Evaluating the effect of intermittent reinforcement on concept learning in canine lung cancer detection

Abstract

Dogs have demonstrated the ability to identify a range of human diseases, including lung cancer, through olfactory analysis of biological samples. Although exhaled breath is a non-invasive and accessible sample type, no single volatile organic compound has been reliably identified as a biomarker. Therefore, the process of identifying lung cancer in human implies a process of concept learning, suggesting dogs’ detection of lung cancer may be relying on the subjects' ability to identify a complex, highly variable pattern of volatile organic compounds (VOCs) in exhaled breath. Concept learning appears to be a special form of generalization, with underlying mechanisms in common with those responsible for perceptual concept learning in the visual domain. While intermittent reinforcement is commonly recommended in scent-detection training to simulate and prepare for operational conditions where reinforcement is not always possible, its effects on conceptual generalization remain poorly understood. This study investigated the effects of intermittent reinforcement on canine concept learning in a lung cancer detection task. Five dogs were trained using a fully automated 17-segment carousel apparatus with breath samples collected from 348 patients who visited respiratory clinic, with 115 tested positive and 233 tested negative in lung cancer. A single-subject reversal design was employed; the reinforcement rate for correct indications to positive samples was systematically thinned from 100% down to ranges of 80% and 60%. The findings demonstrated that thinning the reinforcement schedule to a minimum of 60% did not exert a significant disruptive effect on the dogs' diagnostic accuracy. In addition, an exploratory probe test also provided preliminary evidence that the dogs could successfully differentiate between lung-originated cancer and non-lung-originated (NLO) cancer samples.