The digital profile of a physician as a new approach to improving recommendation systems for prescribing drugs

Abstract

Introduction

Artificial intelligence (AI) is increasingly integral to healthcare’s digital transformation, enhancing areas such as disease prevention and control. Traditional AI models focus on patient-centered data from electronic medical records (EMRs). This study aims to improve drug prescription recommendation systems by developing a physician’s digital profile (DP), incorporating physician characteristics into machine learning models.

Methods

Visit-level EMR clinical data (2018-2023) of Moscow’s children’s polyclinics was collected and contained information on 174,806 appointments with 30 physicians and 53,453 patients. We developed multi-output regression models (LGBMRegressor) using gradient boosting algorithms. Models were trained with and without the physician’s DP, which included personal, drug, and nosological features adapted for machine learning.

Results

Incorporating the physician’s digital profile significantly improved model performance. The LGBMRegressor with DP reduced the average RMSE from 4.20 to 0.74. Models utilizing the DP more accurately predicted drug prescriptions and better captured individual physician behaviors.

Discussion

Developing a physician’s digital profile has been shown to enhance the predictive performance of machine learning models in drug prescribing. This approach supports the optimization of clinical decision support systems, fostering more safe and personalized care and potentially reducing prescribing errors. Moreover, aggregated model outputs may inform drug demand forecasting and assist in optimizing pharmaceutical inventory planning at the population level.