A decision tree analysis to predict clinical outcome of minimally invasive lumbar decompression surgery for lumbar spinal stenosis with and without coexisting spondylolisthesis and scoliosis

Abstract

Background context: Implementing machine learning techniques, such as decision trees, known as prediction models that use logical construction diagrams, are rarely used to predict clinical outcomes.

Purpose: To develop a clinical prediction rule to predict clinical outcomes in patients who undergo minimally invasive lumbar decompression surgery for lumbar spinal stenosis with and without coexisting spondylolisthesis and scoliosis using a decision tree model.

Study design/setting: A retrospective analysis of prospectively collected data.

Patient sample: This study included 331 patients who underwent minimally invasive surgery for lumbar spinal stenosis and were followed up for ≥2 years at 1 institution.

Outcome measures: Self-report measures: The Japanese Orthopedic Association (JOA) scores and low back pain (LBP)/leg pain/leg numbness visual analog scale (VAS) scores. Physiologic measures: Standing sagittal spinopelvic alignment, computed tomography, and magnetic resonance imaging results.

Methods: Low achievement in clinical outcomes were defined as the postoperative JOA score at the 2-year follow-up <25 points. Univariate and multiple logistic regression analysis and chi-square automatic interaction detection (CHAID) were used for analysis.

Results: The CHAID model for JOA score <25 points showed spontaneous numbness/pain as the first decision node. For the presence of spontaneous numbness/pain, sagittal vertical axis ≥70 mm was selected as the second decision node. Then lateral wedging, ≥6° and pelvic incidence minus lumbar lordosis (PI-LL) ≥30° followed as the third decision node. For the absence of spontaneous numbness/pain, sex and lateral olisthesis, ≥3mm and American Society of Anesthesiologists physical status classification system score were selected as the second and third decision nodes. The sensitivity, specificity, and the positive predictive value of this CHAID model was 65.1, 69.8, and 64.7% respectively.

Conclusions: The CHAID model incorporating basic information and functional and radiologic factors is useful for predicting surgical outcomes.

Keywords: Chi-square automatic interaction detection; Clinical outcomes; Decision tree analysis; Lumbar spinal stenosis; Machine learning; Minimally invasive; Risk factors; Surgery.