基于深度学习的乳腺癌术后脂质体多柔比星与表柔比星个性化治疗推荐研究

Research on personalized therapy recommendation of liposome doxorubicin and epirubicin for breast cancer patients after surgery based on deep learning

目的比较机器学习(machine learning,ML)与因果领域个体化干预效果(individualized treatment effect,ITE)评估深度学习这两类方法在真实临床数据集上的个性化推荐性能差异,构建乳腺癌术后聚乙二醇脂质体多柔比星(pegylated liposomal doxorubicin,PLD)与表柔比星(epirubicin,EPI)的个体化药物治疗推荐模型,通过评估药物疗效来指导临床治疗方案。方法回顾性收集浙江大学医学院附属邵逸夫医院收治的904名乳腺癌患者临床资料,其中387例采用PLD治疗,517例采用EPI治疗,通过倾向性评分匹配法比较两组患者5年无病生存期(disease free survival,DFS)结局;应用CFR_WASS等6种ITE模型预测患者在两种药物治疗下5年DFS概率,使用随机森林等6种机器学习模型作为基准进行性能分析比较;根据受试者工作特征曲线下的面积(area under the receiver operating characteristic curve,AUROC)评估预测性能,通过计算实际使用治疗与模型推荐治疗一致组和对照组的5年DFS率差异评估治疗推荐有效性。结果153对匹配病例中,PLD组和EPI组5年DFS结局比较差异无统计学意义,16对病例PLD组临床结局优于EPI组,12对病例EPI组临床结局优于PLD组,验证两种药物存在个体治疗收益差异。CFR_WASS模型获得了最优预测性能(AUROC为0.7368);多数ML组与对照组的5年DFS率无明显差异,ITE组5年DFS率均低于对照组,差异有统计学意义(P<0.01),其中CFR_WASS组5年DFS率较对照组低2.13%。结论相比于ML模型,ITE评估深度学习模型能更准确地估计两种药物的个体化治疗效果,给出有效的个体化治疗推荐,具有一定临床应用价值。

Objective To compare the performance of machine learning(ML)and individualized treatment effect(ITE)models based on deep learning in providing personalized treatment recommendations using real-world clinical datasets,and construct personalized drug treatment recommendation models for pegylated liposomal doxorubicin(PLD)and epirubicin(EPI)in postoperative breast cancer patients,and assist clinical decision-making by evaluating the treatment effects of these drugs.Methods Clinical data of 904 breast cancer patients admitted at Sir Run Run Shaw Hospital,Zhejiang University School of Medicine was collected retrospectively,including 387 cases treated with PLD and 517 cases treated with EPI.The two groups were compared using propensity score matching to assess the 5-year disease free survival(DFS)outcome.Six ITE models,including CFR_WASS,were used to predict the 5-year DFS probability of patients under two drug treatments.Six machine learning(ML)models,including Random Forest,were used as baselines for performance analysis and comparison.Model’s Predictive performance was evaluated based on the AUROC.The effectiveness of treatment recommendations was assessed by calculating the difference of 5-year rates between the group where the actual treatment used was consistent with the treatment recommended by the model and the control group.Results Among the 153 matched cases,there was no statistically significant difference in 5-year DFS outcomes between the two groups.In 16 pairs of cases,the PLD group showed better clinical outcomes than the EPI group,and in 12 pairs of cases,the EPI group had better clinical outcomes than the PLD group,confirming individual differences in treatment benefit between the two drugs.The CFR_WASS model achieved the optimal predictive performance(AUROC value was 0.7368),and there was no significant difference in 5-year DFS rates between most ML groups and the control group;The 5-year DFS rate in the ITE group was lower than that in the control group(P<0.01),showing significant differences.Among them,the 5-year DFS rate in the CFR_WASS group was 2.13%lower than that in the control group.Conclusion The ITE model is more accurate in estimating the individualized treatment effects of two drugs compared to the ordinary ML model,providing effective individualized treatment recommendations,and has certain clinical application value.