A deep learning-based multimodal model with automated body composition analysis predicts prognosis in advanced clear cell renal cell carcinoma.
Authors
Affiliations (5)
Affiliations (5)
- Department of Urology, Yantai Yuhuangding Hospital, Qingdao University, No. 20 East Yuhuangding Road, Yantai, 264000, Shandong, China.
- Department of Urology, Jinan Central Hospital, Shandong First Medical University, No. 105 Jiefang Road, Lixia District, Jinan, Shandong, China.
- Department of Gynecology, Yantai Yuhuangding Hospital, Qingdao University, No. 20 East Yuhuangding Road, Yantai, 264000, Shandong, China. [email protected].
- Department of Urology, Yantai Yuhuangding Hospital, Qingdao University, No. 20 East Yuhuangding Road, Yantai, 264000, Shandong, China. [email protected].
- Department of Urology, Yantai Yuhuangding Hospital, Qingdao University, No. 20 East Yuhuangding Road, Yantai, 264000, Shandong, China. [email protected].
Abstract
Clear cell renal cell carcinoma (ccRCC) is an aggressive malignancy with a high risk of postoperative recurrence. Body composition has emerged as a prognostic marker, but its clinical use is hindered by manual measurement methods and poorly interpretable models. This study aimed to develop an automated, interpretable prediction model integrating deep learning-based body composition analysis to assess prognosis and explore underlying mechanisms in advanced ccRCC. This retrospective multicenter study included patients who underwent radical nephrectomy for advanced ccRCC. Body composition was quantified automatically from preoperative CT using a deep learning model (Comp2Comp). These metrics were integrated with clinicopathological features to build machine learning models for predicting 5-year survival and recurrence. Model performance was rigorously evaluated in an independent external cohort. To explore the biological basis of key imaging features, transcriptomic analysis and experimental validation were performed. Metrics for calibration and discrimination were used to compare the models. The multimodal model demonstrated robust predictive performance. The MLP model excelled in 5-year survival prediction (AUC 0.787 internal, 0.812 external), while the SVC model was optimal for recurrence (AUC 0.740 internal, 0.815 external). SHAP analysis identified visceral (VAT) and subcutaneous (SAT) adipose tissue as top predictive features. Subsequent transcriptomic analysis linked VAT to metabolic pathways and SAT to embryonic development programs. Protein expression of key associated genes was confirmed by immunohistochemistry. This study developed an automated multimodal prediction model by integrating deep learning-derived body composition metrics with clinicopathological features, demonstrating promising predictive performance in assessing survival and recurrence risks in advanced ccRCC, which was further supported by a preliminary external validation cohort. SHAP analysis identified visceral and subcutaneous adipose tissue as key predictors. Transcriptomic analysis linked VAT to pathways of mature renal function and SAT to embryonic developmental programs, providing initial insights into potential biological correlates of these imaging biomarkers.