Abstracts

RETURN TO ABSTRACT LISTING


E2019. AI-Based Assessment of COVID-19 Induced Lung Injury and Relationship to Patient Management
Authors
  1. Shayan Vadie; Houston Methodist Hospital
  2. Myra Cocker; Houston Methodist Hospital
  3. John Nance; Houston Methodist Hospital
  4. Shikha Chaganti; Siemens Healthineers
  5. Dorin Comaniciu; Siemens Healthineers
  6. Sasa Grbic; Siemens Healthineers
  7. Nakul Gupta; Houston Methodist Hospital
Objective:
To evaluate an automated AI-based software for the quantitative assessment of lung injury in COVID-19.

Materials and Methods:
76 non-contrast chest CT’s from COVID-19 RT-PCR positive patients (43 female, 54±15 years) were retrospectively evaluated by the Siemens AIRC CT Pneumonia Analysis Prototype (Siemens healthineers; Princeton, NJ), and lesions such as ground glass opacities (GGO), crazy-paving and consolidation were automatically segmented. Lung parameters of percentage of opacity (%O) and percentage of high opacity (%HO; defined as >-200 HU), were quantitatively assessed. %O reflects percent of diseased lung, while %HO reflects denser consolidation. Number of lesions was assessed. Clinical data included age, sex, disposition (discharged from ER, admitted non-ICU, or ICU), oxygen saturation (O2sat), and requirement for intubation. %O and %HO were compared among groups with a Kruskal-Wallis test. Linear regression was used to predict O2sat from %HO.

Results:
7 patients were discharged; 37 were admitted; 32 required ICU care and 29 were intubated. Men had more lesions than women (32.8±20.6 in males vs. 21.7±14.4 female, p=0.002). Patients requiring ICU had higher %HO (10.8±8.7 for ICU, 6.6±7.1 for admitted, 0.73±0.71 for discharged; p<0.001) and %O (35.6±22.3 for ICU, 24.7±20.3 for admitted, and 3.8±4.0 for discharged; p<0.001). Those who were intubated had greater %O (34.7±23.0 intubated vs. 22.9±20.5 non-intubated, p=0.02) and %HO (10.6±9.0 intubated vs. 6.1±6.9 non-intubated, p=0.007). A significant linear regression equation was found predicting O2sat from %HO (F(1,45)=17.25, p<0.001), with an R2 of 0.28. O2sat was equal to 97.0 - 0.321*%HO.

Conclusion:
The extent of lung injury quantified by AI correlates with need for hospital admission, ICU care and intubation. Additionally, AI assessment of %HO was predictive of oxygen saturation. These results suggest that AI-based quantification of lung injury could be incorporated in decision support tools when triaging patients for admission, ICU care and ventilator support. Future work may incorporate these metrics into personalized risk assessment models in combination with demographics, laboratory values and co-morbidities. AI-based assessment of lung injury induced by COVID-19 may enable development of patient management and triaging algorithms for hospital admission and ventilator-support.