Objective:
Chronic Respiratory Diseases (CRD), such as asthma and Chronic Obstructive Pulmonary Disease (COPD), are incurable diseases that can result in severe respiratory problems.
If not managed properly, these can have long-term consequences that can also lead to death.
Controlling the progression of these diseases is therefore imperative to improve the quality of life of patients and avoid life-threatening events for those suffering from them.
Monitoring the frequency of symptoms is an integral part of the management effort for both asthma and COPD.
However, symptoms monitoring is currently limited to the use of a symptom diary which is unreliable and subjective, while the existing lung sound analysis systems that could overcome this limitation are cumbersome, only provide spot check, and not fully automatic.
Continuous symptoms monitoring is therefore required to overcome these limitations and help assess severity status, predict early exacerbations, and enable timely delivery of treatment and care.
This thesis proposes a wearable symptom monitoring system to improve the effectiveness of asthma and COPD management.
To enable the realisation of such a resource-constrained system, this thesis presents novel characteristic based features and low-complexity algorithms to detect wheeze and cough sounds, two of the common symptoms in asthma and COPD.

Methods:
A novel feature, SNR_e, is proposed which is shown to have a high discriminatory ability in differentiating between breath segments with and without wheezes.
This feature together with an automatic breath segmentation algorithm and a threshold classifier is used to develop a low-complexity wheeze detection algorithm.
For the cough detection, the combination of tonality index, 2nd Linear Predictive Coding (LPC) coefficient, spectral centroid, and spectral flatness is shown to have a high capability in detecting cough sounds in a recording.
These features combined using a logistic regression model is then used for the cough detection algorithm.

Results:
The discriminatory capability of SNR_e was characterised by its significance (a p-value of 1.6e-07 with an effect size of 0.9256), distance between values shown using histogram bins, and linear separability with Area Under Curve (AUC) of 0.9260.
When evaluated against other features using the same dataset and evaluation methods, SNR_e achieved the highest overall performance.
The wheeze detection algorithm employing SNR_e resulted in an average F1-score of 84.23%, better than the use of other features individually.
The cough detection algorithm achieved an average F1-score of 88.74%, which is comparable to other more complex methods reported in previous studies.
The use of these features and algorithms shows great potential in enabling long-term symptoms monitoring to improve the effectiveness in the management of asthma and COPD.