Medical Student Perelman School of Medicine at the University of Pennsylvania Philadelphia, Pennsylvania, United States
Introduction: Surgical treatment of degenerative spondylosis and spondylolisthesis can be effective in properly selected patients. However, even after a successful index operation, a certain subset of patients can suffer setbacks, which may occur after as little as a week or many years following surgery. Previous literature has successfully analyzed mobility data from patient smartphones to characterize these functional outcomes after surgery. In this study, we leverage advanced machine learning (ML) models and mobility data to predict the direction and magnitude of post-operative functional activity.
Methods: Patients were retroactively consented and enrolled. Activity data (steps-per-day) recorded in the Apple Health (Apple Inc., Cupertino, CA) mobile application was used to construct a time series across a 2-year peri-operative window. Inputs into ML algorithms included immediate post-operative patient activity over time, age, and BMI. Three distinct ML models – logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost) were each trained on 80% of the dataset and validated using the remaining data. The occurrence of a post-operative secondary decline, defined as a decrease in physical activity below baseline following post-operative recovery, was the primary endpoint being predicted.
Results: A total of 75 patients were included. Following training, the RF and XGBoost models achieved accuracy values of 86.7% (sensitivity 80%, specificity 90%) and 80% (sensitivity 60%, specificity 90%) respectively for predicting post-operative secondary decline. The LR model was the least effective, with an accuracy of 73.3% (sensitivity 50%, specificity 88.8%). Receiving operator characteristic curves showed an area under the curve, a measure of a classifier model's effectiveness of 0.80 for RF, 0.7 for XGBoost, and 0.693 for LR.
Conclusion : The RF model predicted the direction and magnitude of post-operative functional activity more accurately than the XGBoost and LR models. ML models trained on activity data harvested from spine patients’ smartphones show promise as effective clinical tools to predict post-operative recovery course, though further validation on larger datasets is needed. These ML models may be leveraged to prognosticate patients with various spine pathologies and proposed surgical interventions. Further training of ML models on larger, multi-institutional datasets or national registries is needed to build more generalizable models.
