Abstract
Electronic health records (EHR) present significant challenges for predictive modeling due to their non-stationary, heterogeneous, and sparse nature. While imputation is commonly used to address missing values, existing methods often fail to account for the confidence or fidelity of these imputed estimates. This paper introduces a novel variational recurrent network that estimates the distribution of missing variables to represent uncertainty. By integrating an uncertainty-gating mechanism into a recurrent architecture, the model propagates fidelity information over time to improve downstream mortality prediction. Validated on MIMIC-III and PhysioNet datasets, the proposed model outperforms state-of-the-art methods, demonstrating that modeling imputation uncertainty is critical for clinical decision support.
Introduction
The "messiness" of Electronic Health Records (EHR) is a well-known hurdle in medical informatics. Data is often sparse because clinical tests are ordered selectively, noisy due to sensor errors, and non-stationary as patient states evolve. Traditional approaches rely on imputation—filling missing values with estimates. However, treating these "guesses" as ground truth can lead to error propagation. If a model predicts a patient's mortality based on an incorrectly imputed blood pressure value, the clinical utility of the prediction is compromised.
The core contribution of this work is a model that not only imputes missing values but also estimates its own confidence in those imputations. By quantifying this uncertainty, the model can dynamically weight the influence of imputed data during the prediction process.
Methodology
The proposed framework consists of three integrated components: stochastic imputation, uncertainty-aware recurrence, and joint learning.
Stochastic Imputation via VRNN
Instead of deterministic imputation, the model utilizes a Variational Recurrent Neural Network (VRNN) to predict a probability distribution (mean and variance) for missing variables. The variance serves as a direct proxy for uncertainty: high variance indicates low confidence in the imputed value.
Uncertainty-Gated Recurrence (GRU-U)
The central innovation is the GRU-U cell, a modification of the standard Gated Recurrent Unit. It introduces an attention mechanism ($\alpha_t$) that gates inputs based on their calculated uncertainty. When the model is uncertain about an imputed value, the gate reduces that value's impact on the hidden state update, effectively "muting" unreliable information.
Joint Learning and Optimization
The model is trained end-to-end using a composite loss function. This function balances data reconstruction (VRNN loss), imputation accuracy (MAE), and classification performance (mortality prediction). To address the extreme class imbalance typical of mortality data, the authors employ Focal Loss, which prioritizes learning from rare positive cases (deaths).
Experimental Results
The model was evaluated on two benchmark datasets: MIMIC-III (missingness rate ~94%) and PhysioNet Challenge 2012 (missingness rate ~81%).
Performance Comparison
The proposed model achieved state-of-the-art results in mortality prediction:
- MIMIC-III: AUC of 0.865 and AUPRC of 0.416, outperforming the previous best model, BRITS (0.863 / 0.414).
- PhysioNet: AUC of 0.832, significantly higher than BRITS (0.824).
Ablation Study
To confirm the impact of the uncertainty-gating mechanism, the authors replaced the GRU-U cell with a vanilla GRU. This change led to a substantial performance drop (AUC decreased from 0.832 to 0.794 on PhysioNet), validating that the explicit modeling of fidelity is the primary driver of the model's predictive power.
Discussion and Conclusion
A key insight from this study is the high correlation between the model's predicted uncertainty and the actual Mean Absolute Error (MAE) of its imputations. This confirms that the model successfully "knows what it doesn't know." While other models might achieve slightly higher raw imputation accuracy, the proposed model's focus on fidelity-aware prediction makes it more robust for downstream clinical tasks.
Future work includes extending this uncertainty-aware approach to other clinical outcomes, such as length-of-stay prediction, and exploring how these confidence scores can be visualized for clinicians to improve trust in AI-driven diagnostic tools.
Summary updated