Real-time Prediction for Mechanical Ventilation in COVID-19 Patients using A Multi-task Gaussian Process Multi-objective Self-attention Network

Goal: This paper proposes a robust in-time predictor for in-hospital COVID-19 patient's probability of requiring mechanical ventilation. The data-driven model outputs a highly consistent and robust risk score trajectory for a patient with COVID-19. The patient’s risk score at a particular time point indicates the risk of the patient’s condition worsening to the point of requiring mechanical ventilation. The model could serve as an early warning monitor system for in-hospital patients. The successful and accurate prediction of such risk scores could help physicians to provide earlier respiratory support for the patient and reduce mortality. Methods: A challenge in the risk prediction of COVID-19 patients lies in the great variability and irregular sampling of patient’s vitals and labs observed in the clinical setting. Existing methods have strong limitations in handling time-dependent features' complex dynamics, either oversimplifying temporal data with summary statistics that lose information or over-engineering features that lead to less robust outcomes. We propose a novel in-time risk trajectory predictive model to handle the irregular sampling rate in the data, which follows the dynamics of risk of intubation for individual patients. The model incorporates the Multi-task Gaussian Process (MGP) using observed values to learn the posterior joint multivariant conditional probability and infer the missing values on a unified time grid. The temporal imputed data is fed into a multi-objective self-attention network for the prediction task. A novel positional encoding layer is proposed and added to the network for producing in-time predictions. The positional layer outputs a risk score at each user-defined time point during the entire hospital stay of an inpatient. We frame the prediction task into a multi-objective learning framework, and the risk scores at all time points are optimized altogether, which adds robustness and consistency to the risk score trajectory prediction. The parameters of the MGP are jointly optimized with the downstream multi-objective self-attention network. Result: The model produces in-time robust risk score predictions for each patient – a consistently ascending risk score trend for each patient who would perform mechanical ventilation in the following days and a descending trend for those who would not. On  E-mail addresses: kai.zhang.1@uth.tmc.edu (K. Zhang), Xiaoqian.jiang@uth.tmc.edu (X. Jiang), Siddharth.Karanth@uth.tmc.edu (S. Karanth), Bela.Patel@uth.tmc.edu (B. Patel), Robert.Murphy@uth.tmc.edu (R. Murphy) * Corresponding author the other hand, conventional models’ risk score trajectory predictions fluctuate and often generate self-conflicting results (e.g., the risk at discharge is higher than the risk 12 hours ago). Our experimental evaluation on a large database with nationwide in-hospital patients with COVID-19 also demonstrates that it improved the state-of-the-art performance in terms of AUC (Area Under the receiver operating characteristic Curve) and AUPRC (Area Under the Precision-Recall Curve) performance metrics, especially at early times after hospital admission. We also evaluate the feature importance of each input feature contributing to the final output. Discussion: The ability to distinguish potentially deteriorating patients from the rest also facilitates reasonable allocation of medical resources, such as ventilators, hospital beds, and human resources. We believe there is a need to develop novel risk prediction models to serve as early warning systems to benefit both in-hospital patients and physicians. Our model makes full use of the in-hospital patient’s temporal physiological data, including laboratory tests and vital signs, drug prescriptions, and patient demographic information, etc., to generate robust outcomes. Conclusion: The availability of large and real-time clinical data calls for new methods to make the best use of them for real-time patient risk prediction. It is not ideal for simplifying the data for traditional methods or making unrealistic assumptions that deviate from observation's true dynamics. We demonstrate a pilot effort to harmonize cross-sectional and longitudinal information with multi-objective targets.