A Machine Learning Approach for Predicting the Time Point of Achieving a Negative Fluid Balance in Patients with Acute Respiratory Distress Syndrome

It is beneficial for patients with acute respiratory distress syndrome (ARDS) to achieve a negative fluid balance at an appropriate time, but it is harmful to achieve a negative fluid balance too early or too late. At present, how to determine the appropriate time point is still unclear. In this work, we developed an XGBoost-based machine learning model for the prediction of the time point of reaching a negative fluid balance in ARDS patients, and explored the relevant influencing factors. A total of 8,685 samples were sampled from 494 ARDS patients with negative fluid balance, including 1,441 positive samples and 7,244 negative samples. As a result, our model shows a considerable prediction performance (AUC: 0.950, accuracy: 92.2%). Furthermore, we found the cumulative fluid balance, mild liver disease, dopamine dosage, fraction of inspiration O2 (FiO2), central venous pressure (CVP) and base excess (BE) are important factors for the prediction accuracy of the model. This result shows our method is expected to assist clinicians better determine the time point of negative fluid balance in ARDS patients, thus optimizing the fluid management strategy in time to obtain better treatment effect.