Oral & Maxillofacial Surgery, Faculty of Dentistry, The University of Hong Kong, Hong Kong, Hong Kong.
J Oral Pathol Med. 2020 Nov;49(10):977-985. doi: 10.1111/jop.13089. Epub 2020 Aug 20.
The natural history of oral squamous cell carcinoma (OSCC) is complicated by progressive disease including loco-regional tumour recurrence and development of distant metastases. Accurate prediction of tumour behaviour is crucial in delivering individualized treatment plans and developing optimal patient follow-up and surveillance strategies. Machine learning algorithms may be employed in oncology research to improve clinical outcome prediction.
Retrospective review of 467 OSCC patients treated over a 19-year period facilitated construction of a detailed clinicopathological database. 34 prognostic features from the database were used to populate 4 machine learning algorithms, linear regression (LR), decision tree (DT), support vector machine (SVM) and k-nearest neighbours (KNN) models, to attempt progressive disease outcome prediction. Principal component analysis (PCA) and bivariate analysis were used to reduce data dimensionality and highlight correlated variables. Models were validated for accuracy, sensitivity and specificity, with predictive ability assessed by receiver operating characteristic (ROC) and area under the curve (AUC) calculation.
Out of 408 fully characterized OSCC patients, 151 (37%) had died and 131 (32%) exhibited progressive disease at the time of data retrieval. The DT model with 34 prognostic features was most successful in identifying "true positive" progressive disease, achieving 70.59% accuracy (AUC 0.67), 41.98% sensitivity and a high specificity of 84.12%.
Machine learning models assist clinicians in accessing digitized health information and appear promising in predicting progressive disease outcomes. The future will see increasing emphasis on the use of artificial intelligence to enhance understanding of aggressive tumour behaviour, recurrence and disease progression.
口腔鳞状细胞癌（OSCC）的自然病程较为复杂，包括局部肿瘤复发和远处转移等进行性疾病。准确预测肿瘤的行为对于提供个体化的治疗计划以及制定最佳的患者随访和监测策略至关重要。机器学习算法可用于肿瘤学研究，以改善临床结果预测。
回顾性分析了 19 年间治疗的 467 例 OSCC 患者，构建了详细的临床病理数据库。使用数据库中的 34 个预后特征来填充 4 种机器学习算法，包括线性回归（LR）、决策树（DT）、支持向量机（SVM）和 K 最近邻（KNN）模型，以尝试预测进行性疾病的结果。使用主成分分析（PCA）和双变量分析来降低数据维度并突出相关变量。通过计算接收者操作特征（ROC）和曲线下面积（AUC）来评估模型的准确性、敏感性和特异性，以评估预测能力。
在 408 例特征完全明确的 OSCC 患者中，有 151 例（37%）死亡，有 131 例（32%）在数据检索时出现了进行性疾病。具有 34 个预后特征的 DT 模型在识别“真正的阳性”进行性疾病方面最为成功，其准确性为 70.59%（AUC 为 0.67）、敏感性为 41.98%，特异性为 84.12%。
机器学习模型可帮助临床医生访问数字化健康信息，并且在预测进行性疾病结局方面显示出前景。未来将越来越重视使用人工智能来增强对侵袭性肿瘤行为、复发和疾病进展的理解。
