Advancements and challenges in esophageal carcinoma prognostic models: A comprehensive review and future directions

Abstract

In this article, we comment on the article published by Yu et al. By employing LASSO regression and Cox proportional hazard models, the article identified nine significant variables affecting survival, including body mass index, Karnofsky performance status, and tumor-node-metastasis staging. We firmly concur with Yu et al regarding the vital significance of clinical prediction models (CPMs), including logistic regression and Cox regression for assessment in esophageal carcinoma (EC). However, the nomogram's limitations and the complexities of integrating genetic factors pose challenges. The integration of immunological data with advanced statistics offers new research directions. High-throughput sequencing and big data, facilitated by machine learning, have revolutionized cancer research but require substantial computational resources. The future of CPMs in EC depends on leveraging these technologies to improve predictive accuracy and clinical application, addressing the need for larger datasets, patient-reported outcomes, and regular updates for clinical relevance.

Key Words: Predictive model; Machine learning; Esophageal carcinoma; Survival rate; Factors

Core Tip: Clinical prediction model has great development space and practical value in the medical field. Despite significant efforts to explore the prognosis of esophageal carcinoma, current prognostic models remain imperfect. Traditional predictive models, such as Cox proportional hazards regression and logistic regression, are widely used but often lack effective evaluation mechanisms to determine their optimal performance. Moreover, due to limitations in sample size and predictive factors, the reproducibility of these models is poor, which severely restricts their broad application in clinical practice. Therefore, it is necessary to further explore and select more appropriate analytical methods to construct more accurate and reliable predictive models, thereby better serving clinical needs.

TO THE EDITOR

Esophageal carcinoma (EC) is one of the predominant malignant tumors in the world[1]. While surgical techniques have improved and multimodal therapies have been incorporated, the prognosis for EC remains unsatisfactory[2]. Clinical prediction models (CPMs), also known as clinical prediction rules, risk prediction models, predictive models, or risk scores, are statistical tools that utilize mathematical formulas to predict the probability of a subject having a certain disease or experiencing a specific outcome in the future[3,4]. These models encompass parametric, semi-parametric, or non-parametric approaches. Currently, the primary models that are widely used in clinical settings are logistic regression and Cox regression, although machine-learning algorithms are also being increasingly applied in the construction of CPMs[5]. These models play a crucial role in clinical decision-making by providing quantified risk assessment, which helps strategize disease prevention, diagnosis, and treatment strategies[6]. This letter delves into the current landscape of prognostic models for EC, highlighting key insights and pinpointing promising avenues for future investigative efforts.

THE CURRENT SITUATION AND DIFFICULTIES IN CONSTRUCTING PROGNOSTIC MODELS FOR EC

Several studies are available on prognostic prediction models for EC both domestically and internationally, and they can mainly categorized into four types: Clinical pathological factor models, tissue immunology models, gene molecular models and databases, and machine algorithm models[7]. The tumor staging system of the American Joint Committee on Cancer and the Union for International Cancer Control commonly used in clinical practice, as well as the improved staging scheme based on them, belong to the clinical pathological factor models from among the abovementioned models. The tumor-node-metastasis staging system falls short of providing an accurate survival prognosis for EC patients owing to its omission of clinical traits, treatment options, and laboratory data.

Based on the complex interaction between cancer and the host immune response, immune molecules may be the key factor in determining the prognosis. In 2018, Jiang et al[8] pioneered the development of a gastric cancer prognosis prediction model that leverages tissue immunology indicators. This groundbreaking study utilized data from 879 patients and, notably, employed immunohistochemistry to assess the expression and survival data of 27 immune characteristics across 251 specimens. The resulting prediction model was expertly crafted using a LASSO-Cox regression model, with a focus on five pivotal immune markers. This approach not only expanded the horizons of CPMs but also underscored the potential of integrating advanced statistical algorithms with immunological data to enhance the prognostic accuracy in cancer research. Although there are only a few studies on the prediction of the prognosis of EC by tissue immunology indexes, it provides a new direction for future research.

With the advent of the post-genomic era, the new generation of high-throughput sequencing technologies in transcriptomics, such as gene chips, RNA sequencing, and other technologies, have been widely applied in the field of medical research with some progress[9]. Recent studies[10,11] have shown that the gene expression profiles can be applied in the identification, staging and guiding of treatment and intervention approaches for a variety of cancers. The prognosis of EC is also related to gene molecules and results in differential prognosis. Despite the significant strides made in genetic research, the integration of factors such as CDK8, ARID1A, and certain autophagy-related long non-coding RNAs into prognostic models is challenged by the high costs and complexity associated with genetic testing[10].

With the increasing prevalence of computer technology and the internet, we have entered the big data era, which is distinguished from the traditional clinical case collection based on its sheer volume, diversity, velocity, and value. In recent years, extensive genomic sequencing initiatives like The Cancer Genome Atlas and the International Cancer Genome Consortium have facilitated the analysis of vast datasets, leading to the discovery of the underlying processes involved in the tumor formation, progression, and prognosis[12]. This aspect has provided valuable insights into the development of personalized diagnostic and therapeutic strategies for cancer patients. However, the vast amount of available data demands substantial processing; this challenge has been effectively addressed by the advent of machine-learning algorithms[13]. In a study by Abuhelwa et al[14], a machine-learning algorithm was created to forecast survival rates in urothelial cancer patients undergoing atezolizumab therapy. The research indicated that the Gradient Boosting Machine model demonstrated superior predictive accuracy compared to alternative models, including CoxBoost, random forests, and generalized linear models, in estimating patient survival outcomes. This study employed LASSO regression to select variables, which is a powerful variable selection tool. However, the criteria for variable selection, including the selection process and reasons for penalty coefficient λ, warrants further elaboration. In addition, it is necessary to consider comparing the results of the LASSO Cox model with other variable selection methods (such as forward selection, backward elimination or AIC/BIC standard) to evaluate the robustness of the model. Although the training set and validation set were used to evaluate the performance of the model, the further utilization of k-fold cross-validation is expected to enhance the reliability of model evaluation.

CONCLUSION

The development and application of these models face challenges such as the need for larger and more diverse datasets, the incorporation of patient-reported outcomes, and the requirement for continuous updating to maintain their accuracy and relevance in clinical practice. The future of CPMs lies in leveraging big data, advanced analytics, and multidisciplinary approaches to enhance their predictive capabilities and clinical utility.