• Citer

Multimodal Analysis for Health Assessment

par Rozo, Andrea
Président du jury Bontempi, Gianluca
Promoteur Hendrick, Patrick
Publication Non publié, 2024-12-10

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Healthcare refers to the collective efforts aimed at maintaining or restoring the physical, emotional, and psychological well-being of a person. These efforts include various steps, such as prevention, diagnosis, treatment, and recovery. In resource-limited settings, where access to one or more stages of healthcare is limited, alternative solutions are being explored to address these gaps. An example that highlights the need for innovative tools and techniques to improve health assessment is burn care. Burns can lead to severe complications without prompt and adequate treatment. Methods like laser Doppler imaging (LDI) are used alongside clinical assessments to diagnose the severity of the burns; however, LDI’s costs limit its use.Wearable devices and biomarkers represent other examples of alternative solutions, offering versatile and cost-effective monitoring of vital signs, particularly in resource-limited locations. However, some challenges remain regarding signal quality and interpretability of the interactions found from the collected data. For these reasons, this PhD research is centered around developing innovative and accessible approaches based on machine learning (ML) that can support and facilitate the assessment of health status under different conditions. The first goal of this thesis is to develop an accessible approach for burn assessment. A novel image-to-image translation approach based on ML is proposed to estimate LDI images from standard digital images. Three models are tested, tackling the problem from two different perspectives. The best-performing model achieved promising accuracy in estimating the LDI, depicting a low error rate. This tool is tailored mainly for areas where access to specialized burn care is limited, such as low- and middle-income countries and remote or extreme locations, like in space. By using digital images and ML techniques, this objective addresses the limited availability of LDI devices focusing on the diagnostic part of the care cycle, making burn care more accessible. The second goal of this thesis is to assess the generalization capability of ML algorithms for wearable signal quality indication. Two pre-trained ML models evaluate the quality of respiratory signals obtained with bio-impedance wearable devices from a new patient population. Data augmentation and transfer learning are used to optimize the performance of these models, improving the classification of clean and noisy signals. This objective offers valuable insights for the future development of bio-monitoring devices and data processing protocols, promoting their use in resource-limited locations. The third goal applies and compares algorithms to study the causal interactions between physiological signals. To tackle this objective, two approaches are explored. The first approach examines the linear and nonlinear interactions between neuronal activity and cerebral blood flow velocity in stroke patients using transfer entropy (TE), finding a relation between the strength and the nature of the interactions with the clinical outcomes of the patients. The second approach compares two ML algorithms to compute causal interaction measures, like Granger causality (GC) and TE, with a traditional method to estimate TE to analyze cardio-respiratory interactions during sleep, showing the potential of the ML approaches for this application. The final goal of this thesis is to propose an approach for analyzing biomarkers and questionnaire data from populations in extreme locations, such as submarines and space. This approach allows the study of the data to find patterns or trends that could be useful in predicting health deterioration. It is developed based on submariners' data, considering the limitations of datasets with missing data and outliers and preparing it to receive different input data from other populations, like astronauts. This research introduces various alternatives to address healthcare challenges and contribute to the advancement of more accessible and efficient solutions, particularly in resource-limited areas. These tools can ultimately improve health outcomes where traditional methods are not viable.