Preoperatively Predicting Patient Discharge Disposition After Elective Lumbar Spine Surgery Using a Machine-Learning Classification Model. (16th November 2020)

• Record Type: Journal Article
• Authors: Winkelman, Robert; Habboub, Ghaith; Nault, Rod; Salas-Vega, Sebastian; Chakravarthy, Vikram; Grabowski, Matthew M; Savage, Jason; Pelle, Dominic; Mroz, Thomas E
• Is Part Of: Neurosurgery. Volume 67(2010)Supplement 1
• Publication Date: 2020-11-16
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1093/neuros/nyaa447_152
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract: INTRODUCTION: Prolonged length of stay (LOS) can be a major driver of cost following elective spine surgery. Previously, we have identified patients who required post-acute care services (PACS [i.e. homecare or discharge to a facility]) after lumbar spine surgery at our institution have longer LOS when surgery is scheduled late in the week (i.e. Thursday, Friday) compared to earlier in the week. Therefore, improving the surgical team's understanding of patients' post-acute care needs prior to surgery may provide opportunities to better coordinate care in order to mitigate prolonged LOS. METHODS: A retrospective chart review was performed on 1743 patients who underwent lumbar spine surgery at a single healthcare system from 3/1/2016 through 3/1/2020. Patients' discharge disposition was classified according to the need for PACS (Yes vs. No). Several machine-learning models to predict patients' need for PACS were developed using 51 variables which included patient characteristics (age, gender, Elixhauser comorbidities), health questionnaire data (PROMIS Global Health), and operative characteristics (surgeon, type of surgey, number of levels). RESULTS: The final cohort included 613 patients (35%) who required PACS. The final model selected was a gradient boosted tree model. The model's accuracy, sensitivity, and specificity for detecting the need for PACS in a validation cohort were 78%, 82%, and 77%, respectively. The most important variables as determined by loss from feature dropout were the patient's age, Elixhauser comorbidity score, type of surgery, and PROMIS-GH responses for preoperative physical activity and physical health. CONCLUSION: Our study demonstrates that machine-learning classification models can assist in preoperatively identifying patients who will require PACS after surgery. By understanding patients' post-discharge needs prior to surgery, the surgical team can improve preoperative patient counseling and/or optimize surgical scheduling to potentially avoid unnecessary prolonged LOS.