Feature Ablation Analysis and Alternate Representations Using Machine Learning in Surgical Risk Assessment

This study applies machine learning methods to create models capable of predicting surgical risk. Gradient boosted trees are applied to 20 preoperative variables from the ACS-NQIP surgical dataset to predict 8 postoperative outcomes. This thesis implements variables utilized by the ACS Universal Surgical Risk Calculator to assess the application of machine learning techniques to predict the same problem set. Additionally, this study assesses one-hot encoding as a method to implement multiple surgical risk procedures, known as CPT codes, into risk assessment and assess performance. This study found that gradient boosted trees perform competitively to other surgical risk tools. Moreover, the inclusion of CPT codes as a preoperative variable suggests that it will improve risk assessment.