Leveraging small-sample machine learning for rigorous prediction of JOA recovery in cervical spondylotic myelopathy patients: insights from imaging parameters and modeling strategies.

Abstract

Background: This study investigated how machine learning methods can be applied to small sample sizes to enhance prediction of postoperative functional recovery, as measured by the Japanese Orthopedic Association (JOA) score, in cervical spondylotic myelopathy (CSM) patients undergoing laminoplasty, while leveraging existing research and expert knowledge.

Methods: Data from 143 CSM patients who underwent laminoplasty were analyzed. Eleven key imaging parameters related to cervical alignment and paravertebral muscles were measured. Multiple machine learning algorithms were evaluated using different feature engineering approaches. Model performance was assessed through repeated random sampling and confidence intervals.

Results: Increasing the number of random data splits improved stability of performance metrics. Incorporating fat infiltration parameters enhanced predictive performance. The Gaussian Naive Bayes algorithm achieved the best overall performance, with 76.90% accuracy (65.01-88.78% CI) and 75.24% AUC (59.20-91.28% CI) using the optimal feature set. Logistic regression and support vector machines also performed well. Random forests showed high specificity but low sensitivity.

Conclusions: This study demonstrates that machine learning can effectively predict postoperative outcomes in CSM patients using small samples when combined with expert-informed feature engineering and rigorous evaluation methods. Multiple training iterations and confidence interval reporting enhance result reliability. Machine learning's flexibility in feature selection provides advantages over traditional statistical approaches for such predictive tasks in clinical settings.