User-friendly prognostic model for rectal neuroendocrine tumours: In the era of precision management

Abstract

In this letter, we explore into the potential role of the recent study by Zeng et al. Rectal neuroendocrine tumours (rNETs) are rare, originate from peptidergic neurons and neuroendocrine cells, and express corresponding markers. Although most rNETs patients have a favourable prognosis, the median survival period significantly decreases when high-risk factors, such as larger tumours, poorer differentiation, and lymph node metastasis exist, are present. Clinical prediction models play a vital role in guiding diagnosis and prognosis in health care, but their complex calculation formulae limit clinical use. Moreover, the prognostic models that have been developed for rNETs to date still have several limitations, such as insufficient sample sizes and the lack of external validation. A high-quality prognostic model for rNETs would guide treatment and follow-up, enabling the precise formulation of individual patient treatment and follow-up plans. The future development of models for rNETs should involve closer collaboration with statistical experts, which would allow the construction of clinical prediction models to be standardized and robust, accurate, and highly generalizable prediction models to be created, ultimately achieving the goal of precision medicine.

Key Words: Rectal neuroendocrine tumours; High-risk factors; Prognosis; Clinical prediction models; Precision medicine; Statistical collaboration

Core Tip: Rectal neuroendocrine tumours are rare, and these patients generally have a favourable prognosis; however, the presence of high-risk factors can significantly reduce patient survival. The development of predictive models such as the global alliance for trade in services score is crucial for identifying high-risk patients and guiding precise treatment strategies. However, prognostic models capable of conducting comprehensive and rigorous sample size calculations along with multicentre external validation prior to model construction remain somewhat less common. In future research, close collaboration with experts in medical statistics is imperative, with the aim of balancing clinical utility and predictive accuracy throughout the model development and optimization processes to more effectively guide stratified management of patients with rectal neuroendocrine tumours.

TO THE EDITOR

Neuroendocrine neoplasms are rare tumours that originate from peptidergic neurons and neuroendocrine cells[1]. These tumours undergo demonstrate neuroendocrine differentiation and express corresponding markers. Gastroenteropancreatic neuroendocrine neoplasms account for up to 60% of these tumours[2]. With advances in endoscopic technology, the rate of rectal neuroendocrine tumour (rNET) detection has significantly increased[3]. Most patients with rNETs have a good prognosis, but when patients have high-risk factors, such as larger tumours, poorer differentiation types, and lymph node metastasis, the median survival period significantly decreases[4].

Clinical prediction models are essential for guiding diagnosis and prognosis in health care[5]. In recent years, the number of studies focusing on the development of NET prognostic clinical prediction models based on public databases and multicenter data has increased. Despite the numerous guidelines for clinical prediction models[6-8], the recently created rNET models still have many limitations, including an unclear definition of target patients, insufficient sample size, collinearity among variables, and a lack of external validation. These issues have hindered the clinical application of these models.

Creating an rNET prognostic model with strong predictive ability and clinical utility is vital for guiding the precision treatment of rNETs. Close collaboration with statisticians enables us to consider both statistical significance and practical clinical implications. In the future, we can adopt different interventional measures for high-, medium-, and low-risk patients as identified by high-quality rNET prediction models, which could ultimately improve the prognosis of rNET patients.

SHORTCOMINGS OF THE EXISTING rNET PROGNOSTIC MODELS

Although the recently constructed model for rNETs has strong calibration and discrimination capabilities (Table 1), there are still some deficiencies in the actual construction of the prognostic model. In the process of constructing the global alliance for trade in services (GATIS) model, the target population consisted of patients who could undergo resection treatment. Hence, this model is not suitable for patients with distant metastasis. Given that surgery is generally not recommended once rNET patients develop distant metastases[8], a separate prognostic model needs to be developed for these patients. During the construction of rNET-related models, the issue of collinearity is commonly overlooked[9,10]. In the GATIS scoring system, the authors incorporated both tumour size and T stage simultaneously, which are undeniably correlated, resulting in collinearity between model variables and diminishing the model’s generalizability[11]. In addition to the use of several statistical methods, such as least absolute shrinkage and selection operator regression analysis and examining the correlation between variables and the variance inflation factor, clinicians also need to conduct in-depth analyses on more variables from a professional perspective.

Table 1 Prognostic model for rectal neuroendocrine tumours[9-13,16,17].

Number of patients and centres	Target patients	Variables	Outcome of model	Ref.
346 patients, 5 centres	Patients with endoscopic resection	Tumour size lymphovascular invasion, depth of invasion, positive resection margins, mitotic count	The risk of extracolonic recurrence	Lee et al[16]
199 patients, 8 centres	G1-G2 patients with rNETs	Tumour size, vascular invasion	Preoperative lymph node metastasis	Ricci et al[12]
10580 patients SEER database and 68 patients from 1 centres	Patients with rNETs	Age, sex, race, histologic type, tumour size, tumour number, summary stage, and surgical treatment	5-year survival status	Cheng et al[10]
52 patients, 1 centre	Patients with rNETs	Location and radioactivity uptake	Preoperative lymph node metastasis	Zhou et al[17]
85 patients, 17 centres	Patients with high-grade rNETs	Prognostic nutritional index, alkaline phosphatase, lymph node ratio	1 year, 2 and 3 years relapse-free survival	Zeng et al[13]
1408 patients, 17 centres	Patients with rNETs	Tumour grade, T stage, tumour size, age, and a prognostic nutritional index	1, 3, 5 years overall survival and progression-free survival	Zeng et al[11]
1253 patients from SEER database	Patients with rNETs	Tumour stage, tumour size, tumour grade, age, median income	1, 3, 5 years overall survival and cancer-specific survival, lymph node metastases	Chen et al[9]

SEER: Surveillance, epidemiology, and end results; rNETs: Rectal neuroendocrine tumours.

During model construction, it should be ensured that the sample size is substantial enough to guarantee the reliability of the prediction model when it is applied to a new target population. Consequently, determining an appropriate sample size is one of the crucial factors that influences the stability of the model’s predictive power[5]. Unfortunately, no calculation of sample size was conducted in the research on the rNET prognostic model.

During the construction of predictive models, to increase the practicality of the model for clinical use, continuous variables are often converted into categorical variables. However, this conversion may lead to potential loss of some information; thus, we do not recommend converting continuous variables into binary variables[7]. With respect to the NOVARA score established by Ricci et al[12], these authors binarized age and Ki-67 index on the basis of the receiver operating characteristic curve and constructed a prognostic model using these data[12]. During the construction of the GATIS score, X-tile software was used to binarize age and the prognostic nutritional index[11]. Although this approach simplifies the clinical application of the model, such categorization is applicable only to the specific patient population for which the model was constructed and lacks clinical universality.

OPTIMIZATION OF THE PROGNOSTIC rNET MODEL

rNETs usually have a relatively good prognosis, however[8], in scenarios where the number of positive events is small, a relatively large sample size is needed for model construction. Based on the formula for determining sample size and setting the observation endpoint to 5-year survival rate, the number of candidate predictive parameters to 15, the expected average follow-up time to 5.10 years, the outcome event rate to 0.065, the conservative Cox-Snell R squared statistic value to 0.065, the determined sample size was 1853 individuals[5]. However, in the GATIS scoring model constructed by Zeng et al[11], only data from 1408 patients from 17 centers were incorporated, rendering the sample size slightly insufficient. With the exception of the surveillance, epidemiology, and end results data, this study is thus far the multicenter study with the largest sample size in the construction of rNET prognostic models. Therefore, future rNET prognosis research should focus on expanding to multiple centers, even multinational centers, to achieve a more satisfactory sample size. During the selection of model variables, we usually choose variables with P values less than 0.05 in univariate regression analysis for multivariate regression analysis[13]. However, this raises certain questions. Are variables with P values greater than 0.05 in univariate regression analysis truly irrelevant to prognosis? Is there collinearity among the variables incorporated into the model? Does overfitting occur during the model construction process? Constructing a high-quality rNET model requires the joint participation of clinicians and statisticians to ensure that the resulting model has both statistical and clinical significance.

In addition to the intrinsic characteristics of rNETs, in the establishment of future prognostic models, we should also consider the incorporation of new prognostic indicators related to rNETs, such as chromogranin A[14] and somatostatin receptor 2[15], among others. Preoperative enhanced computed tomography scanning is typically utilized to assess whether patients with rNETs larger than 10 mm have concurrent lymph node metastasis, but studies on rNET prognostic models based on radiomics are still rare. Real-time analysis of artificial intelligence images has been deployed in the diagnosis of early gastric cancer and has high sensitivity and specificity. If this technology can be integrated with endoscopic ultrasound (EUS) and analyzed according to different parts of the EUS image, it may increase the accuracy of rNET prognosis prediction. Notably, in multiomics research, statistical methods such as machine learning are frequently used in model construction, so the issue of sample size still necessitates attention.

CONCLUSION

Creating a high-quality rNET prognostic model enables the grading of rNET patients under varying conditions, thereby achieving refined prognosis management. At present, there are some challenges in the construction of prognostic models for rNETs. We should adhere to the existing guidelines for establishing prognostic models and standardize the model construction process, such as the determination of sample size, the selection and transformation of variables, and internal and external validation of the model. Throughout this process, close collaboration with clinical statistical experts is necessary.