Humanitas AI Center on European Radiology: the importance of chest CT to establish the best therapeutic pathway
Published on European Radiology, the official channel of the European Society of Radiology, here’s the study conducted by the team of radiologists, data scientists and specialists of Humanitas to predict the need for oxygenation and intubation for COVID-19 patients.
Lombardy has been the epicentre of the Covid-19 pandemic since March 2020. The health system suffered from a shortage of beds in intensive care and oxygen therapy devices. Most patients underwent chest CT scans at the time of admission, then interpreted only visually. Given the proven effectiveness of quantitative CT scan analysis in the context of ARDS (Respiratory Distress Syndrome), we decided to test it as an outcome indicator for Covid-19.
How the project started
After having read the first articles on CT characteristics of pneumonia from COVID- 19, we realized that the common thread between CT picture and the seriousness of the disease seemed to be the number and density of lung lesions – parameters that we could easily quantify with a number, instead of just describing them in a report. This information would have made it possible to make very precise statistical analyses. However, extracting the data from the CT scans was only the first step. We needed to relate them to the clinical progress of these patients. We then extracted the clinical data of more than 200 patients in a cloud database – overall, we filled over 11000 spreadsheet boxes in less than ten days.
The statistical analysis immediately gave us encouraging results: the percentage of lung volume “attacked” by the Coronavirus was a powerful index to predict the need for oxygen therapy and intubation of patients. However, to complete the model, we needed to be able to validate it in a real life simulation. We couldn’t wait for a prospective study, so we asked our colleagues in the Artificial Intelligence team to help us with the complex technologies used to train the neural networks. That allowed us to validate our results and obtain precise reference values that were ready to use in the triaging of these patients right from the diagnosis. All this within a few days.
Overall, the study involved a multidisciplinary team of more than ten specialists from different generations – from the young student to the experienced department director. This approach promoted innovation and the achievement of ambitious goals.
We carried out a retrospective study on Covid-19 patients that had been hospitalized from January 25, 2020 to April 28, 2020, and underwent CT scans upon admission due to respiratory symptoms such as dyspnea or desaturation. We then collected clinical data from the patients, including the use of oxygen therapy during admission. The study involved 222 patients – among them, 163 were men with an average age of 66 years. 75% of them received oxygen therapy (20% intubation rate). The compromised lung volume was the most accurate outcome indicator.
The quantitative TAC provides new parameters in relation to Covid-19. The compromised lung volume requires support by oxygen therapy and intubation and is a significant risk factor as it could cause death in hospital. Therefore, the quantitative CT scan can be a useful tool for the triage process of Covid-19 patients.
What impact will the project’s conclusions have on clinical practice? We hope that our research will contribute to a new approach to the diagnosis of COVID-19 pneumonia. The evaluation of the chest CT scan of patients with a quantitative method allows for early identification of patients who will need extra care. This method can also help to better identify those patients who can be treated at home or with non-invasive oxygen therapy. This strategy would also allow an accurate management of the beds available in the hospital, which have proved to be a limited resource during the pandemic.
How do AI and medicine integrate into Humanitas?
The association between AI and medicine is still being investigated, but all the opportunities for collaboration suggest that artificial intelligence could have a wide potential in clinical practice. Tools such as the one developed in our research could – in the near future – be fully automated thanks to artificial intelligence and would lead the way to an “artificial medicine”. Doctors could make decisions not only based on their experience, but also on real prediction; of the clinical course of their patients supported by scientific evidence elaborated by artificial intelligence.