The persistent increase in healthcare expenditure has become a major challenge for the sustainability of public financing worldwide. Therefore, identifying the characteristics of at-risk population and their predictability for healthcare use is crucial to inform targeted policy and interventions to curb with increasing healthcare use and expenditure. Drawing on three waves of the HILDA survey that included a 'health module', the study applied four machine learning (ML) methods-Random Forest, Gradient Boosting Decision Trees, Extreme Gradient Boosting, and multilayer perceptron neural networks and conventional logistic regression for prediction of non-emergency healthcare use (specifically primary and tertiary inpatient hospital care). Predictive performance for the classifiers was evaluated using accuracy, sensitivity, and specificity measures, and area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Matthews Correlation Coefficient (MCC). Calibration of the models was assessed using Brier score, which measures the mean squared difference between predicted probabilities and observed outcomes with lower values indicating better calibration. Finally, Local Interpretable Model-agnostic Explanations (LIME) was conducted to explain the model's predictive behaviour, while SHAP results are provided for each wave along with a representative SHAP plot as a demonstration which uses the probability-contribution scale. Based on 47,899 observations and 741 variables, our model identified socio-economic factors (age, socio-economic status, private insurance status) and health-related variables (e.g. previous contact with healthcare service) and having a designated doctor to see when sick or for health advice were strong predictors of healthcare use. Between the different ML techniques, Gradient Boosting Decision Trees provided better prediction performance on healthcare use compared with logistic regression across all three waves. Although the standard logistic regression produced AUC of 0.69, had 71% positive predictive value (PPV), and 52% negative predictive value (NPV), with 86% sensitivity and 30% specificity, the ML models produced AUC in the range of 0.68 to 0.76, PPV of 75% to 77%, and NPVs of 61% to 63% with sensitivity ranging between 0.86 and 0.89, specificity between 0.40 and 0.44 and brier scores ranging between 0.11 and 0.28. The novelty of using ML techniques on a large, nationally representative longitudinal household survey data that covers a range of different domains provided more robust estimates on factors influencing future healthcare use (primary and inpatient elective care) which are important to inform resource allocation decisions and priority setting.
Lee et al. (Tue,) studied this question.