Mental health is a growing concern in today's society. Mental disorders such as bipolar disorder and depression affects one in four adults at some point in their life, and is one of the leading causes of disability. This greatly affects the individual quality of life as well as the world economy. There is a need for new treatment options, and with the rise of machine learning and wearable electronics over the last years it gives us a new approach. While we have new means of data collection, the data will still have to be labeled by specialists which is a resource-heavy task. The use of general models are less reliable on data collection before it can be taken into use, but suffers in terms of performance. Personalized models require a lot of data, but has better performance. An alternative to this is proposed through hybrid and user adaptive models which tries to get the performance of the personalized model with a reduced amount of data. This thesis explores the classification of mental health states based on motor activity and speech data by using different machine learning classifiers, techniques and models. We create general, personalized, hybrid and user adaptive models to classify migraine attacks based on motor activity from subjects with bipolar disorder and emotional states based on speech activity from actors. From this it is concluded that by using hybrid and user adaptive models it is possible to get close to the performance of a personalized model but with significantly less data.