Machine Learning Methods for Recognizing Brain Disorders

Abstract

Brain disorders represent a significant health challenge. It is estimated that approximately 165 million people suffer from a brain disorder in Europe, while 1 in 3 people will experience such a disorder during their lifetime. Some types of the brain disorders are the following: Alzheimer’s disease, dementias, epilepsy, Parkinson’s disease, Mental disorders, and more. These disorders affect the way people think, feel, or perform daily activities. However, if these disorders are diagnosed early and the person receives suitable medication, their progression may be delayed. For this reason, early diagnosis is crucial. Artificial Intelligence (AI) holds the promise of transforming how we tackle societal issues and enhancing the welfare of both individuals and communities. “AI for Social Good”, also known as “AI for Social Impact” is a new research field aiming to tackle some of the most important social, environmental, and public health challenges that exist today. Another main aim of the “AI for Social Good” is to address the United Nations Sustainable Development Goals (UNSDGs). This PhD thesis aims to contribute to this new field by developing modern machine learning methods, with a particular focus on three major categories (Depression, Alzheimer’s Dementia and Epilepsy). Depression entails a great number of symptoms, including loss of interest, anger, pessimism, changes in weight, feelings of worthlessness, thoughts of suicide, and many more. Social media are used on a daily basis by people, who express their thoughts, feelings by discussing with other users. Prior work employs transformer-based models. However, these models often cannot capture rich factual knowledge. Also, speech is a reliable biomarker for diagnosing depression, since depressed people present decreased verbal activity productivity and “lifeless” sounding speech. However, existing methods employ unimodal models, use early, intermediate, or late fusion strategies to fuse the different modalities, rely on feature extraction, and perform their approaches only in the English language. Alzheimer’s dementia is characterized by loss of memory, while it affects language and speech. Previous work utilizes speech and transcripts for recognizing dementia. However, prior work focuses on just improving the performance of proposed models, relies on feature extraction, while early and late fusion strategies are employed in terms of multimodal approaches, i.e., approaches employing both speech and transcripts. Epilepsy and seizures entail social stigma. Existing works rely on extraction of handcrafted features from electroencephalography (EEG) or dividing the EEG signals into multiple sub-signals and exploiting majority vote approaches. This PhD thesis is the first to systematically investigate various methods for identifying (i) depression by utilizing posts in social media and spontaneous speech, (ii) AD patients and predicting their Mini Mental State Examination scores through spontaneous speech, (iii) epilepsy through single-channel EEG signals. The key contributions of our work are the following: First, we introduce two methods for identifying depression. Regarding the first approach, we present the task of predicting depression in social media and propose a method for injecting external linguistic information into novel pretrained neural language models (e.g. BERT). We show that incorporating linguistic features is beneficial to depression recognition task. In terms of the second approach, we introduce a method which identifies depression based on speech and automatic transcripts. Secondly, for identifying dementia, we fine-tune language models based on transformers and present explainable approaches and linguistic analyses to investigate differences in language between healthy and AD patients. Thirdly, we introduce methods for fusing the different modalities (speech, text), calibrating the proposed models, enhancing the self-attention networks with contextual information, and automatically generating Convolutional Neural Network architectures (Neural Architecture Search). Finally, we present a multimodal approach for detecting epilepsy by exploiting single – channel EEG signals. All experiments are conducted on publicly available datasets. This PhD thesis represents a first, fundamental step among other recent efforts towards improving the performance of automatic systems aiming at recognizing various brain disorders using modern deep learning techniques. This thesis further advances the application of new technologies and sheds light on the emerging fields of text, speech, image and signal processing.