Tremors are one of the most common movement disorders primarily associated with various neurological diseases. Since there are more than 20 different types of tremors, differentiation between them is important from the treatment point of view. In the thesis, we focus on differentiation between three of the most common tremors: Parkinsonian, essential and mixed type of tremor.
Our first goal was to build a diagnostic model for distinguishing between Parkinsonian, essential and mixed type of tremors, based on clinical examination data, family history and digital spirography. The process of building a model was carried out using argument-based machine learning which enabled us to build a decision model through the process of knowledge elicitation from the domain expert (in our case from a neurologist). The obtained model consists of thirteen rules that are medically sensible. The process of knowledge elicitation itself contributed to the higher classification accuracy of the final model in comparison with the initial one.
In the final diagnostic model, attributes derived from the spirography were included in more than half of the rules. This motivated us to build a model based solely on the digital spirography data. For the needs of constructing an understandable model, we first built several attributes which represented domain medical knowledge. We have built more than 500 different attributes which were used in a logistic regression to construct the final diagnostic model. The model is able to distinguish subjects with tremors from those without tremors with 90% classification accuracy.
During the process of attribute construction, we wanted to know what our attributes were detecting. Thus, we have developed a method for attribute visualisation on series. The method not only helped us with attribute construction, but it is also useful for visual interpretation of the diagnostic model's decisions. The visualisation method and consequently the decision model were evaluated with the help of three independent neurology experts. The results show that both the diagnostic model and the visualisation are meaningful and cover medical knowledge of the domain. The final diagnostic model is built into the freely available ParkinsonCheck mobile application.