Clinical Phenotyping with an Outcomes-driven Mixture of Experts for Patient Matching and Risk Estimation.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Cardiology

Background:

• Observational medical data offer valuable insights but require robust matching techniques for patient comparisons.
• Existing methods face limitations in verifying if variables for treatment decisions also predict adverse event risks.

Purpose of the Study:

• To develop a deep mixture of experts approach for simultaneously matching patients and modeling major adverse event risk.
• To validate the model's ability to identify patient phenotypes and predict outcomes using real-world medical data.

Main Methods:

• A deep mixture of experts model was employed to jointly learn patient matching and risk prediction.
• The model was trained on treatment and outcome data, then decomposed into a phenotype-clustering network.
• Validation was performed on a dataset of acute myocardial infarction patients with cardiogenic shock.

Main Results:

• The model achieved an area under the receiver operating characteristic curve of 0.85 ± 0.01 for predicting mortality.
• Five distinct patient phenotypes were identified from pre-treatment information.
• The approach demonstrated effective patient stratification and outcome prediction.

Conclusions:

• The deep mixture of experts model offers a powerful tool for analyzing observational data by jointly addressing patient matching and risk prediction.
• The identified phenotypes can enhance outcomes modeling and support the evaluation of individualized treatment effects.
• This methodology advances the use of machine learning for personalized medicine in critical care settings.