering soon.


Hypoglycemia alert studies bear a resemblance to seizure alert studies in that most results to date have involved self-reporting.  As I noted in the chapter on this subject in Service and Therapy Dogs in American Society, one study that videotaped two patients with dogs in an epilepsy care unit (Ortiz and Liporace, 2005) found that seizure dogs were poor in alerting before a seizure and concluded that seizure dogs “were not as effective as previously thought in predicting seizure activity.”  These authors acknowledged the small size of their sample. Another study (Krauss, Choi, and Lesser, 2007) found that seizure-alert dogs were effective in detecting psychogenic nonepileptic seizures (PNES) but not necessarily in detecting epileptic seizures. The authors stated the cases they analyzed "show that patients with abnormal illness behaviors may seek service animals for support." 

Scientific discoveries often begin with anecdotal accounts, such as the initial reports of dogs that seemed to recognize that dark patches of skin on their owners were somehow disturbing, followed by the discovery that the spots were melanomas.  This led to the phenomenon of cancer sniffers.  Rigorous studies on cancer detection by dogs, however, have led to high levels of specificity and sensitivity of 80% or more, not the anemic 50% shown here.  For scientists to be persuaded that dogs can reliably detect hypoglycemia, there will have to be something more than has been demonstrated by survey studies, no matter how sophisticated, since these studies are inevitably subjective, sometimes little more than a distillation of multiple anecdotal reports.  

It is important that hypoglycemia alerting be verified in strict double-blind settings, and it can be expected that such conflicting results as discussed here, and the ultimately positive benefits if alerting is verified to actually occur, will encourage additional research and open the spigot for funding such research.  The law should sit on the sidelines for a time, as long as dogs that are claimed to have blood sugar level detection abilities are also trained well enough not to be a distraction to other patrons or passengers in public places and transportation.  Nevertheless, even here there should be an expectation that this type of alerting can eventually be supported by highly controlled scientific results. 

This blog was written by John Ensminger and L.E. Papet. Thanks to Dr. Tadeusz Jezierski for comments. 


Dehlinger, K., Tarnowski, K., House, J.L., Los, E., Hanavan, K., Bustamante, B., Ahmann, A.J., and Ward, W.K. (2013).  Can Trained Dogs Detect a Hypoglycemic Scent in Patients with Type 1 Diabetes?  Diabetes Care, 36, e-98-9.

Gonder-Frederick, L., Rice, P., Warren, D., Vajda, K., and Shepard, J. (2013). Diabetic Alert Dogs: A Preliminary Survey of Current Users. Diabetes Care, 36, e47.

Krauss, G.L., Choi, J.S., and Lesser, R.P. (2007).  Pseudoseizure Dogs. Neurology 68(4), 308-9.

Ortiz, R., and Liporace, J. (2005).  “Seizure-alert dogs”: Observations from an Inpatient Video/EEC Unit.” Epilepsy and Behavior, 6(4), 620-622.

Rooney, N.J., Morant, S., and Guest, C. (August 2013). Investigation into the Value of Trained Glycaemia Alert Dogs to Clients with Type 1 Diabetes.  PLOS/One 8(8), 369921. 

No comments:

Post a Comment