Background to the Project
Coronavirus Disease 2019 (COVID-19) is highly contagious, and severe cases can lead to acute failure of the lungs, multiple organs and ultimately death. Chest X-Rays and CT scans provide valuable diagnostic and monitoring information that can complement the laboratory and clinical data. We have formed a Cambridge led collaboration to bring together clinicians, imaging scientists and AI specialists from Cambridge and throughout the world. The ambitions for the project are to develop a clinical tool for deployment in Addenbrooke's hospital in Cambridge which would help with the future COVID-19 cases. This would then be expanded to other respiratory disease areas in future. In addition, we are keen that AI and data science can be more helpful to future pandemics by developing a blueprint for the things we would have done differently in this pandemic and how we can prepare for the next one.
The diagnosis of COVID-19 can be confirmed by a laboratory test, however, the test has high false-negative rates which leads to delayed diagnosis and treatment. Fast and accurate diagnosis of patients is incredibly important, as is prognostication of whether a patient is likely to recover, require intensive care unit (ICU) admission or intensive ventilation. These tools will assist clinicians to allow them to make more informed decisions, leading to better patient outcome and efficient resource allocation (important at times of stretched resources).
Key Analysis Themes
Computed Tomography (CT)
Chest CT imaging is used extensively throughout some countries e.g. China and Russia as a screening tool for COVID-19 patients and is used in the UK for assessing severe or complex COVID-19 cases. In our collaboration we have development AI based segmentation algorithms to extract the ground-glass opacity (GGO) and consolidation regions of the lungs and we are training a variety of classifier networks to use features in the regions of pathology to link to patient outcome. We also have access to a large cohort of non-COVID-19 pneumonia patients and training classification networks to identify COVID-19 specific features to permit fast diagnosis of the COVID-19 patients.
Chest X-Rays (CXR)
CXR imaging is used extensively throughout the UK and Europe as a first line imaging modality for patients in hospitals. We have developed models that use either (a) a single time point image to diagnose or prognosticate, or (b) a time series of CXR images which allow for accurate prognosis based on the changes in the CXR imaging.
We have developed models using detailed longitudinal clinical data for diagnosis and prognostication of patients. These incorporate all clinical readings and test results to provide an evolving prediction for the patient status in 48 hours and 72 hours. These also inform for how particular treatments are likely to affect the patient journey.
Most literature focusses on the features of the lung parenchyma to give a likely prognosis for COVID-19 patients, but we believe that the heart features, such as the calcium score, are of critical importance -- especially given the demonstrated poorer outcomes for patients with cardiovascular issues.
Full Blood Count
We have access to a large dataset of full blood count results for COVID-19 and non-COVID-19 patients. The full blood count is a cheap, routine, test performed in every hospital in the world. Identifying a COVID-19 viral signature in the blood would be revolutionary as it would allow rapid triage of patients in hospitals. We are also identifying prognostic markers in the blood of COVID-19 patients which could determine outcomes.
Quality Control (QC)
"Garbage in, garbage out" is a common refrain in the machine learning community. Therefore, we have developed tools for automated QC of the CXR images for patients to allow us to discard low quality issues quickly and identify other biases in our datasets.
Preparedness for Future Pandemics
It is clear that the imaging and machine learning community could have responded better to the pandemic if the data were available rapidly and the appropriate analysis were performed to develop algorithms quicker. We are writing the blueprint for how to respond to the next pandemic, asking the critical questions:
Where are the bottlenecks in the analysis and how can these be mitigated against in a future pandemic?
How can we rapidly develop prognostic and diagnostic algorithms for a brand new disease?
What is the most efficient method for integrating the algorithms into the clinical workflow?
How would we proceed differently to ensure rapid regulation of the algorithms as a medical device?
Machine learning solutions, with the right data and development, can provide diagnostic and prognostic algorithms within weeks. Next time around we must be better placed to harness this promise.