Scholarlink

Purpose <p>Predicting the likelihood of benign neoplasia in patients with suspected renal cell carcinoma (RCC) is a cornerstone of presurgical planning. We sought to create and validate U.N.I.K., a machine learning (ML) model capable of predicting benign lesions on final histological report.</p> Methods <p>We queried the INMARC database for patients with cT1ab–2ab renal neoplasms. The primary outcome was the development of an ML model to predict benign neoplasia. The secondary objective was to compare ML performance to that of a logistic regression (LR) model. The ML algorithms evaluated included random forest, extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). Model performance was assessed using receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC).</p> Results <p>A total of 3,093 patients were analyzed, with benign histology reported in 7.1% of cases. LR identified the following as significant predictors: female sex (OR 2.14, <i>p</i> &lt; 0.001), diabetes (OR 5.95, <i>p</i> &lt; 0.001), tumor size (OR 0.86, <i>p</i> &lt; 0.001), preoperative CRP &lt; 1&#xa0;mg/L (OR 1.69, <i>p</i> &lt; 0.001), serum calcium (OR 1.81, <i>p</i> &lt; 0.001), Charlson Comorbidity Index (OR 0.52, <i>p</i> &lt; 0.001), cystic lesion (OR 12.31, <i>p</i> &lt; 0.001), De Ritis ratio ≥ 0.9 (OR 1.49, <i>p</i> = 0.03), preoperative proteinuria (OR 0.12, <i>p</i> &lt; 0.001), and Karnofsky Performance Status (OR 0.90, <i>p</i> &lt; 0.001), with an AUC of 0.89. XGBoost showed the best performance (AUC 0.94) and was used to develop the U.N.I.K. model. Limitations include the retrospective design.</p> Conclusion <p>We developed a point-of-care model capable of predicting benign tumors with high accuracy. U.N.I.K. may refine clinical decision-making and reduce the burden of surgical overtreatment.</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

UNIK (Urologic Non-Neoplastic Investigation of Kidneys): a machine learning approach to decode benign lesion

  • 期刊论文

  • 发表日期：

    2025年6月4日

查看全文
  • Cesare Saitta,
  • Giuseppe Garofano,
  • Giacomo Musso,
  • Hajime Tanaka,
  • Dattatraya Patil,
  • Vi Nguyen,
  • Kit L. Yuen,
  • Benjamin Baker,
  • Mai Dabbas,
  • Margaret F. Meagher,
  • Dhruv Puri,
  • Natalie Birouty,
  • Omer Baker,
  • Masaki Kobayashi,
  • Shohei Fukuda,
  • Marco Paciotti,
  • Giovanni Lughezzani,
  • Alessandro Larcher,
  • Umberto Capitanio,
  • Riccardo Melloni,
  • Mario Catanzaro,
  • Sebastiano Nazzani,
  • Nicola Nicolai,
  • Yasuhisa Fujii,
  • Viraj Master,
  • Nicolòmaria Buffi,
  • Ithaar H. Derweesh

摘要

Purpose

Predicting the likelihood of benign neoplasia in patients with suspected renal cell carcinoma (RCC) is a cornerstone of presurgical planning. We sought to create and validate U.N.I.K., a machine learning (ML) model capable of predicting benign lesions on final histological report.

Methods

We queried the INMARC database for patients with cT1ab–2ab renal neoplasms. The primary outcome was the development of an ML model to predict benign neoplasia. The secondary objective was to compare ML performance to that of a logistic regression (LR) model. The ML algorithms evaluated included random forest, extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). Model performance was assessed using receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC).

Results

A total of 3,093 patients were analyzed, with benign histology reported in 7.1% of cases. LR identified the following as significant predictors: female sex (OR 2.14, p < 0.001), diabetes (OR 5.95, p < 0.001), tumor size (OR 0.86, p < 0.001), preoperative CRP < 1 mg/L (OR 1.69, p < 0.001), serum calcium (OR 1.81, p < 0.001), Charlson Comorbidity Index (OR 0.52, p < 0.001), cystic lesion (OR 12.31, p < 0.001), De Ritis ratio ≥ 0.9 (OR 1.49, p = 0.03), preoperative proteinuria (OR 0.12, p < 0.001), and Karnofsky Performance Status (OR 0.90, p < 0.001), with an AUC of 0.89. XGBoost showed the best performance (AUC 0.94) and was used to develop the U.N.I.K. model. Limitations include the retrospective design.

Conclusion

We developed a point-of-care model capable of predicting benign tumors with high accuracy. U.N.I.K. may refine clinical decision-making and reduce the burden of surgical overtreatment.