Bayesian multitask learning for medicine recommendation based on online patient reviews

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Traditional e-commerce recommender systems struggle with new users (cold-start problem).
• Drug recommendation requires integrating diverse patient data, including structured demographics and unstructured reviews.
• Existing methods may not effectively leverage the nuances within patient feedback for personalized recommendations.

Purpose of the Study:

• To develop a novel drug recommendation model that overcomes the cold-start problem.
• To integrate structured patient data with unstructured review texts for enhanced prediction.
• To apply Bayesian multitask learning for predicting patient satisfaction with medications.

Main Methods:

• Utilized multitask learning to predict multiple satisfaction-related measures from patient reviews.
• Employed topic modeling and sentiment analysis to extract information from unstructured review texts.
• Incorporated Bayesian LASSO for variable selection to filter irrelevant features.

Main Results:

• The proposed model demonstrated superior performance compared to benchmark methods in accuracy and AUC.
• The model proved effective even with small sample sizes and limited features.
• The explainability of the model provides valuable insights for healthcare providers.

Conclusions:

• The developed Bayesian multitask learning approach is a pioneering method for ordinal responses in drug recommendation.
• This model offers a robust solution for the cold-start problem in personalized medicine.
• The model's insights can serve as a valuable reference for doctors, complementing their expertise.