Development Of A Deep Learning Based Survival Prediction Model For Advanced Cervical Cancer And A Prospective Correlation Study

Part I Individualized survival prediction model for advanced cervical cancer:a deep survival learning studyBackgroundAt present,clinical staging remains the most commonly used system for advanced cervical cancer(ACC)in clinical practice,and guidance for clinical diagnosis and treatment decision-making were provided.Primary treatment recommended by NCCN(National Comprehensive Cancer Network)for ACC is mainly concurrent chemoradiotherapy(cCRT),which includes external beam radiotherapy(EBRT)/Intensity modulated radiation therapy(IMRT),brachytherapy and cisplatin-based chemotherapy.However,clinical observations and studies have found that there are remarkable heterogeneit in the outcomes,even that the patients may be in same stages.This indicates that clinical staging is not effective enough in predicting prognosis.Therefore,it is urgent and useful to design an ideal individualized survival prediction model to provide more effective treatment plan for patients with poor prognosis.Cox regression model based on linear relationship hypothesis is still the preferred survival analysis method in clinical cancer research.However,there are still a lot of nonlinear correlation phenomena in reality,which have not been fully explored.Deep learning model could capture complex and nonlinear relationships of data,which has not been fully explored in cancer survival analyses.Indeed,deep learning has been used successfully in radiologic image readings,pathological diagnoses,and treatment response predictions.However,utility of deep learning in cancer survival analyses is rather limited.ObjectivesThis study aims to establish a real and reliable large-sample clinical data set of ACC,and design a new individualized survival prediction model,Deep deep survival learn model(DSLM),for ACC patients receiving cCRT with deep learning.We predicted that the performance of DSLM could be significantly better than the traditional model based on linear hypothesis and common machine learning models.It could be more suitable for individualized survival prediction than FIGO stages.Methods1.To establish a large sample clinical data set of patients with advanced cervical cancer.(1)A total of 8,970 cases of cervical cancer from January 2010 to December 2014 in hospital were systematically reviewed.According to inclusion criteria,1,143 advanced cervical cancer patients were retrospectively enrolled in this study.(2)The dataset included forty-nine epidemiological,clinical and hematological clinical characteristic variables.(3)The whole dataset was randomly split into a training set(n=914,80%)and testing set(n=229,20%).In our study,we used the train set for the development of every predictive model,and the test set for the final independent validation of the DSLM.2.DSLM was developed and validated internally and externally.A deep survival neural network in this study was designed based on Python framework and neural multi-task logistic regression model.Cox proportional hazard(CPH)and Random survival forest(RSF)models were established as baseline.Evaluation of models and statistical analyses:(1)Concordance index(c-index)was used to measure the performances of different models,and Integrated Brier Scores(IBS)to evaluate model performances.(2)Difference between predicted and true number of events in each time window was also measured by error curve.(3)In the testing set,Kaplan-Meier curves for patients staged with conventional staging system were drawn.While,Kaplan-Meier curves,Receiver operating characteristics(ROC)and area under curves(AUC)based on new staging system were also plotted.(4)Personalized patient survival curves were used for the comparison of prediction accuracy of conventional and new staging systems.Results1.Clinical cases and follow-up data were collected.Of 1143 ACC patients(FIGO stage ⅡB-ⅣA),903(79.7%)received EBRT combined with afterload irradiation and 201(17.7%)received IMRT combined with afterload irradiation,29(2.6%)received afterload irradiation only.During the follow-up period of>5 years,268 patients(23.4%)died,and the 5-year survival rate was 76.6%,including 86%of patients in stage ⅡB,75%of patients in stage Ⅲ and 52%of patients in stage Ⅳ.2.The assessment to risk factors of ACC dataset were consistent with clinical cognition,which proved that the data set was highly reliable.The 15 factors most associated with survival outcomes were selected,including the lymph node metastasis and pathological types,fibrinogen,albumin,diastolic blood pressure,etc.The result was consistent with clinical experience,and reminded clinical oncology doctors should pay attention to patients’ blood clotting,nutritional status,treatment-related adverse reactions,blood pressure and other risk factors.3.The performance of the deep survival learning model is significantly better than traditional survival prediction models and common machine learning models.C-indexes for the CPH and RSF models were 0.70 and 0.74,and IBS was 0.15 and 0.14,respectively.While,DSLM achieved a c-index of 0.82 in the training set and a c-index of 0.65 in the testing set,IBS was 0.13 and 0.14,respectively.In the calibration death curves,the median absolute error value is only 0.21 and the mean absolute error value is 0.38.In the calibration survival curves,the predicted curve based on DSLM lied mainly within the confidence ranges of the actual curve,the median absolute error value and the mean absolute error value were 2.3 and 3.1,respectively.4.DSLM risk stratification is more scientific and accurate than traditional clinical staging.Patients were divided into four subtypes based on cutoff risk values,and Kaplan-Meier and ROC curves were drawn for each new types.Kaplan-Meier curves for these subgroups were clearly separated by our model,and our new model achieved higher values of AUC(0.669).5.DSLM can provide survival prediction for individual ACC patients.One ACC patient was randomly selected from four subtypes,and 49 characteristic variables were input into DSLM to draw the individual survival prediction curve of the patient.The survival curve can provide the survival probability of individual ACC patients at any time point.Besides,the survival rates of ACC patients would decrease during the follow up time.Also,the decreasing rates for high and low-risk patients were noticeable different.Conclusion1.Deep learning study is an important method to establish the survival prediction model for ACC patients receiving cCRT.2.Compared with CPH and RSF models,DSLM has better satisfactory performances;3.DSLM can provide more accurate risk stratifications and individualized survival prediction for ACC patients receiving cCRT.Part ⅡA prospective study on the correlations between Th17 cells and the efficacy and prognosis of concurrent chemoradiotherapy in advanced cervical cancerBackgroundPersistent infection by high-risk human papilloma virus(HR-HPV)is a necessary risk factor for uterine cervical cancer(UCC).The main mechanism is the integration of HPV DNA in the fragile region of the host chromosome and the expression of E6 and E7 oncogenes,which inhibits the function of tumor suppressor genes such as P53 and Rb,and finaly leads to immortalization and tumorigenesis.In fact,only a few of patients with HPV infection suffer from cervical lesions and cervical cancer.This suggests that immune function,tumor environment changes and other pathogenic factors also play important roles.Therefore,a great deal of researches focus on antitumor immunity and tumor immune escape mechanism,as well as the relationship between development of cervical cancer.The previous researches have been suggested that the adaptive immune response triggered by T-helper 17(Th17)cells may serve an important role in disease development.Th17 cells are recognized as a particular subset of CD4+T helper lymphocytes,which are characterized by high production of interleukin(IL)-17 and other inflammatory cytokines,such as IL-21,IL-22 and transforming growth factor-β(TGF-β).It is believed that the immune response of the Th17 cells during persistent infection of the genital tract with HR-HPV triggers chronic inflammation with the production of IL17 and other pro-inflammatory cytokines,creating a favorable environment for tumor development.However,due to high plasticity,the role of Th17 cells,as well as mechanism,remains ambiguous in the tumor development and progression.Our previous study found that the proportion of Th17 cells in peripheral blood of patients in UCC(FIGO stage Ⅰ-ⅡA)and Cervical Intraepithelial Neoplasia(CIN)was higher than that of normal women,and the proportion and function of Th17/Treg cells were unbalanced,which became worse with tumor progression.Besides,Th17 cells also accumulated abnormally in tissues of UCC and CIN.Up to now,no study has focused on the alterations in Th17 cells and related cytokines during the treatment for advanced cervical cancer(ACC)and follow up.Objectives1.This study aimed to investigate the alterations in the percentage of circulating Th17 cells and related cytokines in ACC patients received concurrent chemoradiotherapy(cCRT)and analyze the correlationship between the alterations in Th17 cells and treatment efficacy and follow up.2.To establish Cox proportional hazard regression model including 49 characteristic variables of ACC patients in the study group,and Th17 cells variable(whether the percentage of Th17 cells decreased significantly after cCRT)were also used as an immunological indicator.Besides,the effects of Th17 cell variables on survival prediction and model performance were investigated by comparing performance parameters.While,the performance advantages of the Deep survival learning model(DSLM)established in the first part were verified again,and the prospect of the combined application of Th17 cell variable and DSLM was discussed.Methods1.A prospective study with 49 ACC(FIGO Stage ⅡB-ⅢB)patients and 23 controls was conducted.ACC Patients received the cCRT schedule recommended by NCCN guidelines and were followed up for three years.Solid tumor efficacy evaluation criteria(RECIST 1.1)was used to evaluate the clinical efficacy with in one month after treatment.2.Circulating Th17 cells(CD3+CD8-interleukin-17+T cells)and related cytokines IL-17,transforming growth factor-β(TGF-β)IL-10,IL-23,IL-6 and IL-22 were detected before and after cCRT by Flow cytometry and enzyme-linked immunosorbent assay(ELISA).The student t-test or Mann-Whitney U test,paired t-test or Wilcoxon paired signed-rank test,Pearson correlation analysis and Kaplan-Meier analysis were applied for data comparison and correlation analysis between alterations of circulating Th17 cells and treatment efficacy.3.Univariate Cox regression model was used to selecte risk factors from 50 characteristic variables including the Th17 cell variables(whether the percentage of Th 17 cells decreased after cCRT,1=decreased,0=non-decreased).Multivariate Cox regression models were established with the patient data sets containing and excluding Th17 cell variables,and the performance was compared with different models,as well as DSLM established in the first part of the paper.Results1.Clinical efficacy and follow-upThe response rate of ACC patients to cCRT was 92.5%.Progression-free survival(PFS)at 6,12,24 and 36 months was 87.5%,82.5%,77.5%and 70.0%,respectively.Overall survival(OS)at 6,12,24 and 36 months was 95.0%,85.0%,82.0%and 70.0%,respectively.2.The alteration of circulating Th17 cell percentage in ACC patients after cCRT The percentage of circulating Th17 cells in the ACC patients was higher than that in the controls,and it significantly decreased after cCRT(P<0.05).Due to the descending rate(DR)of circulating Th17 cells,patients were divided into two groups,obviously decreasing group(OD group)and non-obviously decreasing group(NOD group).A significant decrease of circulating Th17 cells after cCRT was detected in OD group which had a higher treatment efficacy and longer PFS and OS times than that in NOD group.3.The alteration of serum Th17 related cytokine expression levels after cCRT Compared with the control,ACC patients had higher IL-6,IL-10,IL-22,TGF-βlevels and a lower IL-23 level(P<0.05).After cCRT,IL-6,IL-10,IL-17,IL-23 level significantly increased and TGF-β level significantly decreased compared with the levels before cCRT(P<0.05).4.Correlation analysis between Th17 cells and related cytokinesThere were respectively positive correlations between Th17 cells and IL-17(r=0.493,P=0.001),IL-22(r=0.622,P<0.001),IL-23(r=0.347,P=0.028),TGF-β(r=0.358,P=0.023)before cCRT.After cCRT,we analyzed the relationship between the difference value of Th17 cells decreased and the associated cytokines increased/decreased.It indicated that before and after cCRT the difference of Th17 cell changes was positively correlated with the difference of IL-17(r=0.453,P=0.003),IL-22(r=0.528,P<0.001),IL-6(r=0.399,P=0.011),IL-10(r=0.362,P=0.022)and TGF-β(r=0.431,P=0.005)changes.5.Univariate and multivariate Cox regression analysisUnivariate Cox regression analysis identified seven prognostic risk factors(P value<0.05):FIGO stage,paraaortic lymph node metastasis,other complications,marital status,height,cervical appearance after treatment,and mononuclear cell count.The P value of Th17 cell variable was 0.1558.Based on seven prognostic risk factors,multivariate Cox regression models were established.The model excluding Th17 cell variable identified only two independent predictors:other complications and height,and C-index was 0.7159.However Th17 cell variable included in Cox regression model,four independent predictors were identified:paraaortic lymph node metastasis,other complications,marital status and mononuclear cell count,and the C-index increased to 0.7449,but it was still significantly lower than DSLM established in the first part of the paper(C-index:0.82).ConclusionAn obvious decrease of circulating Th17 cells after cCRT correlated with higher treatment efficacy and longer PFS and OS times in ACC patients.Variations of circulating Th17 cells after cCRT may be an important predictor for clinical efficacy and long-term outcomes.we believe that the variable of Th17 cells,as an immune indicator,could significantly improve the performance of survival prediction model for ACC patients.