Achieve Fairness in AI-Assisted Mobile Healthcare Apps through Unsupervised Federated Learning

点击翻译按钮获取中文摘要

Deep learning models have been deployed in an increasing number of edge and mobile devices to provide healthcare in our life, from mobile dermatology assistant, mobile eye cancer (leukoria) detection, emotion detection, to comprehensive vital signs monitoring. All these techniques rely on visual assistance of the cameras that come with mobile devices and inevitably lead to different levels of fairness concerns, due to the inherent gender, race and/or socioeconomic bias in existing AI models. Compounding contributing factors include a lack of medical professionals from marginalized communities, inadequate information about those communities, and socioeconomic barriers to participating in data collection and research. In the absence of a diverse population that reflects that of the U.S. population, potential safety or efficacy considerations could be missed. What is worse, with inadequate data, AI algorithms could misdiagnose underrepresented people, leading to increasing health care disparities. Therefore, there is a critical need to address racial, skin color, and socioeconomic inequities in AI-assisted mobile diagnosis. This project will address the fairness issue in mobile AI assistants, using dermatology diagnosis and skin color inequity as the study case. Instead of collecting equitable demographic dataset in a centralized way, it will develop a federated on-device learning framework for participation inclusion, selective data contribution, and continuous personalization. The framework can continuously learn from new users’ data as they use the mobile apps with little human supervision. An unsupervised federated learning (FL) framework will be developed with heterogeneous hardware (high-end and low-end) and models such that users from all socioeconomic status can participate in the research. While various FL techniques have been developed, how to implement unsupervised FL with both hardware and model heterogeneity is not clear. It is also essential to achieve this goal with as little human supervision as possible since it is impractical to have a doctor constantly label the images when users are using these AI-based apps. In addition, even with FL, data from predominating population will still dominate the data collected. Non-uniform data selection techniques will be developed to automatically weigh the importance of different data for maximum fairness. Finally, not all neural networks exhibit the same inherent fairness even with the same biased data. A fairness-aware neural architecture search framework will be developed to find the networks that can achieve the most fairness. The expected outcome of this project is a holistic framework to mitigate the impacts of inequity by improving the inference performance for minorities. The developed techniques will be implemented as mobile apps with heterogeneous smart phones and evaluated with both public dataset and patients at UPMC. Data and code will be made available for public research. The developed techniques can be easily extended to all AI-assisted diagnosis and account for the inequity in various aspects such as age, sex, racial, etc.