Hyperparameter Optimization with Evolutionary Computing for Mental Health State Prediction

Abstract

Mental health is an important part of how we perceive our state of well-being. It is this state that help us tackle the positive and negative obstacles that life bring. When bad mental health becomes a problem, it can affect the daily task one might have as well as the relationships with the people around in a negative way, which can then cause a reduced quality of life. Depression is the leading cause of disability by the global burden of disease with suicides as the second leading cause of death among people between the age of 15 and 29. With an ever increase in technological devices, we rely more and more on technology to help us in our daily life. Machine learning has become bigger in recent years and proven to be applicable to a lot of different problem areas, as well as in mental health research. In this thesis we explore the extent of predicting migraine attacks from motor activity data and recognizing emotions from audio speech data with the use of machine learning models decision tree, random forest and support vector machine. One challenging aspect of using machine learning models is that they rely on their own set of parameters that need manual tuning before any classification can precede. From this, we will be using grid search optimization and evolutionary algorithm search optimization to find good hyperparameter values that could potentially lead to an increase in classification score. From our best results, we end up with a score of sensitivity/specificity of 75.1%/74.1% for migraine predictions and 41.7%/91% for emotions recognition after hyperparameter optimization. Random forest classifier proved to be the best overall classification model for both data types. This indicates that the choice of hyperparameter values can greatly effect the model performance and produce a better classification.