Abstract
Cardiac arrest is a deadly condition caused by a sudden failure of the heart with an in-hospital mortality rate of ∼80%. Therefore, the ability to accurately estimate patients at high risk of cardiac arrest is crucial for improving the survival rate. Existing research generally fails to utilize a patient’s temporal dynamics. In this paper, we present two dynamic cardiac risk estimation models, focusing on different temporal signatures in a patient’s risk trajectory. These models can track a patient’s risk trajectory in real time, allow interpretability and predictability of a cardiac arrest event, provide an intuitive visualization to medical professionals, offer a personalized dynamic hazard function, and estimate the risk for a new patient.
Introduction
In-hospital cardiac arrest remains a critical challenge in clinical care, characterized by a staggering 80% mortality rate. While physiological deterioration often manifests up to eight hours before an event, traditional Early Warning Scores (EWS) frequently fail to capture these subtle temporal dynamics. These scores are typically static and based on expert opinion rather than data-driven longitudinal analysis. Furthermore, the "rare event" nature of cardiac arrest makes traditional survival analysis difficult, as event times are not easily aligned across heterogeneous patient populations.
To address these limitations, we introduce DYNACARE (Dynamic Cardiac Arrest Risk Estimation). This framework treats a patient's health as a latent, time-varying Cardiac Arrest Trajectory (CAT). By applying dynamic stochastic volatility models—techniques originally developed for financial market forecasting—DYNACARE tracks a patient’s risk in real-time, providing personalized and interpretable hazard functions.
Methodology
The DYNACARE framework models cardiac arrest as a hidden process rather than a discrete, static event. We propose two distinct model architectures to capture different temporal signatures of deterioration:
- Markov Switching Model (MSM): This model assumes a patient transitions between discrete states (e.g., "healthy" and "risky"). It is designed to detect abrupt changes in the gradient of physiological signals, identifying the moment a patient enters a high-risk regime.
- Threshold Model (THR): This model treats risk as a continuous, monotonically increasing value that eventually crosses a critical threshold. It is particularly effective for capturing "slow-burn" deterioration.
Stochastic Volatility and Volatility Clustering
A key innovation in DYNACARE is the use of Stochastic Volatility (SV). In financial econometrics, "volatility clustering" refers to the tendency for large changes in asset prices to be followed by large changes. We observe a similar phenomenon in human physiology: as a patient nears a cardiac event, the variance in their vital signs often spikes and exhibits strong self-correlation.
To implement these models, we utilize Semi-Supervised Particle Filtering. By using backward smoothing on historical data where the arrest time is known, the model learns the latent patterns that precede an event. For new patients, the model employs Stratified Bootstrapping, drawing parameters from demographic strata (age and gender) to account for baseline physiological differences.
Experimental Setup
The study utilized the MIMIC-II database, focusing on adult admissions who experienced an asystole event. To ensure robust temporal modeling, the cohort was filtered to include patients with at least 40 discrete time slices (approximately 6.5 days of data), resulting in a final study group of 108 patients. The models utilized five primary vital signs: heart rate, respiratory rate, body temperature, and systolic/diastolic blood pressure.
Results and Discussion
Our evaluation demonstrates that DYNACARE significantly outperforms traditional static clinical scoring methods.
- Model Performance: The Markov Switching Model (MSM) emerged as the strongest individual performer, effectively capturing the "regime change" from stability to instability. However, the best overall performance was achieved by an ensemble of MSM, THR, and standard Logistic Regression, suggesting that cardiac arrest can manifest through both abrupt jumps and gradual trends.
- Clinical Lead Time: The models successfully identified risk signatures up to 80 hours (20 time periods) before the actual arrest. This provides a substantial window for clinical intervention compared to traditional threshold-based alarms.
- Interpretability: Unlike "black-box" predictive models, DYNACARE provides a visual risk trajectory. This allows medical professionals to see risk climbing over time, making the output more intuitive and actionable.
Conclusion
DYNACARE demonstrates that borrowing techniques from financial forecasting can yield powerful insights into human health. By modeling the heart's "market" and accounting for stochastic volatility, we can provide personalized, dynamic risk estimations for cardiac arrest. Future work will focus on incorporating a broader range of features, such as laboratory results and unstructured clinical notes, and extending the framework to other critical conditions like sepsis.
Summary updated