PO.BCS02.06 · 生物信息与计算
基于AI的可解释性生存建模用于晚期非小细胞肺癌
AI based explainable survival modeling for advanced non small cell lung cancer
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景:晚期非小细胞肺癌(NSCLC)中准确的个体化预后判断因临床和生物学变量间的非线性相互作用而受到阻碍。将可解释的人工智能(AI)与经典生存建模相结合,可在保持透明度的同时提高预测精度。
方法:分析了一个包含62 608例晚期NSCLC患者、52782例观察到死亡事件的真实世界队列。使用单变量Cox比例风险回归(p < 0.05)识别与总生存(OS)显著相关的预后变量。通过采用5折交叉验证(λ = 0.00176)和稳定性选择(≥ 0.7)的LASSO正则化Cox建模实现特征精炼,产生了一个20特征预后特征。随后使用这些变量训练并基准测试了18种回归和集成机器学习算法以进行连续OS预测。模型性能采用R²、RMSE、MAE、校准图和决策曲线分析(DCA)进行评估。特征可解释性采用SHAP(Shapley Additive Explanations)评估,以量化每个预测因子效应的方向和大小。
结果:全部20个变量在多变量Cox分析中仍具显著性。化疗、系统治疗和手术为独立保护因素,而年龄、肿瘤大小、转移负荷、肝/骨转移、区域淋巴结受累和N分期预示较差的结局。集成梯度提升模型优于线性基线(R² ≈ 0.15 对比 0.10;RMSE ≈ 18个月)。LightGBM取得了最高准确率(R² = 0.155;MAE = 11.9个月),具有出色的校准和DCA上最大的净获益。SHAP分析将化疗、诊断年份、器官转移数量和年龄识别为预测生存的主导决定因素,在各折间具有很强的可重复性(Spearman ρ > 0.9)。
结论:一个透明的Cox-LASSO-LightGBM-SHAP框架为晚期NSCLC建立了一个稳健、生物学一致的20因素预后特征。该模型实现了高预测准确性、临床校准和可解释性,揭示了治疗模式、转移范围和时间性治疗进展是主要的生存驱动因素。这一可解释的AI框架能够实现可信的个体化生存预测,并在数据驱动建模与临床肿瘤学之间架起桥梁。
查看英文原文 English abstract
Background: Accurate individualized prognostication in advanced non-small cell lung cancer (NSCLC) is hindered by nonlinear interactions among clinical and biological variables. Integrating interpretable artificial intelligence (AI) with classical survival modeling may enhance predictive precision while preserving transparency.
Methods: A real-world cohort of 62 608 advanced NSCLC patients with 52782 observed death events was analyzed. Prognostic variables significantly associated with overall survival (OS) were identified using univariate Cox proportional-hazards regression (p < 0.05). Feature refinement was achieved through LASSO-regularized Cox modeling with 5-fold cross-validation (λ = 0.00176) and stability selection (≥ 0.7), producing a 20-feature prognostic signature. These variables were then used to train and benchmark 18 regression and ensemble machine-learning algorithms for continuous OS prediction. Model performance was evaluated using R², RMSE, MAE, calibration plots, and decision-curve analysis (DCA). Feature interpretability was assessed with SHAP (Shapley Additive Explanations) to quantify the direction and magnitude of each predictor's effect.
Results: All 20 variables remained significant in multivariate Cox analysis. Chemotherapy, systemic therapy, and surgery were independent protective factors, whereas age, tumor size, metastatic burden, liver/bone metastasis, regional nodal involvement, and N stage predicted worse outcomes. Ensemble gradient-boosting models outperformed linear baselines (R² ≈ 0.15 vs. 0.10; RMSE ≈ 18 months). LightGBM achieved the highest accuracy (R² = 0.155; MAE = 11.9 months) with excellent calibration and the greatest net benefit on DCA. SHAP analysis identified chemotherapy, diagnosis year, organ metastatic number, and age as dominant determinants of predicted survival, with strong reproducibility across folds (Spearman ρ > 0.9).
Conclusions: A transparent Cox-LASSO-LightGBM-SHAP framework established a robust, biologically consistent 20-factor prognostic signature for advanced NSCLC. The model achieved high predictive accuracy, clinical calibration, and interpretability, revealing treatment modality, metastatic extent, and temporal therapeutic progress as principal survival drivers. This interpretable AI framework enables credible, individualized survival prediction and bridges data-driven modeling with clinical oncology.
利益披露 Disclosure
K. Qin, None..
A. Qin, None..
J. V. Heymach, None.