Abstract
We propose a robust in-time predictor for in-hospital COVID-19 patient's probability of requiring mechanical ventilation. A challenge in the risk prediction for 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 performing mechanical ventilation for individual patients. The model incorporates the Multi-task Gaussian Process using observed values to learn the posterior joint multi-variant 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. 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.
Introduction
Predicting which COVID-19 patients will deteriorate to the point of needing mechanical ventilation (MV) is a high-stakes clinical challenge. While Electronic Health Records (EHRs) provide a wealth of data, they are notoriously "messy," with vital signs and lab tests sampled asynchronously and at irregular intervals. Traditional models often treat patient status as a static snapshot, failing to capture the dynamics of disease progression and frequently producing "jittery" or self-conflicting risk scores over time. This paper introduces a novel deep learning pipeline designed to produce a smooth, logical, and accurate risk trajectory in real-time.
Methods
The researchers propose a two-stage end-to-end architecture called MGP-MS, which integrates statistical modeling with deep learning to handle the complexities of clinical data.
Multi-task Gaussian Process (MGP)
To address missing data and irregular sampling, the model uses an MGP. It treats each clinical variable (e.g., blood pressure, oxygen saturation) as a "task" and learns the correlations between them. This allows the model to impute missing values onto a unified, regular time grid while accounting for the inherent uncertainty in clinical measurements.
Multi-objective Self-attention Network
The imputed data is fed into a Transformer-style encoder. To ensure the model remains "in-time" (using only information available at a specific moment), the authors implemented a novel block-wise upper-triangular positional encoding layer. Rather than training the model to predict outcomes only at the final hour, they utilize a multi-objective learning framework. The model is optimized to predict the correct risk score at every time point simultaneously, forcing it to "foresee" the outcome early and ensuring a more stable risk trajectory.
Experimental Setup
The study utilized the Optum® de-identified COVID-19 EHR dataset, focusing on a cohort of 9,532 hospitalized patients. The model processed 16 lab tests, 9 vital signs, 21 medicine categories, and 25 demographic variables, with data aggregated into 4-hour windows over the first 72 hours of admission.
Key Results
The MGP-MS model was compared against several baselines, including XGBoost, Cox Proportional-Hazard, and deep learning models like LSTM and GRU-D.
- Superior Accuracy: The model achieved significantly higher performance in terms of AUC and AUPRC, particularly in the critical early hours after admission (e.g., an AUC of 0.7920 at admission vs. 0.7378 for a standard GRU).
- Consistency and Robustness: The multi-objective approach improved risk score consistency by 160% to 313% over traditional methods, providing a clearer trend for clinicians.
- Clinical Insights: Feature importance analysis identified Urine Output (UROUT), Temperature (TEMP), and Total Calcium as the top three predictors for mechanical ventilation.
Conclusion
By combining Multi-task Gaussian Processes with a multi-objective self-attention network, the MGP-MS model provides a robust solution for real-time risk prediction in COVID-19 patients. The ability to generate consistent risk trajectories from irregular EHR data offers a reliable early warning system for clinicians. Future work will focus on external validation across different hospital systems and exploring the complex dynamics of medicine interactions.
Summary updated