AimsDomestic dogs are trained to a wide variety of roles including an increasing number of medical assistance tasks. Glycaemia alert dogs are reported to greatly improve the quality of life of owners living with Type 1 diabetes. Research into their value is currently sparse, on small numbers of dogs and provides conflicting results. In this study we assess the reliability of a large number of trained glycaemic alert dogs at responding to hypo- and hyper-glycaemic (referred to as out-of-range, OOR) episodes, and explore factors associated with variations in their performance.MethodsRoutine owner records were used to assess the sensitivity and specificity of each of 27 dogs, trained by a single UK charity during almost 4000 out-of-range episodes. Sensitivity and positive predictive values are compared to demographic factors and instructors’ ratings of the dog, owner and partnership.ResultsDogs varied in their performance, with median sensitivity to out-of-range episodes at 70% (25th percentile = 50, 75th percentile = 95). To hypoglycaemic episodes the median sensitivity was 83% (66–94%) while to hyperglyaemic episodes it was 67% (17–91%). The median positive predictive value (PPV) was 81% (68–94%), i.e. on average 81% of alerts occurred when glucose levels were out of target range. For four dogs, PPV was 100%. Individual characteristics of the dog, the partnership and the household were significantly associated with performance (e.g., whether the dog was previously a pet, when it was trained, whether its partner was an adult or child).ConclusionsThe large sample shows that the individual performance of dogs is variable, but overall their sensitivity and specificity to OOR episodes are better than previous studies suggest. Results show that optimal performance of glycaemic alert dogs depends not only on good initial and ongoing training, but also careful selection of dogs for the conditions in which they will be working.