Using Machine Learning to Predict Total Charges and Drivers of Cost following Transsphenoidal Surgery for Pituitary Tumor

Background Despite an increasing focus on decreasing costs in health care, total charges for transsphenoidal pituitary surgery are steadily rising. There currently exist no models that directly predict total charges after pituitary surgery, and drivers of high cost in this patient population are poorly understood. Here, we use machine learning (ML) techniques and a large, national database to build ensemble models that predict total charges following transsphenoidal pituitary surgery. We then interrogate the ensembles to identify variables that best predict total charges.

Methods Using the National Inpatient Sample (NIS), a multi-institution registry of hospitalizations in the United States, we created a training dataset of 15,487 patients who underwent transsphenoidal pituitary surgery between 2002 and 2011. Fifty-four variables, including patient demographics, resection type, tumor histology, preoperative comorbidities, postoperative complications, length of hospitalization (LOH), and hospital characteristics were collected for each hospitalization. Thirty-one ML algorithms were trained to predict total charges and optimized to minimize gamma deviance. The most predictive algorithms were combined to form an ensemble model and the ensemble validated using data excluding from model training. To identify the strongest predictors of total charges, the relative importance (RI) of each variable to the ensemble predictions was calculated. Given the overwhelming influence of LOH on our first ensemble, a second ensemble excluding LOH as a predictor was built to identify drivers of cost whose impact may have been obscured.

Results Mean total charges per hospitalization for transsphenoidal pituitary surgery were $50,0027. An ensemble model comprising a RuleFit regressor and two Gradient Boosted Tree regressors best predicted total charges (gamma deviance = 0.2457). Increasing LOH was the strongest predictor of total charges (RI = 1.0), followed by nonelective admission (RI = 0.17), non-southern hospital region (RI = 0.17), minority race (RI = 0.10), and patient age (RI = 0.03). Each hospital day increased the total predicted charges by approximately $5,000.

A second ensemble excluding LOH as a variable predicted the outcome of interest, though with less fidelity than the initial ensemble (gamma deviance = 0.3176). The strongest predictors of total charges for the non-LOH ensemble were nonelective admission (RI = 1.0), non-southern hospital region (RI = 0.69), minority race (RI = 0.49), the presence of fluid or electrolyte abnormalities (RI = 0.37), and postoperative CSF leak (RI = 0.36).

Conclusion A machine learning ensemble directly predicts total charges for transsphenoidal pituitary surgery with good fidelity, and can be used to give physicians, insurers, and patients a better understanding of expected total charges following surgery.

LOH is the strongest predictor of total charges following transsphenoidal pituitary surgery. Hospital location, admission type, patient age and race, postoperative CSF leak, and fluid or electrolyte abnormalities also predict higher charges, potentially by influencing LOH. Interventions aimed at minimizing the effects of these variables may help decrease charges for this patient population.