Respiratory disease diagnosis for dolphin using breath data

Respiratory disease in marine mammals evokes strong public attention as well as worthwhile scientific interest. Traditional methods for animal disease diagnosis include blood test, ultrasound, and computed tomography scan. These methods require invasive equipment to perform, and cannot be applied to free-swimming animals. Breath data, the measurement of lung inflow and outflow while breathing, can be collected from free-swimming animals in a non-invasive way, and so is less stressful for distressed animals. However, because of new features in the data, new statistical methods are required for the breath data analysis. In this thesis, we investigate one potential method for analyzing breath data. Our method begins by decomposing a raw dataset containing a sequence of breath cycles into a set of individual breath cycles. Incomplete cycles are removed from the dataset. In this research, we consider an entire breath cycle to be one unit of observation. Starting and ending points of breath cycles can be difficult to determine, and cause a large amount of variation in size and shape of breath curves. To reduce cycle to cycle variability, we apply curve registration to synchronize a set of breath cycles. Breath cycles are described using magnitude information and geometric shape information. We propose three shape models, namely, simple oval model, quadratic spline model, and piecewise linear model. Furthermore, principal component analysis is applied to the magnitude/shape descriptors to obtain main features of breath cycles. Criteria for disease diagnosis are developed by identifying key differences among these main features between healthy and unhealthy animals. The proposed methods were applied to check if two testing animals are diseased or not. The results were consistent with the status of both animals.