Data-Driven Clustering for Risk Stratification of Unfavorable Outcomes After Lumbar Fusion Surgery: A Development Study.

OBJECTIVE

Lumbar fusion surgery serves as a crucial option for treating lumbar degenerative diseases. However, patient heterogeneity contributes to suboptimal surgical outcomes in a substantial proportion of cases. Therefore, an accurate classification may provide a powerful tool for personalized treatment and enable the identification of individuals at increased risk for unfavorable surgical outcomes (USO). The study aimed to develop a risk stratification model for USO using cluster analysis.

METHODS

Consecutive patients diagnosed with degenerative lumbar disease who underwent lumbar fusion between April 2019 and January 2023 were enrolled. The outcome of interest was the USO, defined as failure to achieve a minimal clinically important difference in the 36-Item Short Form Health Survey physical component summary score, with the presence of complications. Three machine learning algorithms were employed to identify risk factors associated with USO. Based on these risk factors, we conducted a data-driven clustering analysis to develop a risk stratification model. Furthermore, based on 6 machine learning models, we developed a classification classifier capable of accurately identifying the risk cluster of individual patients.

RESULTS

A total of 662 patients were enrolled for risk stratification model, 219 patients were classified as having an USO. Six features were identified as key prognostic predictors, including frailty, depression, PI-LL (pelvic incidence minus lumbar lordosis) match, surgical levels, functional independence measure, and the relative functional cross-sectional area. The K-prototypes clustering algorithm successfully identified 3 distinct clusters. Furthermore, we developed a classification classifier, in which LightGBM (light gradient boosting machine) demonstrated the highest predictive performance (area under the receiver operating characteristic curve, 0.951; 95% confidence interval [CI], 0.814-0.974; area under the precision-recall curve, 0.927; 95% CI, 0.769-0.969).

CONCLUSION

Based on data-driven clustering analysis, we developed a risk stratification model for predicting USO following lumbar fusion surgery, which demonstrated high predictive accuracy. Further studies in larger and more diverse cohorts are warranted to validate the clinical applicability of clustering analysis in USO risk stratification.