Reducing Bias in the Evaluation of Robotic Surgery for Lung Cancer Through Machine Learning

Abstract

BACKGROUND: Robot-assisted surgery (RAS) is a major innovation in the treatment of lung cancer, offering advantages in surgical precision and reducing postoperative complications. However, its impact on 90-day mortality remains controversial due to methodological biases in comparative studies. This study uses machine learning methods to improve propensity score estimation and reduce selection bias. METHODS: We used the French national hospital database (PMSI) to identify patients who underwent lung resection for cancer between 2019 and 2023. Four models were applied for propensity score estimation: logistic regression, Random Forest, Gradient Boosting Machine (GBM), and XGBoost. Group balancing was achieved through propensity score weighting and matching, followed by logistic regression analysis to estimate the effect of RAS on 90-day mortality. RESULTS: Among the 30,988 patients included, 5717 (18.5%) underwent robot-assisted surgery, while 25,271 (81.5%) underwent thoracotomy. RAS patients had a lower prevalence of comorbidities and earlier-stage tumors. XGBoost was the most effective model for propensity score estimation, with an AUC ROC of 0.9984 and a Brier Score of 0.0119. The adjusted analysis showed a significant reduction in 90-day mortality in the RAS group (OR = 0.39, 95% CI: 0.34-0.45) with weighting and (OR = 0.58, 95% CI: 0.48-0.70) with matching. CONCLUSIONS: The application of machine learning to adjust for selection bias allowed for better control of confounding factors in the analysis of the effect of RAS on 90-day mortality. Our results suggest a potential benefit of robotic surgery compared to thoracotomy, although further studies are needed to confirm these findings.

Ähnliche Arbeiten