OBJECTIVES : This study classified total medical expenditures of patients into high costs and non-high costs categories using two machine learning methods (logistic regression and random forest).

METHODS : Data from the Medical Expenditure Panel Survey (MEPS) in 2018 was used for analyses. High-cost patients were identified with thresholds of 10%, 20%, and 33.3% of total medical expenditures. Predictor variables were selected based on prior review and included demographics, presence of comorbid diseases and insurance related (Medicare, Medicaid and private) healthcare costs. Missing or inapplicable data on any of the variables were excluded. Logistic regression and random forest based classifiers were trained on a random subset (80%) of the data using MATLAB R2020b. Five-fold cross validation was used to prevent overfitting of the trained models. The two classification models were compared on metrics including accuracy, area under curve (AUC), specificity and sensitivity.

RESULTS : From a total of 30,461 individuals, 21,648 individuals were included after removing individuals with missing data. The training dataset consisted of 17,319 patients and the rest were included in the validation dataset. Both random forest and logistic models showed comparably high accuracy (93.6% vs 93.0%), for the training data with classifying threshold of 33.3%. AUC for both models was estimated at 0.9613 and 0.9645, respectively. Similar accuracy and AUC were observed for models with thresholds of 10% and 20%.

For the validation data with 33.3% threshold, random forest model showed similar estimates as compared to logistic model for specificity (97.99% vs 97.13%) and sensitivity (85.07% vs 85.70%), respectively. Comparable results were observed for validation data with 10% and 20% thresholds. Amount paid by private insurance, amount by Medicare, amount by Medicaid, and age had the highest influence on total medical expenditures.

CONCLUSIONS : Both logistic and random forest models showed high accuracy in classifying patients into high and non-high costs categories across different thresholds.