Machine learning for screening and predicting the availability of medications for children: a cross-sectional survey study

TL;DR

This study uses machine learning to identify factors affecting the availability of medications for children in China and develops a predictive model to support policy improvements.

Contribution

A novel machine learning model (XGBoost) is developed and validated for predicting medication availability for children.

Findings

The XGBoost model outperformed other models with an AUC of 0.915.

Five key factors were identified as most influential in medication availability for children.

A nomogram and clinical impact curve were developed to assess model applicability.

Abstract

The aim of the study was to explore the factors influencing the availability of medications for children, and establish a machine learning model to provide an empirical basis for the subsequent formulation and improvement of relevant policies.

Design: Cross-sectional survey. Setting: 12 provinces, China. Medical doctors from 25 public hospitals were enrolled. All data were randomly divided into a training set and a validation set at a ratio of 7:3. Three prediction models, namely random forest (RF), logistic regression (LR), and extreme gradient boosting (XGBoost), were developed and compared. The receiver operating characteristic curve (ROC) and the associated area under the curve (AUC) were used to evaluate the three models. A nomogram and clinical impact curve (CIC) for availability of medication were developed.

Fifteen of 29 factors in the database that were most likely to be selected were considered to establish the prediction model. The XGBoost model (AUC = 0.915) demonstrated better performance than the RF model (AUC = 0.902) and the LR model (AUC = 0.890). According to the Shapley additive explanation values, the five factors that most significantly affected the availability of medications for children in the XGboost model were as follows: the relatively small number of specialized dosage forms for children; unaffordable medications for children; public education on the accessibility and safety of medication for children; uneven distribution of medical resources, leading to insufficient access to medication for children; and years of service as a doctor. The CIC was used to assess the practical applicability of the factor prediction nomogram.

The XGBoost model can be used to establish a prediction model to screen the factors associated with the availability of medications for children. The most important contributing factors to the models were the following: the relatively small number of specialized dosage forms for children; unaffordable medications for children; public education on the accessibility and safety of medication for children; uneven distribution of medical resources, leading to insufficient access to medication for children; and years of service as a doctor.

Figures