Neurological Surgeon Montefiore Health System Bronx, NY, US
Introduction: Adolescent Idiopathic Scoliosis (AIS) is a complex spinal deformity that affects a significant number of adolescents worldwide. Thoracolumbosacral orthosis (TLSO) is a common conservative treatment for AIS, but its effectiveness varies among patients. Predicting treatment outcomes remains challenging, limiting the ability of clinicians to provide personalized care. This study aimed to leverage automated machine learning to develop a risk prediction model for TLSO treatment failure in AIS patients.
Methods: A comprehensive dataset from a multicenter prospective cohort study on AIS patients managed with TLSO was utilized. Treatment failure was defined as either progression of Cobb angle to > 50° or surgery recommended prior to TLSO discontinuation. The current state-of-the-art for aML (MLjar v.1.1) was imported in the Google Colab environment using Python programming language. Nine algorithms, including linear logistic regression (LLR) and Decision Trees (DT), were adopted with hyperparameter tuning. The macro-weighted average Area Under the Receiver Operating Curve (mWA-AUROC) and other evaluation metrics were used to evaluate the discriminating ability of the models.
Results: The dataset consisted of 145 patients (Mean age: 13 years; Female: 95%), with 66% at Risser stage 0 and mean baseline Cobb angle of 33°. Treatment failure occurred in 30% of patients. An LLR algorithmic model predicted treatment failure with an mWA-AUROC of 1, log-loss of 0.24, and 100% positive predictive value (PPV) and negative predictive value (NPV) (95% CI: 39.76-100 and 71.51-100, respectively). Baseline triradiate cartilage (TRC) status, Cobb angle, and orthosis wear time were the most influential predictors. DT models (mWA-AUROC=0.8-0.94) identified favorable prognostic factors: baseline TRC ≤0.5, Cobb angle ≤36.835°, and age ≤12.8 years. A simplified LLR model using the aforementioned three predictors outperformed the recently published conventional statistical model (mWA-AUROCs=1 vs. 0.81; PPVs=100% vs. 68%; NPVs=100% vs. 81%).
Conclusion : This study demonstrated the superior predictive capabilities of aML in forecasting TLSO treatment failure for AIS patients. The developed models outperformed conventional statistical model and pinpointed key prognostic factors that would provide valuable insights for clinical decision-making. The external validation and potential integration of these risk prediction models into clinical practice shall pave the way for more personalized and effective AIS management strategies.