Metabolic dynamic score and machine learning: a novel approach to predicting pathological complete response in rectal cancer after neoadjuvant chemoradiotherapy
Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine
Background This study developed a scoring system based on the dynamic changes in biochemical indicators in advance of and subsequent to neoadjuvant chemoradiotherapy (NCRT) in individuals with locally advanced rectal cancer (LARC). The scoring system, combined with other clinical features, was used to develop a machine learning (ML) model aimed at predicting a pathological complete response (pCR). Methods A review of earlier data was performed on the data of 1300 patients with LARC treated at Center1 and Center 2. To determine factors linked to pCR and create the scoring system, uni and multivariate logistic regression analyses were conducted. The development of predictive models involved the use of 10 ML methods, while model performance was evaluated using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), decision curve analysis, and calibration curves. Additionally, SHapley Additive exPlanations values were applied to enhance model interpretability. Results Multivariate analysis identified the MDS, tumor size, clinical T stage, and clinical N stage as independent predictors, which were incorporated into the ML models. Among the 10 ML models, the XGBoost model demonstrated the best and most generalizable predictive performance (training set: AUC = 0.93, AUPRC = 0.732; external validation set: AUC = 0.92, AUPRC = 0.779) and was selected as the optimal model. Conclusion This study analyzed changes in biochemical in
Abstract
Background This study developed a scoring system based on the dynamic changes in biochemical indicators in advance of and subsequent to neoadjuvant chemoradiotherapy (NCRT) in individuals with locally advanced rectal cancer (LARC). The scoring system, combined with other clinical features, was used to develop a machine learning (ML) model aimed at predicting a pathological complete response (pCR). Methods A review of earlier data was performed on the data of 1300 patients with LARC treated at Center1 and Center 2. To determine factors linked to pCR and create the scoring system, uni and multivariate logistic regression analyses were conducted. The development of predictive models involved the use of 10 ML methods, while model performance was evaluated using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), decision curve analysis, and calibration curves. Additionally, SHapley Additive exPlanations values were applied to enhance model interpretability. Results Multivariate analysis identified the MDS, tumor size, clinical T stage, and clinical N stage as independent predictors, which were incorporated into the ML models. Among the 10 ML models, the XGBoost model demonstrated the best and most generalizable predictive performance (training set: AUC = 0.93, AUPRC = 0.732; external validation set: AUC = 0.92, AUPRC = 0.779) and was selected as the optimal model. Conclusion This study analyzed changes in biochemical indicators pre- versus post-NCRT and developed a promising MDS. By leveraging ML predictive models to estimate pCR, this approach may serve as a convenient tool to support the clinical decision-making process pending further prospective validation.