Post Stroke Impairments and Death Cause Detection Using Machine Learning

Stroke is a common disorder that, aside from having various potentially hazardous side effects, ranks as the third most prevalent cause of mortality globally. The term "stroke" refers to a condition in which blood and nutrients are cut off from the brain due to a clot or a damaged artery, causing brain damage. A stroke causes cell death because oxygen and blood are suddenly cut off to part of the brain that needs it to operate normally. The global annual incidence of new stroke cases is estimated to be over 17 million. On the other hand, knowing ahead of time if a stroke is likely to occur is crucial for receiving prompt care. Stroke risk factors have been studied in the clinic multiple times. Predictive methods are increasingly used in modern clinical decision-making to help with disease incidence or diagnosis, prognosis appraisal, and treatment selection. Nowadays, machine learning-based models are proposed for different assessments of stroke patients. The accuracy of stroke prediction and identification of post-stroke impairments could be increased using machine learning, allowing doctors to deliver preventative therapy better. This thesis explores machine learning as a potential predictive tool for identifying post-stroke impairments and death causes. The thesis begins with a detailed overview of stroke and multiple case studies illustrating the use of machine learning in various aspects of stroke analysis. The main original contribution of this thesis is developing a machine learning model for stroke type classification, post-stroke impairments detection, and death cause identification. Three separate machine learning models have been proposed to achieve this goal: one for determining the type of stroke, another for determining the kind of impairments that develops after a stroke, and a third for determining the reasons for mortality. The first machine learning-based model for stroke type identification can differentiate between ischaemic and haemorrhagic strokes. The second model is used to diagnose conditions that arise after a stroke has occurred. Dysphasia, Hemianopia, visual-spatial difficulties, and signals from the brainstem and cerebellum are all examples of these disorders. Facial, arms/hands, and legs/feet impairments are also considered. Finally, the third model can determine the multiple causes of death, such as initial stroke, recurrent ischaemic stroke, recurrent haemorrhagic stroke, pneumonia, coronary heart disease, pulmonary embolism, and vascular and non-vascular illnesses. The machine learning models in this thesis are all data-driven models, so the international stroke trial dataset has been used to empower the proposed machine learning-based models. A statistical analysis of the International Stroke Trial dataset is performed initially. Then the features' engineering process is executed, leading to the acquisition of the core features required for training the proposed three machine learning-based models. The proposed approach in this thesis delivers promising results regarding stroke analysis, and to make an accurate diagnosis and useful recommendations, it can be useful to ascertain the presence, nature, and severity of cognitive dysfunction. This information can also serve as a baseline for monitoring future changes in cognitive abilities, mood, and personality, as well as treatment effects. Additionally, it might be helpful to comprehend the nature of neuropsychological assessments and choose the kind of test that will provide pertinent data for subsequent planning.