Canine scent detection: Lung cancer target acquisition

A growing body of research has developed over the last 30 years exploring disease scent detection using animals. Research project methodologies and results in this field vary significantly, including some which are quite promising. Critiques of aspects of current and past practice in disease scent detection using animals inform recommendations for the development of scientifically robust standard operating procedures in this field. Greater standardisation and transparency of practice has the potential to strengthen and clarify disease scent detection results. My primary aim was to contribute to this standardisation by using a theoretical understanding of concept formation and learning acquisition to inform three experiments in which pet dogs (Canis lupus familiaris) were trained to detect the presence of lung cancer in human breath and saliva samples. Operant conditioning processes and an automated apparatus were used to train and test the dogs. Data for all three experiments were collected concurrently. The experiments involved: designing a process for evaluating sample comparisons; evaluating the efficacy of sample re-use for training purposes; and developing a mathematical model to support decision-making about the transition from training to testing of detector organisms. Firstly, we evaluated the comparative utility of human breath and saliva samples to train and test dogs for lung cancer detection. Signal detection measures were used to gauge the dogs’ target acquisition and concept formation during the training process. The dogs acquired the lung cancer target concept more quickly from breath samples, but also demonstrated higher-than-chance recognition using saliva samples. Secondly, we systematically evaluated the effect of breath sample re-use on dogs’ performance during lung cancer scent detection training. There were no significant changes associated with the detectability of the target across samples re-used up to four times, and observed changes in performance were small. Finally, we explored methods of evaluating when a detection animal is performing at or near the highest accuracy of which they are capable with a view to identifying the optimal point at which to transition from training to testing. A quantitative model was most informative during our work training dogs to detect lung cancer. Ongoing testing of the dogs’ abilities using novel samples occurred during training. However, the final testing (and intermittent training) needed to measure and maintain the dogs’ performance against asymptotic predictions was beyond of the scope of the current project. Notwithstanding this, each of the experiments described herein provides both specific data on the performance of our dogs, and procedural information about ways in which different components of scent detection methodology using detector organisms could be strengthened. Both of these resources could be used in concert with other methods of scent detection. Relevant theory on concept formation was reviewed and used to plan and interpret the acquisition of the target scent. Likewise, theoretical explanations of target acquisition and associative learning informed our analysis of data. In this way, each of these experiments contribute to improving practice in the field of disease detection by providing model procedures from which other methods for evaluating sample types and the reuse of samples could evolve. The development of a standardised measure for determining when to stop training and start testing a detecting organism might also be applied to a wide range of learners in myriad contexts to good effect.
The University of Waikato
All items in Research Commons are provided for private study and research purposes and are protected by copyright with all rights reserved unless otherwise indicated.
Publisher version