Document Type

Article

Publication Date

5-7-2026

Comments

This article is the author's final published version in Acta Neurochirurgica, Volume 168, Issue 1, May 2026, Article number 148.

The published version is available at https://doi.org/10.1007/s00701-026-06876-6. Copyright © The Author(s) 2026.

Abstract

BACKGROUND: Proximal junctional kyphosis (PJK) and failure (PJF) remain challenging and incompletely predictable adverse event following adult spinal deformity (ASD) surgery. Machine learning (ML) models have been proposed to improve risk stratification, but performance varies and clinical applicability remains uncertain. A systematic review and diagnostic test accuracy meta-analysis was performed to synthesize the predictive performance of ML-based models for proximal junctional pathology after ASD surgery.

METHODS: PubMed, Embase, and CENTRAL were searched from inception to February 2026. Studies developing or validating ML models predicting PJK or PJF following ASD surgery were included. Diagnostic performance was synthesized using random-effects models, pooling sensitivity, specificity, diagnostic odds ratio (dOR), and summary receiver operating characteristic (SROC) curves. A prespecified secondary analysis pooled one highest-performing model per study to reduce within-study dependence. Risk of bias was assessed using PROBAST and certainty of evidence using GRADE.

RESULTS: Sixteen studies comprising 3625 patients and 30 ML models met inclusion criteria. Twenty-two models provided extractable test-set data for diagnostic synthesis. Pooled sensitivity was 0.51 (95% CI: 0.41-0.61) and specificity 0.84 (95% CI: 0.78-0.89), with a pooled AUC of 0.67 (95% CI: 0.62-0.72). In the highest-performing model analysis, pooled sensitivity improved to 0.63 (95% CI: 0.55-0.71) and AUC to 0.77 95% CI: 0.71- 0.83). Specificity remained high (0.86). External validation was rare, and calibration was inconsistently reported. Frequently prioritized predictors included sagittal alignment parameters, construct characteristics, bone quality metrics, and age.

CONCLUSIONS: ML-based models demonstrate biologically coherent but clinically moderate discrimination for predicting proximal junctional pathology following ASD surgery. Current performance supports risk stratification rather than categorical decision-making. Progress toward clinically actionable prediction will require standardized endpoint definitions, robust external validation, and calibration-centered reporting.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Additional Files
ESM 1.docx (823 kB)

PubMed ID

42098465

Language

English

Download

Share

COinS