Investigation of AI-based Image, Video, and Voice Analysis to Assess Clinical Symptoms

The rapid advancement of artificial intelligence (AI) is transforming healthcare, offering the potential to enhance diagnostic accuracy, streamline clinical workflows, and personalize treatment plans. However, the comprehensive application and integration of AI technologies in healthcare face challenges, particularly in enhancing non-invasive screening methods. This thesis investigates the application of AI-assisted tools across three key modalities—video, voice, and image—to improve clinical decision-making and patient outcomes through non-invasive methods. Focusing on neurological conditions such as Parkinson's disease, stroke, and Amyotrophic Lateral Sclerosis (ALS), as well as ophthalmology and wound care, the research is guided by three main questions. While the first two research questions leverage video and voice analysis to detect subtle neurological symptoms in Parkinson’s disease, stroke, and ALS—addressing key challenges of non-invasive diagnostics such as subjective clinical assessments, delayed timeliness, and limited patient monitoring—the third question aims to enhance AI non-invasive methods in ophthalmology and wound care by overcoming data scarcity and advancing image translation techniques. The three questions are: 1. How can AI-assisted facial expression analysis enhance the detection and understanding of neurological conditions such as Parkinson's disease, stroke, and ALS? The study demonstrated that AI-assisted facial expression models could detect subtle symptoms of these disorders, achieving 83% accuracy in identifying hypomimia associated with Parkinson's disease. Similar techniques effectively detected facial weaknesses in Post-Stroke and ALS patients, highlighting the value of AI-driven video analysis for non-invasive assessments. This approach offers a groundbreaking non-invasive way to identify subtle symptoms that might otherwise go unnoticed. Additionally, an AI-driven stroke app can assist in screening cases with just a smile in emergency departments, highlighting the potential of video analysis for rapid and non-invasive assessments. 2. In what ways can AI-based voice analysis tools improve the remote assessment of Parkinson's disease severity and support ongoing monitoring? This work integrate large language models (LLMs) into a chatbot for voice assessment, enabling scalable, remote Parkinson’s disease monitoring. This innovation allowed AI-based voice analysis to accurately categorize symptoms with a classification accuracy of 72%, using vocal biomarkers from phoneme tasks. The chatbot-guided assessments made early detection and continuous care more accessible, particularly in underserved regions. 3. How do AI-powered synthetic imaging techniques contribute to the detection and diagnosis of medical conditions like age-related macular degeneration and venous leg ulcers? In imaging, deep learning models such as StyleGAN-2 achieved 85% accuracy in detecting age-related macular degeneration, outperforming human experts. Additionally, AI-generated thermal imaging achieved promising results for chronic wound assessment with an SSIM score of 0.84, although further validation is necessary. In conclusion, this thesis underscores the transformative potential of AI in healthcare, providing non-invasive solutions that improve early detection, facilitate remote monitoring, and enhance diagnostic precision. Future efforts must address demographic biases, ensure ethical data use, and work with regulatory bodies to integrate these tools into clinical practice, advancing towards more accessible and effective healthcare solutions.<p></p>