The Pitfalls of Using Open Data to Develop Deep Learning Solutions for COVID-19 Detection in Chest X-Rays.

Since the emergence of COVID-19, deep learning models have been developed to identify COVID-19 from chest X-rays. With little to no direct access to hospital data, the AI community relies heavily on public data comprising numerous data sources. Model performance results have been exceptional when training and testing on open-source data, surpassing the reported capabilities of AI in pneumonia-detection prior to the COVID-19 outbreak. In this study impactful models are trained on a widely used open-source data and tested on an external test set and a hospital dataset, for the task of classifying chest X-rays into one of three classes: COVID-19, non-COVID pneumonia and no-pneumonia. Classification performance of the models investigated is evaluated through ROC curves, confusion matrices and standard classification metrics. Explainability modules are implemented to explore the image features most important to classification. Data analysis and model evalutions show that the popular open-source dataset COVIDx is not representative of the real clinical problem and that results from testing on this are inflated. Dependence on open-source data can leave models vulnerable to bias and confounding variables, requiring careful analysis to develop clinically useful/viable AI tools for COVID-19 detection in chest X-rays.

A data flow process for confidential data and its application in a health research project.

BACKGROUND: The use of linked healthcare data in research has the potential to make major contributions to knowledge generation and service improvement. However, using healthcare data for secondary purposes raises legal and ethical concerns relating to confidentiality, privacy and data protection rights. Using a linkage and anonymisation approach that processes data lawfully and in line with ethical best practice to create an anonymous (non-personal) dataset can address these concerns, yet there is no set approach for defining all of the steps involved in such data flow end-to-end. We aimed to define such an approach with clear steps for dataset creation, and to describe its utilisation in a case study linking healthcare data. METHODS: We developed a data flow protocol that generates pseudonymous datasets that can be reversibly linked, or irreversibly linked to form an anonymous research dataset. It was designed and implemented by the Comprehensive Patient Records (CPR) study in Leeds, UK. RESULTS: We defined a clear approach that received ethico-legal approval for use in creating an anonymous research dataset. Our approach used individual-level linkage through a mechanism that is not computer-intensive and was rendered irreversible to both data providers and processors. We successfully applied it in the CPR study to hospital and general practice and community electronic health record data from two providers, along with patient reported outcomes, for 365,193 patients. The resultant anonymous research dataset is available via DATA-CAN, the Health Data Research Hub for Cancer in the UK. CONCLUSIONS: Through ethical, legal and academic review, we believe that we contribute a defined approach that represents a framework that exceeds current minimum standards for effective pseudonymisation and anonymisation. This paper describes our methods and provides supporting information to facilitate the use of this approach in research.