Achieve Fairness in AI-Assisted Mobile Healthcare Apps through Unsupervised Federated Learning - Deep learning models are being increasingly implemented in edge and mobile devices to support healthcare applications. A critical technical challenge in deploying these techniques widely is ensuring their robustness and accuracy for the whole populations in real-world scenarios. Existing AI models can exhibit variations in performance due to unbalance in their training data, which often originates from limited sources and may not fully represent the “real data” seen in the real-world America. This lack of data representativeness can impact the reliability and generalizability of mobile diagnostic tools for all potential users. For example, AI models trained primarily on data from a few well-resourced institutions may not perform optimally for individuals in different geographic locations, healthcare settings. Therefore, there is a significant need to develop AI-assisted mobile diagnostic methods that demonstrate consistent and reliable performance for everyone. This project will address the challenge of achieving generalizable AI in mobile assistants, using dermatology diagnosis as the study case. Instead of relying on centralized data collection, which often struggles with representativeness and data sharing limitations, this project will develop a federated on-device learning framework to enable participation from diverse sources, selective data contribution, and continuous personalization. This framework will continuously learn from new users' data as they use mobile apps with minimal human supervision. An unsupervised federated learning (FL) framework will be developed to accommodate heterogeneous hardware (both high-end and low-end devices) and models, ensuring the framework can be utilized on a wide range of commonly available mobile devices. While various FL techniques exist, how to implement unsupervised FL with both hardware and model heterogeneity in this context is not well-established. Furthermore, even with FL, data volume that cannot represent the whole population may still dominate the learning process. To mitigate this, non-uniform data selection techniques will be developed to automatically weigh the importance of different data contributions for optimal model performance across diverse inputs. Finally, recognizing that different neural network architectures can exhibit varying degrees of robustness, a performance-aware neural architecture search framework will be developed to identify networks that achieve the most consistent and reliable performance. The expected outcome of this project is a novel federated learning framework that enhances the accuracy and reliability of mobile AI-assisted diagnosis for everyone. The developed techniques will be implemented as mobile apps compatible with heterogeneous smartphones and evaluated using both public datasets and patient data at UPMC. Data and code will be made available for public research. The developed techniques can be potentially extended to enhance the generalizability of other AI-assisted diagnostic tools.
