DiagNose.AI: A Novel Explainable Artificial Intelligence Framework for the Identification of Candida spp. from Volatile Organic Compounds using Electronic Noses

• Ano de defesa: 2025
• Autor(a) principal: BASTOS, Michael Lopes
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Pernambuco; UFPE; Brasil; Programa de Pos Graduacao em Ciencia da Computacao
• Palavras-chave em Português: Explainable Artificial Intelligence; Electronic Nose; Candida Infections
• Link de acesso: https://repositorio.ufpe.br/handle/123456789/67534

Fungal infections, especially those caused by Candida spp., represent a critical challenge in intensive care units, being associated with high mortality rates (40–60%). This scenario is aggravated by the slowness and low sensitivity (∼50%) of the current gold standard for diagnosis, the blood culture. To overcome these limitations, this thesis proposes, develops, and validates a new Framework for the identification of microorganisms based on the analysis of Volatile Organic Compounds (VOCs). This approach establishes a systematic workflow that includes: (i) the development of a protocol for experimentation and data acquisition with Elec tronic Noses (E-noses); (ii) the construction and preparation of databases of Candida VOCs, from both culture isolates and in blood broth; (iii) the application and evaluation of tradi tional and time-series classification models; and (iv) the design of a pioneering explainability (XAI) architecture based on an ensemble of techniques, aimed at ensuring the transparency of predictions. The effectiveness of the Framework was validated in the differentiation of Can dida species in different contexts, including culture and blood broth. The results attest to the robustness of the approach, with the classification models achieving accuracies of 97.46% in the culture-based approach and 98.18% in the blood broth context. In this sense, the main contribution of this thesis is the creation of a computational framework that integrates a novel explainability ensemble architecture, based on the combination of multiple methods, in order to provide consistent and multifaceted interpretations of the model’s decisions. The validation of this approach, through ablation and sensitivity studies, confirms its potential to increase confidence in the results and favor the clinical adoption of the solution. Thus, the Framework is established as a significant methodological contribution to computer science, with a direct and relevant impact on healthcare.