Federated Learning for Healthcare Data: The Benefits, Downsides, and Ethics of Cooperation

April 21, 2023

Introduction

In April 2016, the European Union (EU) passed the General Data Protection Regulation (GDPR), which was made enforceable in May 2018. GDPR includes important legislation surrounding an individual’s right to data privacy, an individual’s “right to be forgotten,” directives around appropriate and inappropriate uses of data, and strict guidelines on the types of data that may pass into, out of, and between countries in the Eurozone ¹. In the EU, GDPR is a critical piece of legislation that protects the data privacy of EU citizens and has been mirrored in the privacy laws of other countries ². Despite these clear positive outcomes, GDPR led to many unintended consequences for technology companies. AI-driven systems that freely migrated data across borders for training and inference could no longer do so; instead, data has remained siloed within the Eurozone countries. This obstruction was acutely felt by international companies handling healthcare data and using AI to improve patient care ³.

The AI challenges posed by GDPR, along with other legislation surrounding data transfer across borders, fueled the rise of federated learning (FL), a sub-field of machine learning that designs AI for situations in which passing data freely is not straightforward. FL is now applied to a wide variety of data-sensitive situations but is particularly prevalent in health care and medicine. In theory, federated learning strategies can approach the performance of AI systems without the limitations of data transmission restrictions; however, in practice, this is almost never the case. No matter the application, data privacy will always incur trade-offs in AI model performance ⁴. Though FL has demonstrated a great deal of promise in improving model performance without compromising data privacy regulations, federated learning constructs a system where cooperation could be risky. Thus, an ethical dilemma arises: Should healthcare providers collaborate with FL systems? A Christian perspective on cooperation can help when weighing the risks of collaborating with others, especially when our own interests may suffer.

Basics of Federated Learning

Before discussing the passage of information through FL, let us introduce a standard centralized machine learning (ML) framework. Figure 1 shows a framework where each system within a nation’s borders passes data freely. Orange lines represent the occasional routine porting of new training data (such as updates to patients’ medical histories) from each system to a centralized machine learning model, which is retrained to incorporate the new information. Outside of these retraining procedures, data in the form of queries are regularly routed to the machine learning model, which returns a prediction (also known as an inference). The machine learning model can reach its maximum potential when all systems regularly share new data for retraining. Since sharing data across borders may violate certain regulations, FL offers an alternative model.

Diagram Description automatically generated

Figure 1: A centralized ML architecture diagram.

Figure 2 shows an FL infrastructure that satisfies the criterion prohibiting data distribution outside the data’s country of origin. There are several important changes to consider in this infrastructure. First, each country trains its own separate AI agent on data situated within its borders. Systems in that country occasionally push new training data to the model, which is retrained in turn. Second, models from each nation are shared across borders, such that every nation has a copy of every other nation’s AI, together forming an ensemble of AIs. Because no data has been exchanged outside the country of origin, no regulations are violated. Finally, when a system makes a query, it is processed by the ensemble of AI models held by that country, and an inference is returned – all without crossing borders. While there are different ways to aggregate AI model predictions within the ensemble, the easiest method is to average model predictions (for regression tasks) or take the majority vote (for classification tasks). When one nation retrains its AI, the model updates are passed to the ensembled copy for every other nation. Updates to specific models are now the only information passed between countries.

Diagram Description automatically generated

Figure 2: A simple FL infrastructure diagram.

While FL infrastructures cannot perform as well as centralized ML infrastructures, the FL voting/averaging ensemble can outperform the individual AI agents trained within each country. Because of these properties, FL infrastructures have become the default for preserving data privacy without significantly harming performance. In the healthcare space, FL has found applications in segmenting MRI scans ⁵, identifying brain tumors ⁶, and predicting hospital stay times ⁷. With each of these applications, the factors contributing to decreased FL performance must be carefully assessed.

Downsides of Federated Learning

Ensemble approaches are not new to ML and often lead to state-of-the-art performance. Random forest methods, for example, are among the most effective and easily implemented supervised methods available. The performance of FL models suffers not due to their methods, but due to their data divisions. Under ideal conditions, when data is divided to be independent and identically distributed (iid), FL will perform as well as other ensemble models. However, data divided by national borders is not iid; that is, there are major differences between the patients and healthcare systems in different countries. The datasets are biased within their individual silos and, through their shared models, export their biases to other all ensembles. For example, in 2019 the age-standardized share of adults with diabetes was 3.2% in Ireland, and 10.4% in Germany ⁸. An FL model ensemble covering both countries may over-diagnose diabetes in Ireland and under-diagnose diabetes in Germany. Many solutions have been proposed to overcome the non-iid distribution of data, with various levels of success ⁹; however, these improvements still cannot attain the accuracy of a centralized ML architecture.

In addition to poorer performance, FL architectures can be susceptible to a variety of attacks and violations. Some have informally said that FL requires that no participants ever act “stupidly or adversarially”. For example, suppose that a participant’s diabetes prediction model were erroneously trained to indicate no positive diagnoses (this might be achieved with a simple backdoor attack ¹⁰). While this error persists, it may lead other healthcare providers to under-diagnose diabetes in their own patients. Aside from backdoor attacks, there have been many adversarial strategies leveled against FL architectures, which has turned the space into an arms race between FL engineers, AI security specialists, and hackers of all stripes ¹¹. Given such a volatile and uncertain environment, should participants opt to cooperate at all?

A Christian Image of Cooperation

The state of Christian denominational division offers an excellent substrate when considering the benefits and pitfalls of collaboration. Denominations often set up “borders” which encompass assumed consensus around the issues and beliefs we (Christians) choose to prioritize. Rather than cooperating as a unified body of Christ (1 Corinthians 12:27), it is often easier to vilify fellow Christians over inconsequential differences. Though denominational differences may be firmly entrenched, it is far better to recognize the shared purpose in God’s kingdom on earth. Collaboration between denominations can be costly for both parties and even require difficult compromise, but groups like Evangelicals and Catholics Together ¹² demonstrate the power found in Christians finding common ground. The Christian call to cooperate with others extends not only to fellow Christians, but to those with vastly different beliefs – this mission is especially evident in Jesus’ parable of the good Samaritan (Luke 10: 25-37).

Within the US alone, there are over a dozen non-affiliated Good Samaritan hospitals. The good Samaritan is an image of a social pariah, an apostate, demonstrating radical hospitality towards a Jewish traveler, taking on physical and financial hardship for the sake of a fellow human being. To Christians, this parable exemplifies selfless cooperation. To hospitals and healthcare providers, it may illustrate how patients ought to be cared for. Whether or not hospitals choose to participate in an FL framework, the directive to care for and serve patients, regardless of their identity or status, remains the same. FL has the potential to provide benefits overall, even though individual healthcare providers may face poorer model performance in the future. Healthcare providers opting to cooperate are demonstrating their understanding of patient care as extending beyond their own community of patients.

Conclusion

This article discusses data privacy, and how it drove the adoption of federated learning. After reviewing how FL operates, this inquiry illustrates the pros and cons of cooperating within an FL framework that are difficult to weigh. Then, the Christian example of the good Samaritan is advanced as a model for cooperation. Finally, participation in FL frameworks is considered as a modality through which hospitals could serve patients both within and beyond their locale. I hope that as FL frameworks make technical and security innovations, the choice to cooperate with the goal of improved patient care becomes a simple one.

Acknowledgements

Thanks to Melody Cantwell for her insights into church history and theology. Thanks to Haley Griese for proofreading, editing, and publishing this work.

References

Voigt, Paul, and Axel Von dem Bussche. “The eu general data protection regulation (gdpr).” A Practical Guide, 1st Ed., Cham: Springer International Publishing 10, no. 3152676 (2017): 10-5555.

Goddard, Michelle. “The EU General Data Protection Regulation (GDPR): European regulation that has a global impact.” International Journal of Market Research 59, no. 6 (2017): 703-705.

Forcier, Mélanie Bourassa, Hortense Gallois, Siobhan Mullan, and Yann Joly. “Integrating artificial intelligence into health care through data access: can the GDPR act as a beacon for policymakers?.” Journal of Law and the Biosciences 6, no. 1 (2019): 317-335.

Humerick, Matthew. “Taking AI personally: how the EU must learn to balance the interests of personal data privacy & artificial intelligence.” Santa Clara High Tech. LJ 34 (2017): 393.

Roy, Abhijit Guha, Shayan Siddiqui, Sebastian Pölsterl, Nassir Navab, and Christian Wachinger. “Braintorrent: A peer-to-peer environment for decentralized federated learning.” arXiv preprint arXiv:1905.06731 (2019).

Li, Wenqi, Fausto Milletarì, Daguang Xu, Nicola Rieke, Jonny Hancox, Wentao Zhu, Maximilian Baust et al. “Privacy-preserving federated brain tumour segmentation.” In Machine Learning in Medical Imaging: 10th International Workshop, MLMI 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 13, 2019, Proceedings 10, pp. 133-141. Springer International Publishing, 2019.

Sheller, Micah J., G. Anthony Reina, Brandon Edwards, Jason Martin, and Spyridon Bakas. “Multi-institutional deep learning modeling without sharing patient data: A feasibility study on brain tumor segmentation.” In Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 4th International Workshop, BrainLes 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Revised Selected Papers, Part I 4, pp. 92-104. Springer International Publishing, 2019.

Huang, Li, Andrew L. Shea, Huining Qian, Aditya Masurkar, Hao Deng, and Dianbo Liu. “Patient clustering improves efficiency of federated machine learning to predict mortality and hospital stay time using distributed electronic medical records.” Journal of biomedical informatics 99 (2019): 103291.

Williams, Rhys, Suvi Karuranga, Belma Malanda, Pouya Saeedi, Abdul Basit, Stéphane Besançon, Christian Bommer et al. “Global and regional estimates and projections of diabetes-related health expenditure: Results from the International Diabetes Federation Diabetes Atlas.” Diabetes research and clinical practice 162 (2020): 108072.

Zhao, Yue, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. “Federated learning with non-iid data.” arXiv preprint arXiv:1806.00582 (2018).

Sattler, Felix, Simon Wiedemann, Klaus-Robert Müller, and Wojciech Samek. “Robust and communication-efficient federated learning from non-iid data.” IEEE transactions on neural networks and learning systems 31, no. 9 (2019): 3400-3413.

Wang, Hongyi, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris Papailiopoulos. “Attack of the tails: Yes, you really can backdoor federated learning.” Advances in Neural Information Processing Systems 33 (2020): 16070-16084.

Bhagoji, Arjun Nitin, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. “Analyzing federated learning through an adversarial lens.” In International Conference on Machine Learning, pp. 634-643. PMLR, 2019.

George, Timothy. “Evangelicals and Catholics Together: A New Initiative”, Christianity today, Dec 8, 1997.

Marcus Schwarting

Marcus Schwarting is a PhD candidate in computer science at the University of Chicago, where his research focuses on applying deep learning to important challenges in computational chemistry, materials science, and spectroscopy. Marcus graduated from the University of Louisville in 2018 with degrees in mathematics and chemical engineering. After working for four years at the National Renewable Energy Laboratory and Argonne National Laboratory, Marcus began his PhD studies in 2020 under Drs. Ian Foster and Eric Jonas.