The roadmap to a safe and sustainable AI medical solution, during and beyond COVID-19
By Catalina Barzescu
All hands on deck
The COVID-19 pandemic demands urgent action to manage the influx of patients. For AI companies, this is both a call and an opportunity to deliver solutions that could increase the speed and accuracy of clinical decisions. We too recognised that our experience deploying AI for chest CT should be used to serve the medical staff fighting the disease (a responsibility we took in collaboration with our partners as ICOVAI).
However, some voices have started to question the legitimacy of multiple initiatives creating algorithms on the fly, a process that usually takes months or even years. There are concerns that the AI tools may be built on questionable datasets or circumventing regulatory approvals.
As this article will show, the shortcut-free roadmap to building and deploying a well-performing, certified, and integrated AI solution for clinical practice, during and after COVID-19, is long and testing; to illustrate, we will refer to a possible AI tool for COVID-19 imaging. Our plea is that the urgency of delivering an AI solution does not outweigh the requirements for clinical safety and quality.
Uses of CT for COVID-19
As our understanding of COVID is evolving, so is the debate around the role of medical imaging. Medical societies have issued statements and advice on the use of CT; on the frontlines, clinical practice has followed the availability of tests, staff, and equipment.
The literature in brief
Medical societies — e.g. ACR and ESR-ESTI — advise that CT is not suitable for first-line screening or diagnosing COVID, because it is not sensitive enough (one false negative can cause an outbreak), and not specific enough (findings are not unique to COVID). For the final diagnosis of COVID-19, there is consensus around using the RT-PCR test.
Use cases are identified for CT for a group of patients with COVID-19 (suspicion), complications, or worsening respiratory symptoms. The Fleischner Society considers CT scanning “a major tool if symptoms worsen or in an environment that is resource-constrained for RT-PCR”. Research published in BMJ Journals advocates for a “limited, but important, role” to confirm clinical suspicion in patients with negative RT-PCR.
Other findings indicate that the “quantification of the well-aerated lung at baseline chest CT can be a valuable tool for predicting prognosis in patients with COVID-19 pneumonia”. The use case of using CT to predict a poor outcome or the need for ventilators/ICU requires more research and may prove valuable during future peaks or outbreaks.
In addition, COVID appears to trigger a prothrombotic state (i.e. patients getting blood clots), and some are suggesting a lower threshold for performing a CT pulmonary angiogram.
The real-world practice
Some countries are using CT for diagnosis because it is fast and available. RT-PCR tests are far from perfect: their sensitivity in clinical practice is often lower than in laboratory settings, the turnover time is high, and tests are scarce. As a disadvantage, the cleaning of CT scanners to prevent spreading is time-consuming (whereas, portable X-ray machines are easier to clean).
Hospitals in the Netherlands have been using chest CT scans to assess the pulmonary involvement of COVID-19. To standardize this assessment, The Dutch Radiological Society introduced CO-RADS, a five-point scale guideline for CT reporting. Other reporting templates, such as BSTI, ESTI, or RSNA, provide a similar framework under a different naming convention.
What is the potential of AI for CT reporting?
The clinical need for AI to support radiologists analysing the CT scans of COVID patients has also evolved with the research.
The most useful application seems to be assessing the severity of the disease in order to triage patients and manage risk. An AI algorithm could automatically inform the likelihood of COVID (for example, return the CO-RADS score) and quantify the degree of lung damage. Detection and quantification would be the first steps in predicting the patient course and outcome, should clinical evidence support this development.
AI for COVID-19 CT reporting is expected to relieve the overworked medical staff and increase the confidence in follow-up decisions.
The ICOVAI approach to AI for COVID-19
The following steps are our roadmap to answering a clinical need with an AI solution for COVID-19 on chest CTs:
1. Data collection
AI lives off data and relies on fundamental scientific principles. It is essential to gather high-quality datasets from diverse CT scanners, hospitals, and countries. Quantity matters as well in order to train a well-performing model. Two important considerations come in this step:
- Processing anonymised data only, in line with GDPR. CT scans are considered ‘personal data’. However, they can be anonymised, meaning that there is no realistic possibility to identify the natural person to whom the scan belongs. Should anonymised data still be considered personal data in a specific country or setting, there should be a legal basis for processing based on consent, the research exception, and/or vital interest. In the current pandemic, sharing anonymised medical data has not been consistent across hospitals, despite GDPR relaxation.
- Being mindful of the time required from data providers. Particularly with the current time constraints and workload, data acquisition should be made as easy and efficient as possible for hospitals, and multiple methods are possible to ensure that for scan acquisition. However, there is no automatic way for sites to provide the additional non-imaging data (e.g. RT-PCR result, patient outcome).
Radiologists must annotate the obtained scans (for instance, provide the segmentation and classification for lung damage in COVID scans) in order to ‘teach’ the AI model; for high accuracy, multiple readers are preferred for each scan. Understandably, radiologists are extremely busy at the moment, thus recruiting and training them for annotations requires consideration and flexibility.
To streamline this step, it is necessary to have an annotations tool in place and to define a protocol for the annotations.
This stage comprises the design of the network architecture. Training the model is one of three main aspects. Here, the outcome of the algorithm is also defined, based on recommendations from medical societies and input from radiologists with experience using CT during the outbreak. Thirdly, it is time to decide how results are presented to the user and design the information visualisation.
4. Clinical validation
The performance of an AI model must be validated on a test set. This dataset is independent of the one used to train the algorithm and contains the ground truth — the radiologists’ final assessment and the results of the PCR test. For a hospital evaluating an AI solution, the test set is the most relevant; its diversity determines if the model is applicable to the specific patient population.
To advance research, the results of the validation should be made public. A clinical or academic site who has been involved in the development can lead the study.
5. Product integration
An AI solution can make a difference for radiologists if properly integrated into their workflow. In a previous article, we explored this topic in depth.
All previous stages come together in the technical documentation that is submitted to a notified body (i.e. an organisation designated to assess the conformity of certain products before market release). The submission contains the clinical evaluation report and the risk assessment for the device.
In Europe, medical device regulations distinguish between Class I, Class IIa/b, or Class III devices. A Class I medical device is registered with a local competent authority via an online database, requiring no upfront regulatory reviews. Currently, many AI products on the market are registered as Class I devices, allowing use for research purposes only. From our experience, a Class IIb certification reflects a solid regulatory basis, with strong clinical and risk management data, suitable for clinical decision support software.
Even if exempted from the regulatory certification due to emergency use, an AI tool on the market should be ready for regulatory approval, meaning its safety and performance should be supported by clinical evidence.
Finally, deploying the AI solution in clinical practice requires site integration, piloting, and testing. The process does not end with deployment — engineering and service are as important to maintain the quality of the solution. User feedback and a solid Post-Market Surveillance plan are how the solution can keep improving.
What is, therefore, a realistic timeline for building a robust and compliant AI solution for COVID-19 on chest CT? Collaboration between hospitals, distribution partners, and AI companies, can speed up the above process, particularly if they have experience certifying and deploying AI. Our careful estimate is that it takes at least four or five months to build a solution and deploy it to the first site, without making any compromises along the way.
To conclude, AI has the great potential to support clinical decisions — now and in the long term — but only if the industry does not take shortcuts, and instead uses this moment to deliver solid solutions and establish trust.