Multi-label text mining to identify reasons for appointments to drive population health analytics at a primary care setting

While much research has been conducted regarding population health analytics (PHA), there is limited research related to text mining in that area. In this research, a novel multi-label text mining model is developed to analyze and categorize the reasons for medical appointments at a primary medical center serving a rural population. The model converts an unstructured, unsupervised text corpus to a structured supervised multi-label text corpus by using look-up wordlists defined through expert domain knowledge (EDK). The text dataset contains the reasons patients made appointments in 2019. The appointment reasons were grouped into 27 categories. Each appointment reason text is tagged to its associated group (label) using associated look-up wordlists. Then, the tagged corpus is used to develop a multi-label text classification model using machine learning algorithms. Two resampling models (balanced classifiers and SMOTE) are considered to adjust for the unbalanced created labels. The classifiers and models are tested in three steps using validation, testing, and implementation datasets. Both models performed well, but the SMOTE model is more generalizable, reliable, and consistent than the balanced classifiers. The label-set performance measures are equal to or greater than 77.9% for the balanced classifiers and greater than 80% for the SMOTE model. The label-based testing performance measures using both models and all classifiers are generally greater than 90% for all labels. Finally, the PHA showed that the follow-up and well-check (WC) physical patients are the largest populations (32% and 16.54%, respectively). Besides, the populations are different based on some factors such as age, insurance, show rate, punctuality rate, and scheduling type, while the populations are very similar based on other factors such as ethnicity and gender.