Predicting surgical case durations using ill-conditioned CPT code matrix

Efficient utilization of existing resources is crucial for cost containment in medical institutions. Accurately predicting surgery duration will improve the utilization of indispensable surgical resources such as surgeons, nurses, and operating rooms. Prior research has identified the Current Procedural Terminology (CPT) codes as the most important factor when predicting surgical case durations. However, there have been few attempts to create a general predictive methodology that can effectively extract information from multiple CPT codes. This research proposes two regression-based predictive models: (a) linear regression, and (b) log-linear regression models. To perform these regression analysis, a full-ranked design matrix based on CPT code inclusions in the surgical cases needs to be constructed. However, a naively constructed design matrix is ill conditioned (i.e., singular). A systematic procedure is proposed to construct a full-ranked design matrix by sifting out the CPT codes without any predictive power while retaining useful information as much as possible. The proposed models can be applied in general situations where a surgery can have any number of CPT codes and any combination of CPT codes. Using real-world surgical data, the proposed models are compared with benchmark methods and significant reductions in prediction errors are shown.