PO.BCS02.06 · 生物信息与计算
整合基因组、组织病理和实验室数据的多模态建模预测寡转移性结直肠癌肝切除术后的生存
Multi-modal modeling of genomic, histopathologic, and lab data predicts survival after hepatectomy in oligometastatic colorectal cancer
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：寡转移是转移进展的一个中间阶段，其扩散范围和器官受累有限，此时仍有可能进行根治性治疗。在结直肠癌（CRC）中，约10%的患者表现为局限于肝脏的寡转移性疾病，并接受以根治为目的的肝切除术。然而，临床上仍缺乏能够区分长期生存者与肝切除术后短期内复发（因而未从手术中获益）患者的术前预测性生物标志物。为改善寡转移性CRC的临床决策，我们旨在利用常规采集的临床数据开发一种机器学习（ML）模型，以可靠地预测肝切除术后的结局。
方法：我们分析了在Memorial Sloan Kettering Cancer Center接受治疗的284例CRC患者，这些患者均为局限于肝脏的寡转移性疾病，并接受了经部分肝切除术的转移灶切除。共采集了四种数据模态：临床特征；MSK-IMPACT靶向外显子测序；全切片图像的组织病理学特征；以及肝切除术前30天内采集的实验室数值。我们采用嵌套交叉验证训练了一个XGBoost机器学习模型，以预测哪些患者在肝切除术后能够实现超过三年的总生存。
结果：基于基因组学、组织病理学和实验室数据训练的机器学习模型取得了最高的预测性能（AUROC = 0.75，SE = 0.03），仅使用基因组学和实验室数据时性能相当（AUROC = 0.73，SE = 0.04）。将机器学习预测的风险与已建立的用于CRC复发的临床风险评分（CRS）相结合，预后判别能力有小幅提升（CRS C-index = 0.606，CRS + ML C-index = 0.635）。实验室和组织病理学变量对模型预测的贡献最大，且全身炎症指数升高（（血小板 × 中性粒细胞）/ 淋巴细胞）与更差的预后相关，这与既往研究一致。我们进一步发现，MAPK信号通路调控因子——双特异性磷酸酶DUSP4的深度缺失与长期生存改善显著相关（log-rank p = 0.016）。
结论：使用术前基因组、组织病理、实验室和临床数据的多模态机器学习模型能够预测寡转移性CRC肝切除术后的长期生存。本研究提供了一种整合常规采集数据的策略，以改善寡转移性疾病切除的临床决策和风险分层。
查看英文原文 English abstract
Background: Oligometastasis is an intermediate stage of metastatic progression with limited spread and organ involvement, where curative treatment remains possible. In colorectal cancer (CRC), ~10% of patients present with liver-limited oligometastatic disease and undergo hepatectomy with curative intent. However, there is an unmet need for pre-operative predictive biomarkers that can distinguish long-term survivors from patients who relapse shortly after hepatectomy and thus did not benefit from surgery. To improve clinical decision-making for oligometastatic CRC, we aimed to develop a machine learning (ML) model to reliably predict post-hepatectomy outcomes using routinely collected clinical data.
Methods: We analyzed 284 CRC patients treated at Memorial Sloan Kettering Cancer Center who had liver-confined oligometastatic disease and underwent metastatic resection via partial hepatectomy. Four data modalities were collected: clinical features; MSK-IMPACT targeted exon sequencing; histopathology features from whole slide images; and laboratory values collected within 30 days pre-hepatectomy. We trained an XGBoost ML model with nested cross-validation to predict which patients would achieve overall survival greater than three years after hepatectomy.
Results: An ML model trained on genomics, histopathology and lab data achieved the highest predictive performance (AUROC = 0.75, SE = 0.03), with comparable performance using only genomics and lab data (AUROC = 0.73, SE = 0.04). Combining ML-predicted risk with the established Clinical Risk Score (CRS) for CRC recurrence showed a modest increase in prognostic discrimination (CRS C-index = 0.606, CRS + ML C-index = 0.635). Lab and histopathology variables contributed most to model predictions, and an elevated systemic inflammation index ((platelets × neutrophils) / lymphocytes) was associated with worse prognosis, consistent with prior work. We further found that deep deletion in dual-specificity phosphatase DUSP4, a regulator of MAPK signaling, was significantly associated with improved long-term survival (log-rank p = 0.016).
Conclusions: A multi-modal ML model using pre-operative genomic, histopathologic, laboratory, and clinical data can predict long-term survival following hepatectomy in oligometastatic CRC. This study offers a strategy for integrating routinely collected data to improve clinical decision making and risk stratification for resection of oligometastatic disease.
利益披露 Disclosure
D. Koyyalagunta, None..
M. Liu, None..
S. Chhabria, None..
M. Darmofal, None..
M. Waters, None..
A. C. Wei, None..
T. Kingham, None..
W. R. Jarnagin, None..
J. Shia, None..
M. I. D’Angelica, None..
Q. Morris, None.