| Objectives:To develop a pre-treatment CT-based predictive model to anticipate advanced lung cancer patients’ durable clinical benefits to immunotherapy.Methods:In this study,309 patients with advanced lung cancer receiving programmed cell death-1(PD-1)or programmed cell death ligand-1(PD-L1)immune checkpoint inhibitor(ICI)from Peking Union Medical College Hospital were retrospectively collected.Patients without contrast-enhanced CT images and insufficient follow-up time before treatment in our hospital were excluded,and the final patients were randomly divided into the training set and the test set according to a ratio of 4:1.The clinical outcome was divided into durable clinical benefit group(DCB)and non-durable clinical benefit group(NDCB)according to whether the progression-free survival(PFS)reached 6 months.The lesions were semi-automatically outlined by two radiologists through the Dr.Wise platform(Beijing Deepwise Science and Technology Co.,Ltd,Beijing,China)to generate three-dimensional volumes of interest for the extraction of radiomic features.The original features were weighted summed at the case level by the attention-based multiple instances learning algorithm,and then screened by the inter-reader correlation coefficient,Pearson correlation coefficient and inter-group difference analysis to obtain stable features.These features along with clinical features were input into the ridge-based logistic regression to obtain the final features for modeling.Five classifiers(logistic regression,support vector machine,multilayer perceptron,linear discriminant analysis and extreme gradient boosting)were trained with the five-fold cross-validation method in the training set.The area under the receiver operating curve(AUC)and 95%confidence interval of all models were calculated respectively,and the different models were compared with the DeLong test.Accuracy,sensitivity,specificity,F1 value,positive predictive value,and negative predictive value were calculated for the best model in each classifier.The selected clinical features were added to the radiomic model to form an integrated model.The classifiers with the best discriminative performance were selected as reported results and their prognostic values were analyzed in Kaplan-Meier survival curve and Cox regression.Results:Finally,233 patients were assigned to a training set of 185 and a test set of 48,with DCB accounting for 62.9%and 64.9%,respectively.A total of 33 weighted radiomic features and five clinical characteristics(age,pembrolizumab,line of immunotherapy,clinical stage and pre-treatment bone metastasis)were selected.Compared with the largest lesion method(AUCs:0.70-0.78)and the mean average method(AUCs:0.64-0.80),the prediction model based on weighted summed method performed significantly better on all classifiers except multilayer perception(AUCs:0.75-0.82,P<0.05).Among all radiomic models based on weighted summed method,the efficacy of logistic regression model was the most balanced,with AUCs of 0.87(95%confidence interval 0.84-0.89)and 0.75(0.68-0.82)in the training and cross-validation fold of the training set,respectively.The AUC in the test set was 0.80(0.68-0.92).The addition of clinical features significantly improved the performance of the radiomic model(training:AUC 0.91[0.89-0.93],P=0.042;cross-validation:AUC 0.86[0.800.91],P=0.011;test:AUC was 0.86[0.76-0.96],P=0.026).Additionally,the radiomic model(hazard ratio 2.40-2.5,P<0.05)and the integrated model(hazard ratio 2.9-2.95,P<0.05)were able to sufficiently separate the PFS survival curves.Conclusions:The adoption of weighted radiomic features from multiple intrapulmonary lesions has the potential to predict long-term PFS benefits for patients who are candidates for PD-1/PD-L1 immunotherapies.Adding clinical features significantly improved the performance of radiomic-based predictive models.Objective:We attempted to construct an imaging-based biomarker through radiomic features from pre-treatment CT scans to distinguish patients with advanced non-small-cell lung cancer(NSCLC)that were responsive to immunotherapy.Methods:This study reviewed 436 patients with pathologically confirmed advanced NSCLC who received PD-1/PD-L1 ICI therapy from Peking Union Medical College Hospital from June 2015 to February 2022.Patients who did not undergo contrast-enhanced CT scans in our hospital before treatment and those who did not have recorded PD-L1 expression level were excluded,and the final cases were randomly divided into a training set for model training and a test set for model validation.The expression levels of PD-L1 were classified as negative(<1%),low(1-49%),and high(≥50%)according to the tumor proportion score(TPS).The clinical outcomes were best overall responses as classified by partial or complete response(responders)and stable disease or progressive disease(non-responders).Radiomic features were extracted by PyRadiomics and weighted summed at case level by the attention-based multiple instances learning strategy.Case-level features were then selected by inter-reader correlation coefficient,Pearson correlation coefficient and the least absolute shrinkage and selection operator(LASSO).Finally,the radiomic model score(Radscore)was calculated using the logistic regression classifier.We calculated Radscore’s AUC and 95%confidence interval(CI)for receiver operating characteristic curve to distinguish responders from nonresponders,and compared it with the PD-L1 expression level using DeLong test.Then,Kaplan Meier curve and Cox regression analysis were used to analyze the prognostic value of Radscore and PD-L1 expression levels.Finally,subgroup analyses were performed in patients treated with first line therapy,multiline therapy,ICI monotherapy,and combined therapy.Results:A total of 237 patients were randomly assigned to a training set of 165 patients(mean age±standard deviation,64±9 years)and an independent test set of 72 patients(62±8 years).The number of patients with low PD-L1 expression and high PD-L1 expression in the training set(test set)was 37%(42%)and 21%(15%),respectively.Among the 1454 extracted radiomic features,33 stable,non-collinear and representative features were selected to construct the Radscore.Radscore reached significantly higher AUCs for differentiating responders and nonresponders in the training(0.85)and test set(0.80),respectively,compared to the PD-L1 with threshold at 1%(training set:AUC 0.54,P<0.001;test set:AUC 0.63,P=0.017)and with threshold at 50%(AUC 0.53,P<0.001;AUC 0.53,P<0.001).Kaplan Meier analysis for PFS showed clear difference between patients with high and low Radscore(training set:P=0.017;test set:P=0.11).In the subgroup of 113 patients with first line immunotherapy,no significant difference in response rates was identified among PD-L1 negative,low and high expression(52.5%,48.9%and 53.8%;P=0.91).In the subgroup of 92 patients treated with single-agent immunotherapy,significant difference in response rates was identified among PD-L1 negative,low and high expression(8.6%,31.7%,and 25%;P=0.047).Patients with high Radscore had significantly lower risk of experiencing disease progression compared to those with low Rascore in the subgroup of first line immunotherapy(P=0.008)and combined therapy(P=0.0087),respectively.Conclusions:The weighted radiomic model based on pre-treatment CT can be used as an alternative biomarker of PD-L1 to more accurately and effectively predict tumor response and long-term risk of disease progression in patients with advanced NSCLC receiving immunotherapy.In NSCLC patients with single-agent immunotherapy,the tumor response rate was higher in the PD-L1 positive patients compared to the PD-L1 negative patients.However,in first-line immunotherapy,the expression level of PD-L1 was not significantly associated with the tumor response. |
PDF Full Text Request