Background: A method for non-invasive, instantaneous, and cost-effective estimation of glucose level for diabetic patients, from their voice, is being developed to alleviate the issues in the estimation of blood glucose level. [1] It takes speech as an input and it evaluates another variable of interest and various error metrics can be used to evaluate the biomarker (accuracy, regression coefficient, unweighted average recall and other specific ones). We propose a clinical-based assessment called parkes error grid (PEG) to evaluate the performance.

Objectives: The main objectives of this thesis are to empirically obtain the performance metrics of different classifiers. To incorporate parkers error gird to categorical and numerical cases and obtain the PEG results. To compare the performance metrics and check which classifier gives better accuracy. To statistically compare parkes error grid error measure with statistical error measure. [1] To statistically compare parkes error grid error measure with unweighted average recall error measure and check whether the two metrics exhibit contradictions.

Methods: Through the course of this project, we use a pipeline that does feature extraction feature selection and classification for a commercial database which includes the voice of three diabetic patients. We later use two methods the numerical method and the categorical method to obtain further performance metrics and PEG results. For numeric predictions, we run the pipeline with different machine learning numerical regression classifiers and incorporate parkes error grid (PEG) to the output predictions obtained. We also obtain different statistical errors for the model.For categorical predictions, we run the pipeline with different machine learning categorical classifiers and we device an analogous parkes error grid (PEG). We also obtain different performance metrics and error metrics.

Result: The results obtained from applying the various supervised machine learning techniques are explained. We compare parkes error grid (PEG) output to the statistical error output and see which best error measure is to be considered for the given pipeline which runs on medical data. Different machine learning algorithms that are used in the thesis are assessed based on the output so that the best algorithms could be used for similar biomarkers that work with glucose analysis.

Conclusions: We conclude that it is good to use parkes error grid (PEG) more than the statistical error metric. We also see that the support vector machine is a good machine learning algorithm for numerical prediction and Navis Bayes is a good machine learning algorithm for categorical prediction based on clinically assessed parkers error grid and unweighted average recall.

Keywords: machine learning, parkers error grid (PEG), supervised machine learning.