Sileshi Nibret Zeleke

From 01/12/2024 to 31/05/2025

Open Ideas

Explainable and Privacy-Preserving Federated Learning Based System for Trustworthy Aging-Related Disease Diagnosis

Pennsylvania State University

I am a PhD student in Computer Science at the University of Bari, Italy currently conducting research at Pennsylvania State University, USA through the NGI Transatlantic Fellowship. My work focuses on explainable AI and federated learning for predictive healthcare, with a particular emphasis on aging-related diseases. I have developed advanced machine learning models, including hybrid deep networks and large language models, to improve diagnosis and interpretability. My contributions have been recognized through publications and presentations at IEEE Big Data and similar venues. I am passionate about building transparent, reliable AI systems that address real-world medical challenges across decentralized environments.

Advancements in artificial intelligence help diagnose health conditions and exact intensive age-related diseases. However, preserving patient privacy and the trustworthiness of the models are of utmost importance. Most adopted centralized data storage and processing methods raise data privacy and security concerns. Therefore, there is a need for an explainable federated learning (FL) approach that ensures the privacy of data owners. Given this, we are developing a method that leverages the potential of conventional deep learning models and large vision language models (VLM) for the diagnosis of heart conditions. Deep learning models are efficient in terms of computational requirements for a federated approach but are difficult to explain the results in a reasoning approach. Given this, we leverage the potential of vision language models. First, we finetune a pre-trained vision language model on ECG data.

Then, we implement deep learning models and apply conformal prediction to identify the misclassification of the deep learning model to be improved using the VLM and to produce a textual description or explanation of why the model thinks it is that specific heart condition. In that way, we can produce natural language-based explanations in addition to visual explanations from deep learning models. Applying VLM in this scenario has two major advantages. The first is that it helps to improve the diagnosis decision of the deep learning model, and the other is providing natural language-based explanations.

Over the past months, I have made substantial progress on my research and development objectives, primarily focused on explainable intelligence FL and biomedical signal analysis. To mention specifics, we fine-tuned a VLM for ECG interpretation and began integrating it with deep learning models to improve multi-class classification performance and explainability.

The fellowships have significantly enhanced my research capacity by facilitating international collaboration and providing access to advanced resources at Pennsylvania State University. It enabled me to explore cutting-edge topics like VLM and explainable federated learning for healthcare, deepen my expertise in privacy-preserving AI, and accelerate the integration of novel architectures into real-world biomedical applications. The program also broadened my professional network and provided valuable exposure to transatlantic innovation ecosystems, aligning my work with EU and US digital research priorities. Moreover, it helps me build a research collaboration with different individuals, such as professors, PhD students, and others in the US. I submitted an international conference paper, and it is under review. Moreover, I attended and presented a paper at the IEEE Big Data conference.