Global Deep Forecasting with Patient-Specific Pharmacokinetics.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Healthcare time series forecasting is crucial for patient monitoring and early detection of adverse outcomes.
• Challenges include variable medication administration and patient-specific pharmacokinetic (PK) properties.
• Existing models struggle to capture individual patient variability effectively.

Purpose of the Study:

• To develop a novel hybrid global-local architecture and a PK encoder for improved healthcare time series forecasting.
• To enhance deep learning models with patient-specific treatment effects.
• To improve accuracy in blood glucose forecasting, particularly during critical glycemic events.

Main Methods:

• Proposed a novel hybrid global-local deep learning architecture.
• Developed a pharmacokinetic (PK) encoder to integrate patient-specific treatment information.
• Evaluated the model on simulated and real-world blood glucose data.

Main Results:

• The PK encoder improved accuracy by up to 16.4% on simulated data and 4.9% on real-world data for individual patients during critical events.
• The hybrid global-local architecture outperformed patient-specific PK models by 15.8% on average.
• Significant accuracy gains were achieved in blood glucose forecasting.

Conclusions:

• The proposed hybrid architecture and PK encoder effectively address challenges in healthcare time series forecasting.
• This approach enhances prediction accuracy for patient monitoring and early detection of adverse events.
• The model shows promise for personalized medicine and improved clinical decision-making.