SBILab, Department of ECE, Indraprastha Institute of Information Technology, New Delhi, India.
Vehant Technology Pvt. Ltd., Noida, India.
Comput Biol Med. 2022 Oct;149:106048. doi: 10.1016/j.compbiomed.2022.106048. Epub 2022 Aug 24.
In this study, we present an efficient Graph Convolutional Network based Risk Stratification system (GCRS) for cancer risk-stage prediction of newly diagnosed multiple myeloma (NDMM) patients. GCRS is a hybrid graph convolutional network consisting of a fusion of multiple connectivity graphs that are used to learn the latent representation of topological structures among patients. This proposed risk stratification system integrates these connectivity graphs prepared from the clinical and laboratory characteristics of NDMM cancer patients for partitioning them into three cancer risk groups: low, intermediate, and high. Extensive experiments demonstrate that GCRS outperforms the existing state-of-the-art methods in terms of C-index and hazard ratio on two publicly available datasets of NDMM patients. We have statistically validated our results using the Cox Proportional-Hazards model, Kaplan-Meier analysis, and log-rank test on progression-free survival (PFS) and overall survival (OS). We have also evaluated the contribution of various clinical parameters as utilized by the GCRS risk stratification system using the SHapley Additive exPlanations (SHAP) analysis, an interpretability algorithm for validating AI methods. Our study reveals the utility of the deep learning approach in building a robust system for cancer risk stage prediction.
在这项研究中，我们提出了一种基于图卷积网络的高效风险分层系统（GCRS），用于预测新诊断的多发性骨髓瘤（NDMM）患者的癌症风险分期。GCRS 是一种混合图卷积网络，由多个连接图的融合组成，用于学习患者之间拓扑结构的潜在表示。该风险分层系统集成了从 NDMM 癌症患者的临床和实验室特征中准备的这些连接图，将它们分为三个癌症风险组：低、中和高。广泛的实验表明，GCRS 在两个公开的 NDMM 患者数据集上的 C 指数和危险比方面优于现有的最先进方法。我们使用 Cox 比例风险模型、无进展生存（PFS）和总生存（OS）的 Kaplan-Meier 分析和对数秩检验对我们的结果进行了统计学验证。我们还使用 SHapley Additive exPlanations（SHAP）分析（一种用于验证 AI 方法的可解释性算法）评估了 GCRS 风险分层系统中使用的各种临床参数的贡献。我们的研究揭示了深度学习方法在构建癌症风险阶段预测的稳健系统中的应用。
