Dynamic Bayesian network for predicting physiological changes, organ dysfunctions and mortality risk in critical trauma patients

Qi Chen, Bihan Tang, Jiaqi Song, Ying Jiang, Xinxin Zhao, Yiming Ruan, Fangjie Zhao, Guosheng Wu, Tao Chen, Jia HeBMC Medical Informatics Decis. Mak 2022 · DOI

TL;DR

This paper introduces a Dynamic Bayesian Network (DBN) model to predict physiological changes, organ dysfunctions, and mortality risk in critical trauma patients in real-time. By leveraging temporal patient data, the DBN demonstrates superior predictive accuracy compared to traditional scoring systems, offering a promising tool for clinical decision-making and early intervention.

What Problem Does This Paper Address?

Critical trauma patients face a high risk of mortality, making accurate and timely prediction of their condition crucial for effective management. Existing prognostic scoring systems (like SAPS II and APACHE II) often rely on static, baseline data and can be complex to use, limiting their real-time applicability. Furthermore, patient status in the ICU is dynamic, with physiological variables constantly changing. This necessitates a predictive model that can adapt to these changes and provide continuous risk assessment, moving towards a more personalized, precision medicine approach. The challenge lies in modeling the complex, temporal relationships between numerous physiological variables to accurately forecast future outcomes.

Methods

The core of this research is the development and validation of a Dynamic Bayesian Network (DBN). DBNs are an extension of Bayesian Networks that explicitly model temporal dependencies, making them suitable for time-series data.

  • Structure Learning: The DBN structure was learned using a data-driven approach with logical constraints, specifically employing the PC algorithm. This algorithm identifies conditional dependencies between variables.
  • Parameter Learning: Once the network structure was defined, parameters (conditional probabilities) were estimated using maximum likelihood estimation.
  • Dynamic Prediction: The DBN models the state of physiological variables at time t based on their states at previous time slices (t-1). This allows for predictions of future physiological states, organ dysfunctions, and mortality risk.
  • Model Evaluation: The DBN's predictive accuracy was evaluated for:
  • Physiological Changes: Measured by absolute differences and difference ratios between true and predicted values.
  • Organ Dysfunctions: Assessed using sensitivity, specificity, and accuracy based on criteria derived from the Multiple Organ Dysfunction Score.
  • Mortality Risk: Evaluated using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values, compared against traditional scoring systems (SAPS II, APACHE II, ISS) and a logistic regression model.

Datasets

  • Training Dataset: MIMIC-III (Medical Information Mart for Intensive Care database), a publicly available ICU database from Beth Israel Deaconess Medical Center.
  • Validation Dataset: Patient data from the Burn and Trauma ICU of Changhai Hospital (ICU-CH) in East China.

The study utilized existing, large-scale ICU datasets, rather than introducing a new benchmark dataset.

Key Insights and Ideas

  • Temporal Dynamics Matter: The DBN's ability to capture time-dependent relationships is key to its improved predictive power over static scoring systems. Patient status evolves, and a model that reflects this evolution is more accurate.
  • DBNs as a Robust Tool: The study demonstrates that DBNs are a powerful and robust method for modeling complex, non-linear, and interactive relationships in temporal medical data, outperforming traditional methods in critical care prediction.
  • Real-time Clinical Utility: The development of an interactive web-based calculator makes the DBN model accessible for practical clinical use, enabling physicians to input current patient data and receive real-time predictions.
  • Predicting Beyond Mortality: The model's capability to predict not only mortality but also physiological changes and organ dysfunctions provides a more comprehensive picture for clinical decision-making and early intervention.

Key Results

  • Physiological Variable Prediction: The DBN model achieved high prediction accuracy for physiological variables, with difference ratios between real and predicted values generally within 5% for both 24-hour and 48-hour predictions in both datasets.
  • Organ Dysfunction Prediction: The model showed good predictive ability for renal, hepatic, cardiovascular, and hematologic dysfunctions, with accuracies exceeding 0.8 in both MIMIC-III and ICU-CH.
  • Mortality Risk Prediction: The DBN model achieved superior AUC scores for mortality risk prediction compared to SAPS II, APACHE II, and ISS. For example, in MIMIC-III at 24 hours, the DBN model achieved an AUC of 0.977 (95% CI: 0.967–0.988), outperforming SAPS II (0.954) and APACHE II (0.948). Similar high performance was observed at 48 hours and in the ICU-CH validation set.
  • Calibration: Calibration plots indicated that the predicted mortality from the DBN model closely approximated actual outcomes.

Limitations and Future Work

  • Variable Selection: The model included variables from SAPS II and APACHE II. While this simplifies the model, more recent versions of these scoring systems exist, and future work could explore incorporating additional variables.
  • Complications Data: The study did not include the occurrence time of complications like sepsis, which could further enhance predictive accuracy.
  • External Validation Scope: While validated in an external dataset, the data came from two high-level hospitals. Future work should involve broader validation across diverse hospital settings to assess generalizability.
  • Individual Uniqueness: The DBN model provides deterministic outputs for identical inputs. Future research could explore incorporating more individual patient characteristics to account for unique patient variability.
  • Web Calculator Enhancement: The authors suggest further improving the web calculator to enhance user convenience for physicians.

Summary updated

BibTeX
@article{chen2022dynamic,
  author = {Qi Chen and Bihan Tang and Jiaqi Song and Ying Jiang and Xinxin Zhao and Yiming Ruan and Fangjie Zhao and Guosheng Wu and Tao Chen and Jia He},
  title = {{Dynamic Bayesian network for predicting physiological changes, organ dysfunctions and mortality risk in critical trauma patients}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2022},
  doi = {10.1186/s12911-022-01803-y},
}
Figure