Forecasting Trajectories of Physiological Mechanics with Sparse Clinical Data Using a Data Assimilation and Machine...

Area of Science:

• Physiological modeling
• Computational biology
• Machine learning in healthcare

Background:

• Clinical decision-making relies on understanding complex physiological dynamics, but faces challenges from sparse observational data and patient heterogeneity.
• Current data assimilation (DA) methods struggle to forecast future physiological states because they cannot predict evolving model parameters.

Purpose of the Study:

• To develop a hybrid data assimilation (DA) and machine learning (ML) pipeline for estimating and forecasting individual physiological processes.
• To overcome limitations in current DA workflows for predicting evolving physiological parameters.
• To enhance clinical decision support systems by providing accurate, patient-specific forecasts.

Main Methods:

• Developed a hybrid pipeline integrating DA with longitudinal ML forecasting models.
• Stacked DA-estimated posterior empirical distributions of physiological parameters with ML models.
• Validated the pipeline using both synthetic and real-world EHR data in an ICU glycemic management use case.

Main Results:

• The integrated pipeline successfully estimates and forecasts individual future physiological processes by predicting ordinary differential equation (ODE) model parameters.
• Demonstrated the pipeline's effectiveness in a glycemic management scenario using EHR data.
• Quantified uncertainties associated with the forecasts, providing reliable predictions.

Conclusions:

• The novel DA-ML hybrid pipeline offers a robust method for forecasting patient-specific physiological dynamics.
• This approach significantly enhances the potential for next-generation clinical decision support tools.
• Accurate forecasting of physiological parameters can address data sparsity and heterogeneity challenges in clinical predictions.