Artificial Intelligence Systems for Microcirculation Analysis

Cardiovascular diseases are the leading cause of death and disability world- wide, with major implications on patients’ quality of life and healthcare systems. In this context, it is crucial to have effective and accessible tools for the early assessment of cardiovascular function. This thesis investigated photoplethysmography (PPG), a low-cost, non-invasive optical technique, as a tool to characterise macro- and microcirculatory alterations. The aim was twofold: to explore the use of PPG as a tool for characterising vascular ageing (VA), and to evaluate its potential for identifying microcirculatory alterations in acute conditions such as sepsis and Covid-19. The main methodological contribution was the development of two novel mathematical models for PPG waveform decomposition (multi-exponential and skew-gaussian) which enabled accurate signal reconstruction (R2 > 0.98) on multiple datasets while introducing physiologically interpretable param- eters. Additional pipelines were implemented for signal quality assessment, morphological and dynamic feature extraction, and artificial intelligence (AI)-based analysis, including a portable Raspberry Pi prototype for real- time acquisition and processing. Experimental results confirmed the potential of PPG as a cardiovascular biomarker. In VA studies, PPG decomposition features showed significant differences (p < 0.05) between different stiffness levels in vitro. In vivo, PPG- based classification between two levels of ageing achieved a mean accuracy of 87%, 83% sensitivity and 92% specificity. Regression analysis further indicated encouraging results in chronological age estimation (Mean Absolute Error of 6.8 years). In acute settings, a deep learning model identified Covid- 19 patients with 84% sensitivity and 83% specificity, while several machine learning approaches (ResNet neural network, feature-based XGboost, and a hybrid pipeline) discriminated sepsis with accuracies of up to 78%. Overall, this thesis demonstrated that PPG, integrated with advanced signal processing and AI, can provide physiologically meaningful indices of cardiovascular function. Potential applications range from risk stratification to early triage in critical illness. Important open challenges still exist related to the standardisation of acquisition protocols, the diversity of datasets, and the need for large-scale multicentre validation. However, the results obtained support the integration of PPG into multimodal and AI-based systems, open- ing up concrete prospects for more accessible, scalable, and personalised car- diovascular monitoring, even in resource-limited settings.