Deep Learning Algorithm Identifies Risk of ARDS or In-Hospital Mortality in Mechanically Ventilated Patients and Validation on COVID-19 Patients

Introduction: ARDS is often unrecognized. We aimed to developed and validate a deep learning algorithm using EHR structured data to identify patients at risk for ARDS or in-hospital mortality in pre-COVID and COVID-19 inpatient population. Method: The initial cohort consisted of adult ICU patients on mechanical ventilation (MV) from 2017 to 2018 at 3 hospitals at Montefiore Medical Center. Independently trained physician reviewed all chest x-rays in patients with PaO2/FiO2 ratio ≤ 300 to determine ARDS as per Berlin criteria. We used 66 discrete EHR variables to generate 135 features based on variable types. We longitudinally sampled hourly and missing data were forwardly imputed. We trained a preliminary Long Short-Term Memory network to determine the probability of ARDS or in-hospital mortality in the 2017-2018 cohort. Using active learning process, we applied this preliminary model to additional MV patients admitted between 2016-2017 and 2018-2019 to expand the sample size with additional ARDS and non-ARDS patients. The expanded pre-COVID cohort (2016-2019) was split into training and validation cohort (80 and 20%). A new and final LSTM model were trained on the pre-COVID training cohort. Validations were performed on the pre-COVID validation cohort and a population of mechanically ventilated and non-ventilated COVID-19 patients between March-April 2020. Result: The pre-COVID cohort included 3905 MV patients in the ICU from final 2016-2019. 1646 (42.4%) had ARDS and 1033 (26.5%) died during hospitalization. COVID-19 cohort included 5672 patients and 907 (16%) died. 803 (14.2%) were mechanically ventilated including 583 ARDS patients(10.3% of COVID cohort and 72.6% of MV COVID patients) and 418 (52.1%) who died (Table 1). The model ROC was 0.78 for pre-COVID validation cohort. At model cutoff 0.90, the sensitivity was 86% and specificity was 57% (Table 1) In the COVID-19 cohort, the ROC was 0.83 (sensitivity of 70% and specificity of 84%). On the COVID-19 +MV sub-cohort, the model ROC was 0.70 and accurately diagnosed 91.6% of ARDS or death. The model was able to identify ARDS or death close to the time of intubation and ARDS label time in both pre-COVID and COVID-19 cohort (Table 1). Conclusion: Our LSTM algorithm accurately and timely identified the probability ARDS or in-hospital in mechanically ventilated patients. When we applied the model to a larger cohort of hospitalized patients with COVID-19 which had a higher prevalence of ARDS, the sensitivity, and positive predictive value remains high but specificity is much lower among MV patients.