Finalist, Italy, for AI-SCoRE (Artificial Intelligence – Sars Covid Risk Evaluation) Initiative
In the initial phase of the SARS-CoV-2 pandemic, while most hospitalized patients did not progress towards the most serious (and often fatal) forms of the disease, others presented with a rapid and unpredictable course, transitioning from mild symptoms to severe respiratory failure within a very short time. Indeed, timing, efficiency, missed diagnoses and inaccurate prognostic evaluation were frequently occurring during the initial care of COVID-19 patients. As a result, in the first wave of pandemic most patients were hospitalized, subtracting resources from other diseases. Furthermore, the outcome for these patients became clear only at the end of the disease’s natural history.
Research groups at the Experimental Imaging Center in our hospital demonstrated that CT imaging was pervasively altered in COVID-19 patients even in the early stages of the disease, and outperformed RT-PCR in detecting SARS-CoV-2 lung infection. However, assessment based only on absolute aerated lung volume is suboptimal because prognosis is influenced by other factors, including COVID-19 related cardiac injury, host immune response to SARS-CoV-2 infection and general frailty of the patient.
We posited that cluster analysis of relevant clinical and imaging information from COVID-19 patients would improve predictability of patient outcomes and care. Follow-up clinical and imaging data from the same patients at 6-12 months after discharge could further increase prognostic accuracy. The use of AI algorithms designed to integrate clinical and imaging data could provide personalized and actionable diagnostic and prognostic COVID-19 scores, towards improved, more cost-effective management of COVID-19 patients. Also, this AI-driven integrated approach would provide other short term and longer-term benefits including improved collaboration and integration among different medical specialties, hospitals and hi-tech stakeholders, a more sustainable approach to diagnosis and importantly, the great potential to extend the methodology to other pathologies with similar clinical needs (see item 26 in this application).
Main objectives.
The specific goal was to develop models that did not offer the mere replication of the diagnostic activity of a single specialist, but information on the staging and possible evolution of the disease, integrating data usually available only to different experts, to provide an overall picture.
Ad hoc algorithms were developed to analyze multiple and heterogeneous inputs (images and structured data) and obtain an easy to interpret single, coherent output for the definition of a diagnostic and prognostic score for COVID-19 patients and prediction of disease risk in healthy and symptomatic individuals. The deployment of this AI platform would provide a rapid evaluation of risk of mild or severe disease in symptomatic COVID-19 patients thus predicting the need for hospitalization and intensive care.
Other important expected benefits were to:
- Decrease pressure on regional/national health systems, increase diagnostic and prognostic efficacy and reduce patient care costs and times.
- Improve risk quantification and therefore individual patient management.
- Carry out a population level risk analysis to develop strategies to deal with a possible new SARS virus epidemic.
- Factor in scalability of the platform to other fields including oncological, cardiovascular, metabolic, and neurological diseases.
The potential positive impact of AI in healthcare is undoubtedly enormous but must be channeled into real-world benefits for healthcare systems and ultimately patients.
A completely automated, and successful, approach such as that of AI-SCoRE represents a model for the development of further solutions in the health sector, allowing to easily export the knowledge gained in a certain region of the world thousands of kilometers away in a moment. The cloud-based platform, in fact, can potentially work for a patient anywhere in the world, without the need to develop local infrastructures, but simply through an internet connection that allows the transfer of the machine learning model. This type of federated integration between different worlds and skills will be increasingly important, including to increase preparedness with respect to local and global health crises.
The application of AI-SCORE at the national level would have prospectively reduced the number of Covid 19 hospitalized patients by 27%. However, the development of effective vaccines resulted in a reduction of hospitalized patients thus making the deployment of the platform on a regional or a national base not timely and convenient. However, this (fortunate) contingency does not decrease the validity and power of our approach, which is now being rolled out over different clinical settings.
Expected results.
By leveraging the power of AI, based on data from thousands of patients of the first wave of pandemic and data from a given patient at the admission to the triage, we expected to generate a simple, actionable and effective prognostic risk score for each patient. The solution was prospectively validated on a cohort of patients of the second wave of pandemic and is now published (Radiol Med. 127(9):960-972. doi: 10.1007/s11547-022-01518-0).
How were patients selected
This project, approved by the ethical review boards of each participating institution, included a retrospective series of 1575 consecutive COVID-19 adults, admitted to the emergency department of 16 hospitals in Northern Italy during the first wave of pandemic and a cohort of 213 consecutive COVID-19 adults prospectively enrolled during the second wave of the pandemic at IRCCS San Raffaele Hospital, Milan, Italy, for prospective external validation.