Implementing rapid and effective automated tools to profile COVID-19 patients with respect to their risk of death or hospitalization represents a fundamental challenge for a better allocation of patients and health resources. General clinical risk scores used in the emergency department, such as SOFA and MEWS, when applied to COVID‐19 infection, unfortunately lack adequate sensitivity and specificity to predict mortality associated with COVID‐19 infection [27, 28].

Therefore, in the last months several attempts to improve patients risk stratification was performed developing clinical risk scores also based on ML algorithms. However, the real clinical applicability of the proposed methods is unclear, mainly for methodological issues concerning scarce quality of raw data, heterogeneity and lack of standardization of collected variables, biases in outcome definition and unclear resolution of bias [6, 29].

In order to fulfill the clinical need, overcoming the aforementioned methodological limitation, we have developed the fully automatic AI-SCoRE platform, able to provide a patient risk score in a 0–100 scale, based on the evaluation of only five variables: two demographic data (age and gender), one standardized clinical data of very fast and easy measurement (oxygen saturation) and two quantitative imaging features automatically extracted by a conventional non-contrast chest CT scan (the well-aerated lung volume and the total cardiovascular thoracic calcium).

Outcome was defined as survival, considered as the most reliable data during COVID-19 pandemic, for low reliability of information about oxygen therapy due to fragmented collection of data in emergency and for bias in deployment of treatment (e.g., ICU access) according to hospital resources and pandemic phases.

The AI-SCORE was developed on a retrospective series of 1125 patients referred to 16 Italian hospitals in a limited time period and prospectively validated on 214 consecutive patients during the second wave.

This model showed good performances in the prediction of patient’s outcome in both the first and second waves (AUC = 0.842 and AUC = 0.808), despite the significant improvement of treatment during second wave with subsequent reduction in the overall mortality rate. Notably, the AI-SCORE showed a non-inferior performance compared to models (Vars24 and Vars13) including a larger set of patients’ clinical and laboratory test features, highlighting its clinical value and applicability.

The AI-SCoRE algorithm and platform was able to identify the three risk classes, with only 1.8% of patients misclassified as low risk in the external prospective validation on second wave, all of them with preexisting severe condition determining a strongly reduced expectancy of life.

Our final algorithm included common demographics as age and sex [3, 4, 8, 15], and oxygen saturation, which are all well-recognized predictors of patients’ outcome and crucial parameter to guide patient’s treatment and management [1]. The AI score platform integrates these parameters with chest CT metrics automatically extracted from the entire volume of lung parenchyma and thoracic vessels. The automatic volumetric analysis of lung involvement guarantees a more realistic and accurate measurement of pneumonia severity score [30], in comparison with analysis of isolated 2D slices or even 2D patches used in some previous studies [6], as well as in comparison with semiquantitative score derived from radiologist reading [4, 18], which are affected by limited panoramicity or reader subjectivity.

Moreover, the use of chest CT images instead of XR images guarantees higher sensitivity in the identification of lung parenchyma involvement, with full consideration of slight inflammatory changes [31], and the possibility of a deeper patients’ phenotyping through the quantification of calcium deposits in cardiac valves and thoracic vessels [32].

AI-SCoRE is the first ML COVID-19 risk model integrating cardiovascular calcium. This provides a more comprehensive assessment of patients’ risk. Coronary calcium score is a marker of coronary artery disease and is an established independent predictor of mortality and cardiovascular events in the general population [33]. It was associated with critical illness, adverse major cardiovascular events and death in COVID-19 patients [10, 11, 34, 35]. Total thoracic cardiovascular calcium, which includes also aortic valve and thoracic aorta calcium, resulted a stronger predictor of prognosis in COVID-19 patients if compared to coronary calcium score alone, suggesting that total calcium provides a more comprehensive assessment of systemic atherosclerosis and cardiovascular senescence and left ventricle overload [36]. Its prognostic value may originate from several factors. First of all, endothelium is a target of SARS-CoV-2 infection and diffuse endothelitis has a pivotal role in determining multiple organ damage, hence the chronic endothelial dysfunction and endothelial inflammatory state occurring in atherosclerosis may increase susceptibility to COVID-19 systemic injury [10, 37, 38].

The development of AI-SCoRE was based on a clinical machine learning perspective. This consisted in a first step based on pure ML approach in which interactions between clinical and imaging covariates, and patient outcomes were obtained in a fully data-driven manner. Then, according to recent criticisms about limited predictive power of complete data-driven approaches to COVID-19 [6], a further clinical-driven reduction of the variables was performed with the exclusion of comorbidities, laboratory tests and subjective measurements, potentially affected by limited generalizability due to challenge collection in emergency, inter-laboratories differences in reference values or inter-reader variability for manual measurement.

Differently from most of previous predictive models [3, 4, 6, 30], in our study the outcome was defined as patients survival, considered as the most reliable endopoint during COVID-19 pandemic, due to scarce reliability of other endpoints affected by local protocols and hospital resources.

AI-SCoRE requires only a few and easy to be collected variables, also for poorly equipped hospitals facing a pandemic in a overwhelmed condition. This architecture of the algorithm allowed to avoid missed data, differently from most of previously developed algorithms in which from 30% to more than 50% of patients enrolled did not have all required values collected [18] with imputation of missing data often used [18, 39]. AI-SCoRE model was developed and validated only on a complete dataset, avoiding imputation of missing data, with subsequent more realistic performance metrics and higher applicability in clinical practice.

An important novelty in our procedure is the considerable reduction of time to diagnosis consistent with the urgent public health needs of optimizing health resources. AI-SCoRE may support clinical decision making (home care, mobile hospital quarantine, hospitalization or access to ICU) at hospital admission.

Similarly to previous studies on COVID-19, one limitation of our study consists in potential heterogeneity of data collected in a short time interval from multiple centers in an emergency setting. However, the multicenter approach is mandatory for reducing biases and increase generalizability of the prediction model. Notably, CT parameters have been centrally analyzed in the first step of the study and fully automatically extracted in the final step, significantly reducing the risk of bias. Moreover, validation on second wave cohort, under different public health conditions, confirmed the effectiveness of AI-SCoRE in prediction of patients’ outcome also with optimized treatment.

Although the vaccine significantly reduces the infection rate and COVID-19 severity, the variability of adherence to vaccination policy with persistent spread of infection in non-vaccinated people suggests the potential usefulness of AI-SCoRE platform to improve the allocation of resources based on patients’ risk stratification.