Predictive Modeling for Complications in Patients Undergoing Spine Surgery Using Mobile Health Technology

Introduction: Approximately 16% of patients who undergo spine surgery experience a complication, including surgical-site infection, wound dehiscence, thrombus formation, or hardware failure, resulting in re-operation, pain, and prolonged hospital stays. Despite the use of systems like frailty scores, there remains substantial difficulty in identifying patients at greatest risk. Obtaining more granular preoperative assessments using mobile health technology may support more individualized evaluations that improve predictions of postoperative complications.

Methods: This data was collected through a single-center, prospective cohort study of adult patients undergoing spine surgery for degenerative disease at our tertiary academic medical center between February 2021 and June 2023. Participants completed questionnaires and were provided a Fitbit to wear before surgery. Complications data for 90 days post-operatively was extracted from the electronic medical record. Predictive modeling was performed using 3 sets of predictors: clinical (frailty and comorbidity scores), Fitbit, and combined. Models were validated using leave-one-out cross-validation (LOOCV) and 10-Fold Cross-Validation to assess the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC).

Results: Of 247 patients who underwent spine surgery, 32 (12.9%) experienced a post-operative complication, including readmission (4.0%), unplanned reoperation (3.6%), SSI (1.6%), DVT (1.6%), intubation (1.2%), hematoma (1.2%), hardware failure (0.8%), neurologic injury (0.8%), PE (0.4%), sepsis (0.4%), and wound dehiscence (0.4%). When modeling this dataset, the best-performing model was logistic regression using the combined metrics (AUROC: 0.7176, AUPRC: 0.2709), compared to clinical (AUROC: 0.6610, AUPRC 0.2333) and fitbit (AUROC: 0.6932, AUPRC 0.2561) alone. These findings are consistent with existing studies using clinical parameters to predict post-operative outcomes, with similar AUC ranges for validated models using clinical predictors.

Conclusion : This study highlights the added value of preoperative Fitbit data when combined with clinical predictors, which are extensively studied in outcome prediction, in order to ultimately improve outcomes in this patient population.