Scalable federated learning for emergency care using low cost microcomputing: Real-world, privacy preserving development and evaluation of a COVID-19 screening test in UK hospitals (preprint)

Background Tackling biases in medical artificial intelligence requires multi-centre collaboration, however, ethical, legal and entrustment considerations may restrict providers' ability to participate. Federated learning (FL) may eliminate the need for data sharing by allowing algorithm development across multiple hospitals without data transfer. Previously, we have shown an AI-driven screening solution for COVID-19 in emergency departments using clinical data routinely available within 1h of arrival to hospital (vital signs & blood tests; CURIAL-Lab). Here, we aimed to extend and federate our COVID-19 screening test, demonstrating development and evaluation of a rapidly scalable and user-friendly FL solution across 4 UK hospital groups. Methods We supplied a Raspberry Pi 4 Model B device, preloaded with our end-to-end FL pipeline, to 4 NHS hospital groups or their locally-linked research university (Oxford University Hospitals/University of Oxford (OUH), University Hospitals Birmingham/University of Birmingham (UHB), Bedfordshire Hospitals (BH) and Portsmouth Hospitals University (PUH) NHS trusts). OUH, PUH and UHB participated in federated training and calibration, training a deep neural network (DNN) and logistic regressor to predict COVID-19 status using clinical data for pre- pandemic (COVID-19-negative) admissions and COVID-19-positive cases from the first wave. We performed federated prospective evaluation at PUH & OUH, and external evaluation at BH, evaluating the resultant global and site-tuned models for admissions to the respective sites during the second pandemic wave. Removable microSD storage was destroyed on study completion. Findings Routinely collected clinical data from a total 130,941 patients (1,772 COVID-19 positive) across three hospital groups were included in federated training. OUH, PUH and BH participated in prospective federated evaluation, with sets comprising 32,986 patient admissions (3,549 positive) during the second pandemic wave. Federated training improved DNN performance by a mean of 27.6% in terms of AUROC when compared to models trained locally, from AUROC of 0.574 & 0.622 at OUH & PUH to 0.872 & 0.876 for the federated global model. Performance improvement was more modest for a logistic regressor with a mean AUROC increase of 13.9%. During federated external evaluation at BH, the global DNN model achieved an AUROC of 0.917 (0.893-0.942), with 89.7% sensitivity (83.6-93.6) and 76.7% specificity (73.9- 79.1). Site-personalisation of the global model did not give a significant improvement in overall performance (AUROC improvement <0.01), suggesting high generalisability. Interpretations We present a rapidly scalable hardware and software FL solution, developing a COVID-19 screening test across four UK hospital groups using inexpensive micro- computing hardware. Federation improved model performance and generalisability, and shows promise as an enabling technology for deep learning in healthcare. Funding University of Oxford Medical & Life Sciences Translational Fund/Wellcome

Mode Connections For Clinical Incremental Learning: Lessons From The COVID-19 Pandemic (preprint)

Dynamic distribution shifts caused by evolving diseases and demographic changes require domain-incremental adaptation of clinical deep learning models. However, this process of adaptation is often accompanied by catastrophic forgetting, and even the most sophisticated methods are not good enough for clinical applications. This paper studies incremental learning from the perspective of mode connections, that is, the low-loss paths connecting the minimisers of neural architectures (modes or trained weights) in the parameter space. The paper argues for learning the low-loss paths originating from an existing mode and exploring the learned paths to find an acceptable mode for the new domain. The learned paths, and hence the new domain mode, are a function of the existing mode. As a result, unlike traditional incremental learning, the proposed approach is able to exploit information from a deployed model without changing its weights. Pre-COVID and COVID-19 data collected in Oxford University hospitals are used as a case study to demonstrate the need for domain-incremental learning and the advantages of the proposed approach.

RapiD_AI: A framework for Rapidly Deployable AI for novel disease & pandemic preparedness (preprint)

COVID-19 is unlikely to be the last pandemic that we face. According to an analysis of a global dataset of historical pandemics from 1600 to the present, the risk of a COVID-like pandemic has been estimated as 2.63% annually or a 38% lifetime probability. This rate may double over the coming decades. While we may be unable to prevent future pandemics, we can reduce their impact by investing in preparedness. In this study, we propose RapiD_AI: a framework to guide the use of pretrained neural network models as a pandemic preparedness tool to enable healthcare system resilience and effective use of ML during future pandemics. The RapiD_AI framework allows us to build high-performing ML models using data collected in the first weeks of the pandemic and provides an approach to adapt the models to the local populations and healthcare needs. The motivation is to enable healthcare systems to overcome data limitations that prevent the development of effective ML in the context of novel diseases. We digitally recreated the first 20 weeks of the COVID-19 pandemic and experimentally demonstrated the RapiD_AI framework using domain adaptation and inductive transfer. We (i) pretrain two neural network models (Deep Neural Network and TabNet) on a large Electronic Health Records dataset representative of a general in-patient population in Oxford, UK, (ii) fine-tune using data from the first weeks of the pandemic, and (iii) simulate local deployment by testing the performance of the models on a held-out test dataset of COVID-19 patients. Our approach has demonstrated an average relative/absolute gain of 4.92/4.21% AUC compared to an XGBoost benchmark model trained on COVID-19 data only. Moreover, we show our ability to identify the most useful historical pretraining samples through clustering and to expand the task of deployed models through inductive transfer to meet the emerging needs of a healthcare system without access to large historical pretraining datasets.

RapiD_AI: A framework for Rapidly Deployable AI for novel disease & pandemic preparedness (preprint)

COVID-19 is unlikely to be the last pandemic that we face. According to an analysis of a global dataset of historical pandemics from 1600 to the present, the risk of a COVID-like pandemic has been estimated as 2.63% annually or a 38% lifetime probability. This rate may double over the coming decades. While we may be unable to prevent future pandemics, we can reduce their impact by investing in preparedness. In this study, we propose RapiD AI : a framework to guide the use of pretrained neural network models as a pandemic preparedness tool to enable healthcare system resilience and effective use of ML during future pandemics. The RapiD AI framework allows us to build highperforming ML models using data collected in the first weeks of the pandemic and provides an approach to adapt the models to the local populations and healthcare needs. The motivation is to enable healthcare systems to overcome data limitations that prevent the development of effective ML in the context of novel diseases. We digitally recreated the first 20 weeks of the COVID-19 pandemic and experimentally demonstrated the RapiD AI framework using domain adaptation and inductive transfer. We (i) pretrain two neural network models (Deep Neural Network and TabNet) on a large Electronic Health Records dataset representative of a general in-patient population in Oxford, UK, (ii) fine-tune using data from the first weeks of the pandemic, and (iii) simulate local deployment by testing the performance of the models on a held-out test dataset of COVID-19 patients. Our approach has demonstrated an average relative/absolute gain of 4.92/4.21% AUC compared to an XGBoost benchmark model trained on COVID-19 data only. Moreover, we show our ability to identify the most useful historical pretraining samples through clustering and to expand the task of deployed models through inductive transfer to meet the emerging needs of a healthcare system without access to large historical pretraining datasets.