Scholarlink

Background <p>We aimed to create a Machine learning (ML) model using patient demographic, clinical and pathological data for prediction of overall survival in patients treated with radical cystectomy (RC). Secondly, we evaluated whether inflammatory markers contributed to this model.</p> Methods <p>We conducted a retrospective analysis of the institutional cystectomy database and identified consecutive RC patients. Dataset-1 (DS-1) was analyzed in ML models using 30 original features (including the target feature) encompassing preoperative, intraoperative, and postoperative data of the patients. All derived inflammatory markers were cumulatively added to DS-1 to create DS-2, and to test the specific contribution of inflammatory markers, they were systematically integrated in an ordinary order based on their predictive ability (DS-3). Markers without predictive contribution were excluded from the DS-3 model. In addition, the Shapley Additive Explanations (SHAP) method was used to examine the importance of each clinical feature and inflammatory marker.</p> Results <p>The median age of the 241 patients was 65 years. The mortality rate was 60.2% (145/241). Two- and five-year overall survival (OS) rates were 54.7% and 37.2%, respectively. According to DS-1, F<sub>1</sub> scores were between 0.72 and 0.78. Random Forests and XGBoost models achieved the highest score of 0.78. DS-2 including all inflammatory markers; however, no significant improvement in F<sub>1</sub> scores was observed (0.73–0.78). In DS-3, firstly adding the systemic inflammatory response index (SIRI) to the original features and then neutrophil/lymphocyte ratio (NLR) and platelet/lymphocyte ratio (PLR), achieved the highest F<sub>1</sub> score (0.80) in the Random Forest model. SHAP analyses showed that Tumor (T) stage, preoperative albumin and presence of lympho-vascular invasion (LVI) contributed most to model predictivity.</p> Conclusions <p>ML models derived using demographic/clinical features resulted in a maximum F<sub>1</sub> score of 0.78. However, adding the most predictive inflammatory markers in the sequence SIRI, NLR and PLR to demographic data achieved the highest F<sub>1</sub> score of 0.80. Furthermore, T stage and preoperative albumin were the strongest predictive factors in the ML models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluation of inflammatory markers in survival analysis of patients undergoing radical cystectomy using machine learning

  • 期刊论文

  • 发表日期：

    2025年10月21日

查看全文
  • Naci Burak Çınar,
  • Hasan Yılmaz,
  • Efe Yılmaz Taşyürek,
  • Meltem Kurt Pehlivanoğlu,
  • Sevinç İlhan Omurca,
  • Muhlis Ünal,
  • Kerem Teke

摘要

Background

We aimed to create a Machine learning (ML) model using patient demographic, clinical and pathological data for prediction of overall survival in patients treated with radical cystectomy (RC). Secondly, we evaluated whether inflammatory markers contributed to this model.

Methods

We conducted a retrospective analysis of the institutional cystectomy database and identified consecutive RC patients. Dataset-1 (DS-1) was analyzed in ML models using 30 original features (including the target feature) encompassing preoperative, intraoperative, and postoperative data of the patients. All derived inflammatory markers were cumulatively added to DS-1 to create DS-2, and to test the specific contribution of inflammatory markers, they were systematically integrated in an ordinary order based on their predictive ability (DS-3). Markers without predictive contribution were excluded from the DS-3 model. In addition, the Shapley Additive Explanations (SHAP) method was used to examine the importance of each clinical feature and inflammatory marker.

Results

The median age of the 241 patients was 65 years. The mortality rate was 60.2% (145/241). Two- and five-year overall survival (OS) rates were 54.7% and 37.2%, respectively. According to DS-1, F1 scores were between 0.72 and 0.78. Random Forests and XGBoost models achieved the highest score of 0.78. DS-2 including all inflammatory markers; however, no significant improvement in F1 scores was observed (0.73–0.78). In DS-3, firstly adding the systemic inflammatory response index (SIRI) to the original features and then neutrophil/lymphocyte ratio (NLR) and platelet/lymphocyte ratio (PLR), achieved the highest F1 score (0.80) in the Random Forest model. SHAP analyses showed that Tumor (T) stage, preoperative albumin and presence of lympho-vascular invasion (LVI) contributed most to model predictivity.

Conclusions

ML models derived using demographic/clinical features resulted in a maximum F1 score of 0.78. However, adding the most predictive inflammatory markers in the sequence SIRI, NLR and PLR to demographic data achieved the highest F1 score of 0.80. Furthermore, T stage and preoperative albumin were the strongest predictive factors in the ML models.