Background: Acute respiratory distress syndrome (ARDS) is a serious conditionthat occurs in around 10 % of mechanically ventilated patients in intensivecare units. Delayed or even missing diagnosis contributes to the reasons for thevery high mortality rate of up to 40 %. To address this shortcoming, Artificial Intelligence(AI) methods for the prediction and classification of ARDS have beenproposed in literature. For the training of certain AI models, it is necessary todetermine an onset of ARDS in the data to use associated data points as trainingor test data. This is particularly important for secondary databases, which oftendo not come with respective documentation.
Methods: In this study, the ARDS onset time of 200 patients from 4 differentdatabases was determined by two intensive care physicians and by a rule-basedalgorithm whose rules are based on the criteria of the Berlin Definition, which is considered the standard for ARDS diagnosis. For manual annotation, a customweb tool was used to visualize the selected patient data. Finally, the results ofthe automatic and manual onset determination for ARDS are compared usingstatistical means.
Results: Depending on the configuration, the automatic onset of the rule-basedalgorithm matches the manual onsets in 71.50 % - 86.36 % of the cases. Overall,it was shown that the respective results of the two methods are close to eachother on average.
Conclusions: In our analysis, we were able to show that a rule-based algorithmwith a temporal component can be used to determine an ARDS onset in timeseriesdata with sound accuracy in comparison with the manual onset created bytwo physicians. In summary, it was shown that it is reasonable to generate theonset automatically to make big databases accessible for the development of AImodels in the context of ARDS.