Addressing machine learning challenges with microcomputing and federated learning

Abstract

In this issue of The Lancet Digital Health, Andrew A S Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar address a major challenge of artificial intelligence (AI) development through the use of microcomputing and federated learning. Access to diverse training data, rather than computing power, is the limiting factor for many machine learning problems, which Soltan and colleagues overcome using a full-stack pipeline on a preconfigured and inexpensive microcomputer that enabled federated learning across four hospital sites for COVID-19 prediction. Concerns that machine learning predictions are biased because of limitations of the available data mean that many of the transformational changes promised by artificial intelligence have not yet been realised. Machine learning models trained with insufficient or unrepresentative data have the potential to cause patient harm as a result of inaccurate predictions.2Wiens J Saria S Sendak M et al.Do no harm: a roadmap for responsible machine learning for health care.Nat Med. 2019; 25: 1337-1340Crossref PubMed Scopus (370) Google Scholar Data availability and diversity can be improved through federated learning to improve model robustness. Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar used federated learning to address data needs by federating four sites to increase patient recruitment numbers and improve model training, while simultaneously preserving local data control and patient privacy. Their full-stack pipeline achieved the largest secondary-care federated learning study to date in terms of patient numbers, using data from 130 941 patients (1772 of whom were COVID-19-positive) for training and 32 986 patients (3549 COVID-19-positive) for evaluation. Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar deployed federated learning at four hospital groups, in Oxford, Birmingham, Bedfordshire, and Portsmouth. Federated learning improved the area under the receiver operating characteristic curve (AUROC) from 0·574 [95% CI 0·560–0·589] to 0·872 [0·862–0·882] at Oxford and from 0·622 [0·608–0·637] to 0·876 [0·865–0·886] at Portsmouth—a substantial improvement through multi-site training. The final federated learning model showed robustness during model evaluation, with a high AUROC of 0·917 [0·893–0·942] at Bedfordshire. In addition to addressing the large amount of data required for machine learning, model training requires the use of advanced computing architectures that incorporate relevant hardware and software.3Hu Y, Liu Y, Liu Z. A survey on convolutional neural network accelerators: GPU, FPGA and ASIC. 2022 14th International Conference on Computer Research and Development; Jan 7–9, 2022 (pp 100–07).Google Scholar Machine learning frequently involves expensive computer hardware using a Compute Unified Device Architecture (CUDA)-enabled graphics processing unit for highly parallelised computing to optimise training time. The model developed by Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar runs on a microcomputer, the Raspberry Pi, which is inexpensive and can be shipped to additional sites. An inexpensive microcomputer can also easily be physically destroyed at the end of a clinical project for patient privacy and security, which is occasionally a requirement for clinical studies, such as in this work. The algorithm used by Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar analysed data that are routinely collected within 1 h of a patient arriving at the hospital: vital signs, full blood count, liver function tests, urea and electrolytes, and C-reactive protein concentrations. They concluded that the most impactful features for model prediction were the oxygen delivery device for the logistic regression model and the eosinophil count for the deep neural network model, similar to results obtained in clinical studies that did not involve AI.4Lindsley AW Schwartz JT Rothenberg ME Eosinophil responses during COVID-19 infections and coronavirus vaccination.J Allergy Clin Immunol. 2020; 146: 1-7Summary Full Text Full Text PDF PubMed Scopus (224) Google Scholar This finding suggests that such machine learning methods could be useful for biomarker refinement to optimise clinical trials. Federated learning removes the requirement for identifiable patient data to leave a site for multi-site model training, alleviating a number of ethical, security, and privacy concerns. Because these concerns—in addition to data and model licensing considerations—cannot be entirely removed through federated learning, the number or rate of new studies that can be conducted with machine learning and federated learning is still limited. These concerns are a probable reason why additional sites were not incorporated to match the numbers used in several other federated learning studies.5Dayan I Roth HR Zhong A et al.Federated learning for predicting clinical outcomes in patients with COVID-19.Nat Med. 2021; 27: 1735-1743Crossref PubMed Scopus (246) Google Scholar, 6Bai X Wang H Ma L et al.Advancing COVID-19 diagnosis with privacy-preserving collaboration in artificial intelligence.Nat Mach Intell. 2021; 3: 1081-1089Crossref Scopus (22) Google Scholar Further work incorporating additional sites would more fully realise the potential of the method developed by Soltan and colleagues.1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar The microcomputing approach used here has two major drawbacks. Raspberry Pi microcomputers can have security vulnerabilities that require consideration when using sensitive data.7Sainz-Raso J Martin S Diaz G Castro M Security vulnerabilities in Raspberry Pi—analysis of the system weaknesses.IEEE Consum Electron Mag. 2019; 8: 47-52Crossref Scopus (8) Google Scholar Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar address security through a firewall and a pre-agreed port. Health-care data, such as imaging5Dayan I Roth HR Zhong A et al.Federated learning for predicting clinical outcomes in patients with COVID-19.Nat Med. 2021; 27: 1735-1743Crossref PubMed Scopus (246) Google Scholar, 6Bai X Wang H Ma L et al.Advancing COVID-19 diagnosis with privacy-preserving collaboration in artificial intelligence.Nat Mach Intell. 2021; 3: 1081-1089Crossref Scopus (22) Google Scholar or genomics8Stephens ZD Lee SY Faghri F et al.Big data: astronomical or genomical?.PLoS Biol. 2015; 13e1002195 Crossref PubMed Scopus (870) Google Scholar data, can easily overwhelm the computing cores and memory capacity of a Raspberry Pi, making microcomputing impractical; however, high-end computing architecture was not necessary for achieving the results in this work. With these considerations in mind, the work of Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar enabled multi-site machine learning to improve model training, which will help to achieve transformational changes in health-care. Full-stack microcomputing solutions can reach hospitals that otherwise do not conduct dedicated research yet hold patient information that is important for improving disease prediction. This inexpensive solution can enable rapid deployment and eliminates the need for specialist technical expertise, which is advantageous considering that hospitals vary in their data infrastructure.9Hertzum M Ellingsen G The implementation of an electronic health record: comparing preparations for Epic in Norway with experiences from the UK and Denmark.Int J Med Inform. 2019; 129: 312-317Crossref Scopus (27) Google Scholar As the code is provided online, researchers can fully detail the installation and setup processes involved in the work of Soltan and colleagues1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar to replicate or develop it further. This illustration of microdevice federated learning helps to pave the way for further applications of this technology, such as with wearables that collect physiological data.10Wang W Li X Qiu X Zhang X Brusic V Zhao J A privacy preserving framework for federated learning in smart healthcare systems.Inf Process Manage. 2023; 60103167 Google Scholar In summary, federated learning, microdevices, and open-source code, as used by Soltan and colleagues,1Soltan AAS Thakur A Yang J et al.A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitals.Lancet Digital Health. 2024; 6: e93-104Summary Full Text Full Text PDF Google Scholar are welcomed for developing clinical studies with federated multi-site model training for the purpose of improving patient outcomes. I receive support from the NIHR Cambridge Biomedical Research Centre (BRC-1215-20014), the Myrovlytis Trust (MT22_15_Marciniak), Cancer Research UK (A25922), and Addenbrooke's Charitable Trust, and have industrial partners including Rhino Health, the Asher Orion Group, GE Healthcare, GlaxoSmithKline, and AstraZeneca. I have recently held funding from the European Union's Horizon 2020 research and innovation programme (grant agreement no 761214). I am a former president of the Postdoctoral Society of Cambridge and held a committee role in the UK Research Staff Association. I do not receive personal fees nor benefits that could be considered a conflict of interest. The views expressed are my own and not those of the funders and partners. A scalable federated learning solution for secondary care using low-cost microcomputing: privacy-preserving development and evaluation of a COVID-19 screening test in UK hospitalsWe developed an embedded system for federated learning, using microcomputing to optimise for ease of deployment. We deployed full-stack federated learning across four UK hospital groups to develop a COVID-19 screening test without centralising patient data. Federation improved model performance, and the resultant global models were generalisable. Full-stack federated learning could enable hospitals to contribute to AI development at low cost and without specialist technical expertise at each site. Full-Text PDF Open Access

Ähnliche Arbeiten