| Description |
- Background: Chronic lower back pain (LBP) is a leading cause of disability and economic burden worldwide. Despite numerous treatment modalities, patient outcomes remain poor. Identifying the correct treatment for a given patient remains presenting LBP remains an unsolved clinical problem.
Objective: This study aimed to develop and validate a machine learning (ML) tool to predict patient outcomes following both surgical and non-surgical interventions for chronic LBP.
Methods: Using data from 137,915 patients within the Duke University Health System, we built 6 ML models to predict patient outcomes using patient demographics, past medical history, medication and treatment details as input features. Models were developed for patients who underwent surgery within 90 days of presenting with LBP and those who did not. Key outcomes included active opioid use at 3-, 6- and 9-months following intervention.
Results: The models showed moderate predictive power, with the best performing model exhibiting an AUROC of 0.69 for predicting opioid use at 180 days in the non-surgical cohort. This initial iteration of a clinical decision support tool demonstrates potential in aiding the direction of patients to the most appropriate treatment pathways.
Conclusion: Our ML tool offers a novel approach to guiding the management of chronic LBP by predicting patient outcomes across different index interventions. Future work must focus on refining the models to optimize performance, predicting additional outcome labels, including ED utilization and reduction in pain, and integrating the tool into clinical workflows.
|
|---|
