Comparative effectiveness of arthroscopic-assisted uniportal surgery versus percutaneous endoscopic decompression in...

Single-level lumbar spinal stenosis (LSS) that does not respond to conservative therapy is now of the standard care given due to minimal invasive decompression, but comparative 12-month outcome data of arthroscope-aided single-portal surgery of the lumbar spine (AUSS) and percutaneous endoscopic lumbar decompression (PELD) are scarce. There are also no individual outcome prediction models of these technique specific populations. To compare perioperative and 12-month patient-reported outcomes between AUSS and PELD and to develop and evaluate exploratory machine-learning (ML) models for predicting 12-month favorable outcome, disability (Oswestry Disability Index [ODI]), and back pain (visual analog scale [VAS]) after surgery. This retrospective comparative cohort study analyzed a de-identified patient-level dataset of 865 adults with single-level LSS and complete 12-month follow-up (AUSS n = 445; PELD n = 420). The dataset was provided as a de-identified patient-level file generated directly from source clinical records and was not reconstructed from published aggregate summary statistics; key assumptions include single-center retrospective treatment allocation, imaging-confirmed single-level LSS eligibility, and complete 12-month follow-up as an inclusion criterion. Clinical, perioperative, and complication variables were compared using appropriate nonparametric and parametric tests. Baseline group imbalance was quantified using standardized mean differences (SMDs). Classification and regression ML models were developed using stratified five-fold cross-validation on an 80% development set and evaluated on an independent 20% holdout set. Class imbalance was addressed with inverse-frequency class weighting. Given the exploratory nature of the analysis and dataset provenance, all ML results are interpreted as internal exploratory findings requiring external prospective validation. Mean age was 65.0 years; 89.6% of patients (775/865) achieved a favorable 12-month modified MacNab outcome. AUSS was associated with shorter total operating time (45.47 ± 3.19 vs 54.39 ± 5.24 min; P < 0.001; SMD - 2.07), shorter intracanal decompression time (21.40 ± 2.31 vs 35.49 ± 3.55 min; P < 0.001; SMD - 4.73), and markedly lower fluoroscopy exposure (7.57 ± 2.35 vs 38.41 ± 7.59 s; P < 0.001; SMD - 5.55). PELD showed a less access-traumatic profile with smaller incisions (7.87 ± 1.14 vs 19.71 ± 2.03 mm; P < 0.001; SMD 7.13), lower blood loss (9.41 ± 1.38 vs 17.57 ± 6.35 mL; P < 0.001; SMD 1.75), and lower cost (17 496 ± 603 vs 21 956 ± 581 CNY; P < 0.001). Baseline age imbalance was substantial (SMD = - 0.64), and all group comparisons should be interpreted in this context. Favorable 12-month outcome rates were 93.7% for AUSS and 85.2% for PELD (P < 0.001). Among ML classifiers, random forest showed the most balanced holdout performance: ROC-AUC 0.596, PR-AUC 0.909, sensitivity 0.922, specificity 0.222, and Brier score 0.143. Calibration was suboptimal (intercept 1.72; slope 0.53). Regression performance was limited: holdout R2 = - 0.000 for 12-month ODI and R2 = 0.053 for 12-month back-pain VAS. High rates of 12-month favourable outcome in both procedures had been attained. Operative efficiency and lesser radiations were enhanced by AUSS; PELD was linked with tissue disturbance minimization, a decrease in the amount of blood loss, and low cost. There was no significant difference in overall complication rates. Exploratory ML models demonstrated little predictive power especially in outcomes related to disability, and in unfavorable minority group, and can only be prospectively validated before any personalised clinical usage. Such findings are hypothesis-generating and they should not be relied on to make clinical judgments alone.