Improve individual treatment by comparing treatment benefits: cancer artificial intelligence survival analysis system for cervical carcinoma

Study dataset was downloaded from Surveillance Epidemiology and End Results (SEER) database (2010–2015). All patients were diagnosed with cervical carcinoma through pathological examination. The diagnostic criteria for cervical carcinoma was in accordance with the suggestions of American Joint Committee on Cancer (AJCC 7 edition). In order to eliminate the effects of confounding factors, living patients with survival time less than 12 months were excluded from the present study. In the study of tumor prognosis, 5 years or 10 years is the most common follow-up period for tumor prognostic study. For a well-designed prognostic study with good patient compliance, the survival time of “living patients” should be infinitely close to the longest follow-up time. The living patients with a survival time shorter than 12 months in the study dataset should consider the following two different situations: the first one is that this patient died within 12 months and can’t continue to follow up. In this case, this died patient defined as a living patient in dataset will has an adverse impact on the study conclusion, so it should be excluded from the current study accordingly. The other one is that this patient is still alive, but can’t be followed up and provide subsequent survival information due to other special reasons. In this case, the survival time of this patient is obviously underestimated, and it will has a significant adverse impact on the study result. Therefore, the living patients who were followed up for less than 12 months were excluded from the current study. Meanwhile, patients who died of causes other than cancer were excluded from the current study. All patients’ privacy information and identity information were anonymized in SEER database. All patients in SEER database signed the informed consent form at the enrollment stage. For the above reasons, ethical review and informed consent were exempted by our institutional review board. There were 14,946 cervical carcinoma patients included in the final survival analysis.

Cox proportional hazard regression model algorithm was performed according to the advices in original articles [22, 23]. Accelerated failure time (AFT) algorithm was performed according to the previous studies [21, 24]. Multi-task logistic regression (MTLR) algorithm was performance in line with the suggestions of the previous articles [25, 26]. The restricted mean survival time is the sum of the areas under the survival curve in a specific time period [27‐31]. As a valuable prognostic index, restricted mean survival time was widely applied to different prognostic studies [27‐31].

Statistical analyses were carried out by SPSS Statistics 21.0 (SPSS Inc., USA). Artificial intelligence algorithms were performed through R software 3.6.0 and Python language 3.7.2 according to previous studies [11‐18]. P value < 0.05 was defined as significant statistical difference.

The current study finally enrolled 14,946 eligible cervical cancer patients. The enrolled patients were randomly divided into model dataset (n = 7536) and validation dataset (n = 7410). The baseline characteristics of patients in model dataset and validation dataset were shown in Table 1.

The current study performed random survival forest algorithm to evaluate the variable importance and explore the error rate with different number of trees. Error rate chart assessed by Out-Of-Bag method was presented in Fig. 1A. Figure 1B listed the most important variables on survival outcome from high to low: stage, PT, chemotherapy, PM, age, and radiation_surgery. Multivariable Cox regression identified stage, PM, chemotherapy, age, PT, and radiation_surgery as independent prognostic factors for cancer specific survival (CSS) of cervical carcinoma in Table 2.

A prognostic nomogram mortality risk predictive chart based on Cox regression model was presented in Fig. 1C. The prognostic score could be calculated by using the following equation: prognostic score = (0.787 * Stage) + (− 0.208 * Chemotherapy) + (0.015 * Age) + (0.299 * PT) + (− 0.263 * Radiation_Surgery) + (0.245 * PM).

The current study further developed a novel Cancer artificial intelligence survival Analysis system (CAISAS) for predicting the prognosis of cervical carcinoma patients. CAISAS was developed based on six previous influence factors through Cox proportional hazard regression model algorithm, accelerated failure time model (AFT) algorithm, and Multi-task logistic regression (MTLR) algorithm. CAISAS could be freely used at: https://​zhangzhiqiao15.​shinyapps.​io/​Tumor_​Artificial_​Intelligence_​Survival_​Analysis_​System/​.

By six major parameters and three artificial intelligence algorithms, CAISAS could provide individual mortality risk predicted curves for a special patient under different treatments.

Cox proportional hazard regression model could provide individual survival predicted curves for a special patient under different treatments (Fig. 2A). The concordance indexes of Cox model were 0.860, 0.849, and 0.848 for 12-month, 36-month, and 60-month in model dataset (Fig. 2B), whereas it were 0.881, 0.845, and 0.841 in validation dataset (Fig. 2D). The higher the C index, the better the diagnostic accuracy. Survival curve charts demonstrated that Cox model could discriminate high mortality risk patients from low mortality risk patients in model dataset (Fig. 2C) and validation cohort (Fig. 2E; Additional file 1).

Accelerated failure time model could provide individual survival predicted curves for a special patient under different treatments (Fig. 3A). The concordance indexes of AFT model were 0.861, 0.852, and 0.851 for 12-month, 36-month, and 60-month in model dataset (Fig. 3B), whereas it were 0.882, 0.847, and 0.846 in validation dataset (Fig. 3D). Survival curve charts demonstrated that AFT model could discriminate high mortality risk patients from low mortality risk patients in model dataset (Fig. 3C) and validation cohort (Fig. 3E).

Multi-task logistic regression model could provide individual survival predicted curves for a special patient under different treatments (Fig. 4A). The concordance indexes of MTLR model were 0.860, 0.863, and 0.861 for 12-month, 36-month, and 60-month in model dataset (Fig. 4B), whereas it were 0.880, 0.860, and 0.861 in validation dataset (Fig. 4D). Survival curve charts demonstrated that MTLR model could discriminate high mortality risk patients from low mortality risk patients in model dataset (Fig. 4C) and validation cohort (Fig. 4E).

The lower the Brier score, the more consistent the predicted results with the actual results. Brier scores of Cox model were 0.126, 0.127, and 0.137 in model dataset, whereas it were 0.116, 0.125, and 0.134 in validation dataset for 12-month, 36-month, and 60-month, respectively. Brier scores of AFT model were 0.133, 0.128, and 0.136 in model dataset, whereas it were 0.124, 0.126, and 0.133 in validation dataset for 12-month, 36-month, and 60-month, respectively. Brier scores of MTLR model were 0.124, 0.126, and 0.137 in model dataset, whereas it were 0.116, 0.124, and 0.133 in validation dataset for 12-month, 36-month, and 60-month, respectively.

Limited by the special requirements for chemotherapy information and radiotherapy information, the current study failed to obtain effective external validation datasets from public databases other than SEER database. Therefore, according to the recommendations of transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) [32], we used the self-help guide resampling method to build different internal validation datasets for evaluating the accuracy of three prognostic models. We re-sampled 14,946 patients from the original 14,946 patients in the way of put back re-sampling to build 5 internal validation datasets. Then we used these 5 internal validation datasets to evaluate the accuracy of three predictive models (Table 2). The evaluation results showed that the C-indexes of MTLR model were the best, and its highest C-indexes of 12-month, 36-month, and 60-month were 0.828, 0.830, and 0.830 respectively, suggesting that MTLR model has the best diagnostic efficiency in three prognostic models. At the same time, Brier scores of MTLR model of 12-month, 36-month, and 60-month were 0.075, 0.121, and 0.130 respectively, showing good consistency between the actual mortality and predicted mortality predicted by MTLR model.

As shown in Fig. 5, AFT algorithm provided predicted mortality percentage and 95% confidence interval at specific time points. This predictive function could provide individual mortality predicted percentage and 95% confidence interval for patients receiving different treatments at 12-month (Fig. 5A), 36-month (Fig. 5B), and 60-month (Fig. 5C). Through comparison of treatment benefits at different time points, this predictive function could provide valuable prognostic information for personalized treatment decision.

To explore the treatment benefits in different stages, CAISAS provided predictive function in providing individual mortality risk predicted curves under different treatments in different stages. Treatment benefits under different treatments were presented in Fig. 6A for stage I, Fig. 6B for stage II, Fig. 6C for stage III, and Fig. 6D for stage IV. Figure 6B, Fig. 6C, and Fig. 6D demonstrated that radiation/surgery and chemotherapy could improve the cancer specific survival in stage II, stage III, and stage IV, whereas Fig. 6A suggested that radiation/surgery and chemotherapy did not improve the cancer specific survival in stage I. The restricted mean survival time could provide lateral prediction of survival time for tumor patients, so as to help patients better understand the survival benefits brought by different treatments. The current predictive system provided the restricted mean survival times for patients receiving various treatments in four tumor stages (Fig. 6).

To explore the treatment benefit of chemotherapy in different stages, CAISAS provided predictive function in providing individual mortality risk predicted curves for patient without chemotherapy and with chemotherapy in different stages. Treatment benefit of chemotherapy under different treatments were presented in Additional file 2: Fig. S1A for stage I, Additional file 2: Fig. S1B for stage II, Additional file 2: Fig. S1C for stage III, and Additional file 2: Fig. 1D for stage IV. As shown in Additional file 2: Fig. S1A, the survival of patients with chemotherapy was significantly poor than that of patients without chemotherapy in stage I (HR = 4.115, P < 0.001), whereas the survival of patients with chemotherapy was significantly higher than that of patients without chemotherapy in stage II, stage3, and stage IV, indicating that chemotherapy did not improve the cancer specific survival in stage I. The current predictive system provided the restricted mean survival times for patients receiving chemotherapy or not in four tumor stages (Additional file 2: Fig. S1).

To explore the treatment benefit of radiation/surgery in different stages, CAISAS provided predictive function in providing individual mortality risk predicted curves for patient without radiation/surgery and with radiation/surgery in different stages. Treatment benefit of radiation/surgery under different treatments were presented in Additional file 3: Fig. S2A for stage I, Additional file 3: Fig. S2B for stage II, Additional file 3: Fig. S2C for stage III, and Additional file 3: Fig. S2D for stage IV. As shown in Additional file 3: Fig. S2A, the survival of patients with radiation/surgery was significantly poor than that of patients without radiation/surgery in stage I (HR = 2.077, P < 0.001), whereas the survival of patients with radiation/surgery was significantly higher than that of patients without radiation/surgery in stage II, stage 3, and stage IV, indicating that radiation/surgery did not improve the cancer specific survival in stage I. The current predictive system provided the restricted mean survival times for patients receiving radiation/surgery or not in four tumor stages (Additional file 3: Fig. S2).

To explore the differences of prognostic factors in different stages, the current study performed multivariable Cox regression in different stages. In stage I, univariable Cox regression identified radiation/surgery and chemotherapy as risk factors for cervical carcinoma (P < 0.001). Multivariable Cox regression demonstrated that chemotherapy was an independent risk factor for cervical carcinoma in stage I subgroup (P < 0.001). For stage II subgroup, stage III subgroup, and stage IV subgroup, radiation/surgery and chemotherapy were proved to be independent protective factors for cervical carcinoma by univariable Cox regression and multivariable Cox regression (Table 3).

First, the current study was not able to further explore the treatment benefits of specific radiotherapy, chemotherapy, and surgery because the SEER database did not provide the detailed radiotherapy, chemotherapy, and surgery information. Second, because the SEER database did not provide the information of the eighth AJCC tumor staging system, the pathological criteria was in accordance with the seventh AJCC tumor staging system in the current study. Third, in order to improve the clinical generality of CAISAS in different regions and hospitals with different medical levels, several valuable diagnostic biomarkers (such as CA242 and CA199) were not included in CAISAS. The addition of serum tumor biomarkers might be helpful to improve the predictive accuracy of the prognostic models. Fourth, CAISAS provided individualized mortality risk prediction based on the current research dataset of 14,946 cervical cancer patients. As far as the prognostic model is concerned, all individual predictive results are closely related to the clinical characteristics of the enrolled patients, so the predicted results have certain limitations and can’t represent an absolute survival predicted result, which is only for the reference of clinicians. Fifth, the current research lacked independent external validation study. Large sample size independent external validation study is very important for tumor long-term prognostic study.

In conclusion, the current study developed a novel cancer artificial intelligence survival analysis system to provide individual mortality risk predictive curves for cervical carcinoma patients based on three different artificial intelligence algorithms. Cancer artificial intelligence survival analysis system could provide mortality predicted percentage at specific time points and explore the actual treatment benefits under different treatments in different stages, which could help patient determine the best individualized treatment. Cancer artificial intelligence survival analysis system was available at: https://​zhangzhiqiao15.​shinyapps.​io/​Tumor_​Artificial_​Intelligence_​Survival_​Analysis_​System/​.