Development of a multimodal deep learning-based risk prediction model integrating clinical and radiomic features for short-term acute kidney injury following partial nephrectomy
10.3760/cma.j.cn112330-20250317-00096
- VernacularTitle:基于深度学习网络整合临床及影像组学特征的肾部分切除术后短期AKI风险多模态预测模型的建立
- Author:
Jiangting CHENG
1
;
Jiayi XU
1
;
Chenyang SHEN
1
;
Guanwen YANG
1
;
Yaohui LI
1
;
Li LIU
1
;
Jiajun WANG
1
;
Xiaoyi HU
1
;
Jianming GUO
1
;
Hang WANG
1
Author Information
1. 复旦大学附属中山医院泌尿外科，上海 200030
- Publication Type:Journal Article
- Keywords:
Acute kidney injury;
Partial nephrectomy;
Deep learning;
Risk prediction model
- From:
Chinese Journal of Urology
2025;46(5):349-355
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop and validate a deep learning-based multimodal model integrating clinical and radiomic features for predicting acute kidney injury（AKI）risk after partial nephrectomy.Methods:A retrospective analysis was conducted on 416 patients who underwent partial nephrectomy at Zhongshan Hospital，Fudan University from January 2023 to January 2025. The cohort included 100 AKI patients［defined by a ≥ 25% reduction in postoperative evaluated glomerular filtration rate（eGFR）within 48 hours sustained for >24 hours］and 316 non-AKI patients（1∶3 ratio，randomly matched with 16 additional cases for redundancy）. Clinical and radiomic features were extracted from preoperative contrast-enhanced CT scans using PyRadiomics. Demographics included 259 males and 158 females，with a median age of 57（49，65）years，body mass index of（24.1 ± 3.3）kg/m2，preoperative eGFR of（88.5 ± 18.3）ml/（min·1.73 m2），postoperative eGFR（48-hour）of（76.0 ± 21.9）ml/（min·1.73 m2），Zhongshan Score（ZSscore）of 7.34 ± 2.01，and R.E.N.A.L. score of 7.50 ± 1.71. All tumors were T 1a stage. Patients were divided into training（n = 312）and test（n = 104）sets（3∶1 ratio）. A clinical model was constructed via multivariate logistic regression，while radiomic and combined（clinical + radiomic）models utilized an artificial neural network（ANN）with 1 input layer，5 hidden layers，1 output layer，and 10 5 training epochs. Model performance was evaluated by using receiver operating characteristic（ROC）curves and area under the curve（AUC），and was compared to the Martini model. Feature contributions were interpreted via SHapley Additive exPlanations（SHAP）. Results:In the test set，the results of multivariate logistic regression showed that patient’s weight，preoperative eGFR，R.E.N.A.L. score，surgical approach，and operation time were risk factors for AKI（ P < 0.05）. The AUC of the clinical feature prediction model constructed based on the above factors was 0.852（95% CI 0.775?0.929）. In the test set，the AUC of the Martini model was 0.725（95% CI 0.565?0.791）. The radiomic model，trained on 1 315 imaging features，achieved an AUC of 0.898（95% CI 0.804?0.993）with 94.2%（98/104）accuracy. The combined clinical and radiomic model，integrating 1 315 radiomic features and clinical features，demonstrated superior performance with an AUC of 0.946（95% CI 0.887?1.000）and 96.2%（100/104）accuracy，outperforming both the clinical model（ P = 0.03）and the Martini model（ P < 0.01）. SHAP analysis identified the top five predictors in the combined model：ZSscore（SHAP value：0.78），long-run low gray-level emphasis（SHAP value：0.61），run-length non-uniformity（SHAP value：0.58），size-zone non-uniformity（SHAP value：0.46），and gray-level co-occurrence matrix joint energy（SHAP value：0.36）. Conclusions:The deep learning-based multimodal model integrating clinical and radiomic features accurately predicts AKI risk after partial nephrectomy，offering a novel strategy for preoperative risk stratification and personalized intervention.
