Artificial Intelligence (AI), particularly advancements in Machine Learning (ML) and Deep Learning (DL), has revolutionized the way clinical data are managed and processed to provide healthcare operators with decision support and context awareness. Today, AI can support clinicians in diagnosis, prognosis, surgical treatment, and follow-up, potentially providing undeniable benefits to patients. Despite its potential, the design and development of AI algorithms in healthcare present numerous challenges, including data collection biases, model generalization across diverse clinical settings, and barriers to practical implementation. These challenges may hamper the translation of the developed methodologies into actual clinical practice, posing barriers to the full exploitation of AI potentiality. To overcome these barriers, this PhD thesis builds upon the European “Ethical Guidelines for Trustworthy AI”, which outlines key principles such as fairness, robustness, privacy, and human oversight. By focusing on some of these requirements - particularly “diversity, non-discrimination, and fairness”, “technical robustness and safety”, “privacy and data governance”, and “human agency and oversight”, - the thesis proposes a pipeline that operationalizes these principles in real-world clinical applications. The work addresses three interconnected areas: 1. Dataset collection and bias mitigation: This stage addresses the identification and mitigation of dataset biases to enable fair and diverse AI models. By analyzing public datasets, the thesis highlights biases, such as minority, informativeness and selection biases. Representation learning techniques are proposed to reduce the reliance on labeled data and mitigate selection bias. 2. Privacy preservation and multicentric variability management: The second stage focuses on privacy-preserving methodologies, such as Federated Learning (FL), to address data-sharing constraints across diverse clinical datasets and manage variability across clinical datasets through domain adaptation techniques. 3. AI integration into clinical workflows: The third stage emphasizes the development of human-centric solutions into clinical workflows. By engaging clinicians in the design process and aligning with real-world practices, integrated AI systems can enhance usability, reduce operator dependency, and deliver reliable diagnostic support. These stages are demonstrated through diverse case studies, spanning from Electronic Health Records (EHRs) to Ultrasound (US) imaging and X-ray Coronary Angiography (XCA), showcasing the pipeline’s adaptability across healthcare applications. By aligning the pipeline with principles of fairness, generalizability, and human-centric design, this thesis seeks to close the gap between algorithm development in the lab and real-world application in clinical environments, offering actionable strategies for trustworthy AI in healthcare.
L’Intelligenza Artificiale (AI), con le sue applicazioni in Machine Learning (ML) e Deep Learning (DL), sta rivoluzionando il settore sanitario. Queste tecnologie forniscono strumenti avanzati per supportare i processi di diagnosi, prognosi, trattamento e follow-up, al fine di migliorare così l’assistenza e la cura del paziente. Nonostante i notevoli progressi, l’introduzione di sistemi basati su ML e DL nella pratica clinica presenta ancora sfide rilevanti, come la presenza di bias nella raccolta dei dati, la difficoltà di generalizzazione dei modelli in ambienti clinici diversificati e le problematiche di integrazione nei flussi di lavoro esistenti, che limitano il pieno potenziale dell’AI in questo ambito. Per superare tali barriere, questa tesi di dottorato trova il suo fondamento nello studio della Commissione Europea “Orientamenti etici per un’AI affidabile”, che delinea i requisiti fondamentali che i sistemi di AI dovrebbero soddisfare per essere considerati affidabili. In particolare, la tesi si concentra sui requisiti di “diversità, non discriminazione ed equità”, “riservatezza e governance dei dati”, “robustezza tecnica” e “intervento e sorveglianza umani”, adattandoli alle specificità e alle sfide dell’ambito clinico. Questo approccio ha portato alla definizione di una pipeline innovativa che rende operativi tali principi, consentendo la loro applicazione concreta in contesti clinici reali. La ricerca si articola in tre fasi interconnesse: 1. Raccolta dei dataset e mitigazione dei bias: Questa prima fase si concentra sull’identificazione e sulla mitigazione dei bias presenti nei dataset, con l’obiettivo di promuovere modelli di AI equi e rappresentativi della diversità delle popolazioni cliniche. 2. Preservazione della privacy e gestione della variabilità multicentrica: La seconda fase affronta le sfide legate alla condivisione di dati clinici sensibili, proponendo metodologie rispettose della privacy, come il Federated Learning (FL), e strategie specifiche per gestire la variabilità dei dati provenienti da più centri sanitari. 3. Integrazione dell’AI nei flussi di lavoro clinici: La terza fase si focalizza sulla progettazione di soluzioni AI che si integrino nei flussi clinici esistenti, coinvolgendo i clinici durante il processo di progettazione per garantire usabilità e allineamento alle pratiche operative quotidiane. Attraverso diversi casi studio, che spaziano dall’utilizzo di dati estratti da cartelle cliniche elettroniche fino a immagini ecografiche e angiografiche, questa tesi dimostra la flessibilità e l’applicabilità della pipeline proposta in diversi contesti clinici. Allineando tale pipeline ai principi di equità, generalizzabilità e human-centered design, questa tesi contribuisce a ridurre il divario tra lo sviluppo algoritmico in laboratorio e la sua applicazione nella pratica clinica, offrendo strategie concrete per un’AI affidabile e innovativa nel settore sanitario.
Moving AI into Clinical Practice: A Trustworthy Approach to Overcome Barriers in Real-World Implementation / DI COSMO, Mariachiara. - (2025 Mar 21).
Moving AI into Clinical Practice: A Trustworthy Approach to Overcome Barriers in Real-World Implementation
DI COSMO, MARIACHIARA
2025-03-21
Abstract
Artificial Intelligence (AI), particularly advancements in Machine Learning (ML) and Deep Learning (DL), has revolutionized the way clinical data are managed and processed to provide healthcare operators with decision support and context awareness. Today, AI can support clinicians in diagnosis, prognosis, surgical treatment, and follow-up, potentially providing undeniable benefits to patients. Despite its potential, the design and development of AI algorithms in healthcare present numerous challenges, including data collection biases, model generalization across diverse clinical settings, and barriers to practical implementation. These challenges may hamper the translation of the developed methodologies into actual clinical practice, posing barriers to the full exploitation of AI potentiality. To overcome these barriers, this PhD thesis builds upon the European “Ethical Guidelines for Trustworthy AI”, which outlines key principles such as fairness, robustness, privacy, and human oversight. By focusing on some of these requirements - particularly “diversity, non-discrimination, and fairness”, “technical robustness and safety”, “privacy and data governance”, and “human agency and oversight”, - the thesis proposes a pipeline that operationalizes these principles in real-world clinical applications. The work addresses three interconnected areas: 1. Dataset collection and bias mitigation: This stage addresses the identification and mitigation of dataset biases to enable fair and diverse AI models. By analyzing public datasets, the thesis highlights biases, such as minority, informativeness and selection biases. Representation learning techniques are proposed to reduce the reliance on labeled data and mitigate selection bias. 2. Privacy preservation and multicentric variability management: The second stage focuses on privacy-preserving methodologies, such as Federated Learning (FL), to address data-sharing constraints across diverse clinical datasets and manage variability across clinical datasets through domain adaptation techniques. 3. AI integration into clinical workflows: The third stage emphasizes the development of human-centric solutions into clinical workflows. By engaging clinicians in the design process and aligning with real-world practices, integrated AI systems can enhance usability, reduce operator dependency, and deliver reliable diagnostic support. These stages are demonstrated through diverse case studies, spanning from Electronic Health Records (EHRs) to Ultrasound (US) imaging and X-ray Coronary Angiography (XCA), showcasing the pipeline’s adaptability across healthcare applications. By aligning the pipeline with principles of fairness, generalizability, and human-centric design, this thesis seeks to close the gap between algorithm development in the lab and real-world application in clinical environments, offering actionable strategies for trustworthy AI in healthcare.File | Dimensione | Formato | |
---|---|---|---|
Tesi_DiCosmo.pdf
accesso aperto
Descrizione: Tesi_DiCosmo
Tipologia:
Tesi di dottorato
Licenza d'uso:
Creative commons
Dimensione
42.24 MB
Formato
Adobe PDF
|
42.24 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.