Synthetic Model Combination: A new machine-learning method for pharmacometric model ensembling.

Area of Science:

• Pharmacology
• Machine Learning
• Data Science

Background:

• Data-driven pharmacological predictions rely on labeled examples, but data sharing limitations often result in multiple models trained on disparate populations.
• Existing methods for combining models (global selection or ensembling) create a single model, which performs poorly on out-of-domain data due to covariate shift.

Purpose of the Study:

• To introduce a novel instance-wise ensembling method, Synthetic Model Combination (SMC), for improved pharmacological predictions.
• To address the challenge of combining multiple models trained on disparate datasets when data sharing is not feasible.

Main Methods:

• Developed Synthetic Model Combination (SMC), an instance-wise ensembling technique.
• Incorporated a representation learning step within SMC to manage sparse, high-dimensional data.
• Applied SMC to vancomycin dosing predictions as a case study.

Main Results:

• SMC demonstrated effective model ensembling by weighting individual instances based on their domain relevance.
• The method showed improved prediction performance in the vancomycin dosing example.
• The approach is adaptable to various scenarios requiring the combination of multiple predictive models.

Conclusions:

• Synthetic Model Combination (SMC) offers a robust solution for ensembling machine learning models in pharmacological settings with limited data sharing.
• The instance-wise weighting mechanism effectively mitigates issues arising from covariate shift.
• SMC is a versatile method applicable to diverse prediction tasks involving multiple, independently trained models.