Prediction models of colorectal cancer prognosis incorporating perioperative longitudinal serum tumor markers: a retrospective longitudinal cohort study

In this study, we constructed prediction models to evaluate the prognostic value of serum tumor markers in a more intuitive way and found that the predictive performance of CRC prognostic models improved with the incorporation of preoperative CEA, CA19-9, and CA125. The functional data analysis method was innovatively applied to sparse and irregular repeated marker measurements to extract individual-specific longitudinal features, and longitudinal models were created, which were proven to have better discrimination and calibration than preoperative models. The proposed longitudinal prediction model based on MFPCA can provide a personalized dynamic prediction for a new patient, with estimated survival probability updated when a new measurement is collected during 12 months after surgery. The dynamic prediction tool developed based on the model had high clinical applicability, enabling the repeated measurement of perioperative CEA, CA19-9, and CA125 to be fully utilized, which will benefit the prognosis assessment and treatment management of CRC patients.

The improved prediction accuracy of preoperative models with the inclusion of multiple prognostic tumor markers was supported by previous studies. It has been reported that CRC patients with simultaneously positive preoperative CEA, CA19-9, and CA125 tended to have the highest rate and the shortest survival time for death [31] and recurrence [32]. The number of elevated preoperative tumor markers was concluded to significantly predict the prognosis of patients with stage II and III CRC [33]. Recently, attention has been paid to prediction models incorporating multiple tumor markers. Tang et al. developed a predictive nomogram to identify early recurrence, with CA19-9 and CA125 included. Zhu et al. confirmed the clinical implications of CEA, CA19-9, CA125, and positive lymph node scheme (LODDS) in predicting OS of CRC patients and conducted a novel nomogram incorporating the three tumor markers and LODDS. This novel proposed model performed better than the model containing CEA only, consistent with our results [13]. However, the extent to which the further addition of CA19-9 and CA125 to the model including CEA can improve the prediction has not been quantified. In our study, based on the random survival forest model, two preoperative models including different numbers of tumor markers were constructed. Compared with the model with only preoperative CEA, the model including all three tumor markers has significantly improved predictive accuracy. Our results supplement previous research, making the prediction improvement of the combination of markers more intuitive.

In addition to the preoperative level, the dynamic measurements of tumor markers during follow-up after surgery also provide important prognostic information. The changing patterns of postoperative tumor markers are important prognostic predictors [14]. Patients whose CEA continues to rise after surgery have a higher risk of death or recurrence than those whose CEA declines and remains stable [15, 16]. Ma et al. studied the impact of dynamic changes in inflammation and biochemical indicators on the prognosis of perioperative patients and constructed a new prognostic model using dynamic changes to achieve a more accurate prediction of the overall survival and disease-free survival of patients with CRC [34]. Though this predictive model contains dynamic information, limited measurements of indicators were used. And few previous studies incorporated dynamic measurements of tumor markers into prognostic models of CRC. In our prediction model, the longitudinal changing features of perioperative CEA, CA19-9, and CA125 within 12 months after surgery were expressed as principal component scores and included in the survival model for dynamic prediction. Compared with the model including preoperative levels, the predictive performance of the longitudinal model has been greatly improved. Through the comparison between the preoperative CEA&CA19-9&CA125 model and the longitudinal CEA&CA19-9&CA125 model, the prognostic value of the longitudinal changing patterns of the three markers was quantified.

For a specific individual, we expect to update the estimate of prognostic risk based on a newly obtained measurement of tumor marker, for which a dynamic prediction model is needed. It is worth noting that although the longitudinal prediction model incorporating repeated measurements of CEA, CA19-9, and CA125 within 12 months after surgery is not dynamic, it can be used for dynamic prediction. That is, dynamic prediction can be achieved by MFPCA-RSF. In the field of dynamic prediction, the joint model is a commonly used model that can process longitudinal data and survival data simultaneously [35]. Based on the joint model, Cao et al. have established a model to dynamically predict OS of patients with ovarian cancer, with longitudinal CA125 values considered [36]. However, the joint model fits longitudinal data using parametric models, whose parameter distribution is difficult to determine for longitudinal markers with complex patterns. Moreover, the traditional joint model cannot handle multiple longitudinal data. This limits the application of the joint model in our data. The functional data analysis method used in this study is non-parametric, making no assumption about the trajectory and being flexible in non-linear situations. It has been widely used in the feature extraction of longitudinal data. Yan et al. applied FPCA to longitudinal measurements of BCR-ABL gene expression levels to extract features, and use these features as covariates in a Cox proportional hazard model to conduct dynamic predictions of chronic myeloid leukemia [23]. And Luo et al. have extended FPCA to multivariate principal components analysis (MFPCA) and implemented dynamic predictions for Alzheimer’s disease and Parkinson’s disease [24, 25, 37, 38]. On the other hand, the joint model predicts future survival probability based on the Cox proportional hazard regression model, requiring the data to satisfy the proportional hazard assumption. The estimation of Cox model is unstable when there are complex interactions between variables. Therefore, we chose the RSF model, proven to remain robust in the presence of interactions between covariates [20]. However, FPCA-RSF is poor in interpretability. The meaning of the eigenfunctions and corresponding FPCs estimated by FPCA and MFPCA is difficult to explain, and the effect of the covariates on the survival outcome cannot be quantified in RSF [16]. Although we measured variable importance and found that the principal components were important predictors, the contribution of CEA, CA19-9, and CA125 to each principal component was not clear. MFPCA only focused on the common components of CEA, CA19-9, and CA125, with marker-specific components ignored [25].

Our study had several strengths. The performance improvement, brought by the further inclusion of CA19-9 and CA125 in the prognosis prediction model with CEA only, quantified the clinical value of simultaneous measurements of CEA, CA19-9, and CA125. The FPCA and MFPCA methods used to extract longitudinal features in our study are flexible, suitable for various and nonlinear changing patterns of tumor markers. Different from the trajectory analysis methods, the functional data analysis method can estimate longitudinal trajectory features which are individual-specific and time-varying. By extracting individual-specific longitudinal features, longitudinal prediction models containing more prognostic information were proposed, which allows clinicians to make full use of available repeated measurements of tumor markers to make more accurate prognostic predictions. Based on the time dependence of the extracted features, the longitudinal model can be used for dynamic prediction. For a target patient, the predicted risk can be updated timely to reflect the latent prognosis. In addition to time-dependent variables, time-varying effects of time-independent prognostic factors may also contribute to the changed prognosis [21]. And demographic and clinicopathological covariates with possible time-dependent effects have been considered in our study, using the random survival forest method which was robust when the proportional hazards assumption was violated [19].

Our study is subject to the limitations and bias inherent in observational retrospective studies. The inclusion criteria were patients who had at least three marker measurements within 12 months after surgery, which may lead to selection bias. Besides, the proposed dynamic prediction model was constructed and validated using data from South China only, and whether the model can be applied to a wider population still needs to be further verified. A large-scale, multicenter prospective cohort study with regular follow-up will provide more evidences to validate the model for dynamic prediction and increase the impact. In external verification, limited by the number of CRC patients in the SYSU6 cohort, we only predicted the survival probability at 60 months after surgery, to ensure a sufficient number of outcomes. The CRC patients in SYSU6 had a low mortality rate, related to the high economic and medical level of Guangdong province. Despite the differences between patients in YNCH and SYSU6, our proposed dynamic prediction model constructed using data from YNCH performed well for patients in SYSU6, indicating that the model had certain stability and extrapolation. MSI status is a well-known factor associated with the prognosis of colorectal cancer [39]. However, since MSI status was not widely detected until 2017 in Yunnan province with backward economic conditions, it had a high missing rate in the current study, and it was not considered when constructing our models. An exploratory analysis based on available MSI data showed that the predictive performance of prediction models including MSI status had a trend of improvement with the incorporation of CEA, CA19-9, and CA125 (Additional File 2: Figure S9). However, the value of including CEA, CA19-9, and CA125 in prediction models with MSI status needs to be further clarified in a large cohort. It should be mentioned that the proposed model is not suitable for patients with stage IV colorectal cancer. Patients with stage I–III and IV CRC receive different treatment regimens, with stage I–III patients mainly treated with surgical resection and stage IV patients mainly treated with chemotherapy and targeted therapy, resulting in heterogeneity in their changing patterns of tumor markers. Therefore, to reduce the complexity of prediction models, stage IV patients were not included in this study. However, the development of dynamic prediction models for stage IV CRC patients is also valuable and should be studied specifically in the future.

In conclusion, the predictive performance of CRC prognostic models improved with the incorporation of preoperative CEA, CA19-9, and CA125, as well as their perioperative longitudinal measurements. The longitudinal measurements of CEA, CA19-9, and CA125 are recommended to predict the prognosis of CRC patients.