This thesis investigates the potential of non-invasive detection of knee Osteoarthritis (OA) using the sounds emitted by the knee joint during walking and captured by a single microphone. This is a novel application since, until now, there are no other methods that considered this type of signals. Clinical detection of knee OA relies on imaging techniques such as X-radiology and Magnetic Resonance Imaging. Some of these methods are expensive and impractical while others pose health risks due to radiation. Knee sounds on the other hand may offer a quick, practical and cost-effective alternative for the detection of the disease.

In this thesis, the knee sound signal structure is investigated using signal processing methods for information extraction from the time, frequency, cepstral and modulation domains. Feature representations are obtained and their discriminant properties are studied using statistical methods such as the Bhattacharyya distance and supervised learning techniques such as Support Vector Machine. From this work, a statistical feature parameterisation is proposed and its efficacy for the task of healthy vs OA knee condition classification is investigated using a comprehensive experimental framework proposed in this thesis.

Feature-based representations that incorporate spatiotemporal information using gait pattern variables, were also investigated for classification. Using the waveform characteristics of the acoustic pulse events detected in the signal, such representations are proposed and evaluated. This approach utilised a novel stride detection and segmentation algorithm that is based on dynamic programming and is also proposed in the thesis. This algorithm opens up potential applications in other research fields such as gait analysis.