Frequent Emergency Medical Services Utilization Among Older Patients: Exploration and Automatic Identification Using...

An ageing population is associated with increased health care utilization and higher likelihood of older adults becoming frequent users of emergency medical services (EMS). This study aims to explore a natural language processing (NLP) approach using textual notes from EMS to facilitate the identification of older patients at risk of frequent usage.

Retrospective cohort study.

Routinely collected patient data of patients aged ≥65 years from EMS records 2013-2019.

Frequent EMS users were defined as patients with ≥3 EMS uses per year. Descriptive statistics were used for exploratory analysis. For NLP model development, data were split into training (80%) and test (20%) sets, stratified by frequent user incidence. Raw EMS text was preprocessed and transformed using term frequency-inverse document frequency. These vectors served as input to an Extreme Gradient Boosting (XGBoost) algorithm, trained with 5-fold cross validation. Model performance was evaluated using c-statistic, calibration (slope/intercept), sensitivity (recall), precision, and F1 score. Insights into word importance and example predictions were examined and provided.

Among 97,736 patients (median age, 78 years; interquartile range, 72-85), 9.8% were frequent users, accounting for 28.6% of EMS use. On the training set, the supervised machine learning XGBoost algorithm achieved a c-statistic of 0.97 (95% CI, 0.97-0.97), recall of 0.93, precision of 0.47, and F1 score of 0.63. Test set performance was slightly lower, with a c-statistic of 0.92 (95% CI, 0.90-0.92), recall of 0.82, precision of 0.41, and F1 score of 0.54. Calibration metrics indicated good model fit.

An NLP-based model using EMS text data could improve the identification of frequent users. Integration of such tools into prehospital care systems may support earlier identification and tailored intervention strategies for older people frequent in use of acute care.