Prognosis values of modied Lauren classication in gastric cancer: a validation from SEER database

Background: It remains controversial as to which pathological classication is most valuable in predicting overall survival (OS) in patients with gastric cancer (GC). We assessed the prognostic performances of three pathological classications in GC and developed a novel prognostic nomogram individually predicting OS. Methods: Patients were identied from the Surveillance, Epidemiology and End Results program. Univariate and multivariate analyses were performed to identify the independent prognostic factors. Model discrimination and model-tting were evaluated by receiver operating characteristic curves and Akaike information criteria. Decision curve analysis was performed to assess clinical usefulness. The independent prognostic factors identied by multivariate analysis were further applied to develop a novel prognostic nomogram. Results: A total of 2,718 eligible GC patients were identied. The modied Lauren classication was identied as one of the independent prognostic factors of OS. It showed superior model discriminative ability and model-tting performance over the other pathological classications, and similar results were obtained in various patient settings. In addition, it showed superior net benets over the Lauren classication and tumor differentiation grade in predicting 3- and 5-year OS. A novel prognostic nomogram incorporating the modied Lauren classication showed superior model discriminative ability, model-tting performance, and net benets over the American Joint Committee on Cancer (AJCC) 8 th Edition TNM classication. Conclusion: The modied Lauren classication showed superior net benets over the Lauren classication and tumor differentiation grade in predicting OS. A novel prognostic nomogram incorporating the modied Lauren classication showed good model discriminative ability, model-tting performance, and net benets. pN stage, pathological N stage; pT stage, pathological T stage. McNeil test comparing AUCs of modied Lauren classication vs. differentiation grade.


Introduction
Gastric cancer (GC) is the fth most prevalent and the third leading cause of cancer death worldwide [1]. It is a complex, heterogeneous entity that encompasses tumors with varying histopathologies, molecular pro les, and behaviors; however, GC is considered as a single entity for the purpose of clinical management and treatment, without regard to its subtype [2,3]. To date, the gold standard for GC prognostication and treatment guidance is the anatomical American Joint Committee on Cancer (AJCC) tumor-node-metastasis (TNM) classi cation [4,5]. It has been widely applied in many clinical practices without reference to its histopathology because the value of the morphological features of GC in determining clinical outcomes is still limited [6]. In addition, many investigators are still trying to identify a more valuable classi cation with better prognostic value [3,7,8].
Due to the wide variations in the morphological features of GC, many histological classi cations have been proposed, and they are currently in wide use [3,[9][10][11][12][13]. One of these classi cations is the tumor differentiation grade. GC can be classi ed as well differentiated, moderately differentiated, poorly differentiated, and undifferentiated, according to the degree of differentiation exhibited by the tumor [10]. The tumor differentiation grade has been identi ed as a prognostic risk factor for GC in some studies [14,15]. However, several recent studies have reported that the tumor differentiation grade is not signi cantly associated with the prognosis of GC patients [16][17][18][19]. Another classi cation is the Lauren classi cation [13]. Although it dates back to 1965, it remains one of the most commonly used pathological classi cations in GC. This classi cation categorizes GC into intestinal, diffuse, or mixed types, according to its histology, and each type has a distinct pathology and prognosis [13,[20][21][22]. However, several studies have reported that the Lauren classi cation is not signi cantly correlated with patient survival because anatomic and corresponding epidemiologic distinctions were not taken into account [23,24].
Recently, it has been proposed that the Lauren classi cation be modi ed to include both the Lauren classi cation and the anatomical location of GC, thus yielding at least three entirely distinct types, including the proximal nondiffuse type, distal non-diffuse type, and diffuse type [3]. Molecular biology analyses further showed that there were marked differences in the mRNA expression pro les of the three types. Recent studies performed in Asia also suggested that the modi ed Lauren classi cation could be a reliable prognostic factor for patients with GC [25,26].
However, it remains controversial as to which pathological classi cation is most valuable in predicting the overall survival (OS) in GC patients. Therefore, we aimed to assess the prognostic values of the tumor differentiation grade, Lauren classi cation, and modi ed Lauren classi cation in GC patients. We compared model discriminative ability, model-tting performance, and net bene ts to identify the optimal prognostic pathological classi cation for GC based on the updated Surveillance, Epidemiology, and End Results (SEER) program. We also developed a novel prognostic nomogram for individually predicting the 3-and 5-year OS by applying the optimal pathological classi cation.

Data source
We included data of eligible primary operable gastric cancer patients from the SEER program (https://seer.cancer.gov/). Data were extracted by SEER*Stat 8.3.6 software (www.seer.cancer.gov/seerstat). The data-use agreement for the SEER program data le was approved. Ethical review was not required because the SEER program is a publicly available database with anonymized data.

Clinicopathologic features
The analyzed clinicopathologic features included gender, age, tumor size, depth of tumor invasion (pT stage), number of retrieved lymph nodes, number of positive lymph nodes (pN stage), tumor differentiation grade, and Lauren classi cation. Patients were uniformly reviewed and re-staged (pT or pN stage) according to the AJCC 8 th Edition TNM classi cation [4]. The last follow-up was in November 2016. The OS was de ned as the time of diagnosis to the time of death from any reason.

Statistical analysis
The OS was calculated from the time of diagnosis to the time of death from any reason. Kaplan-Meier survival curves with log-rank tests were applied to analyze the difference in the OS among the groups. Factors with P-values less than 0.1 in univariate analysis were considered potential prognostic factors and included in the Cox proportional hazards regression model. Hazard ratios (HRs) with 95% con dence intervals (CIs) were applied.
The model discriminative ability of different pathological classi cations was assessed by receiver operating characteristic (ROC) curves (AUCs) [27]. The model-tting performance was evaluated by Akaike information criteria (AIC). A higher AUC value indicated a better model discriminative ability, and a lower AIC value indicated a superior model-tting performance. The differences in AUC values were assessed by Hanley and McNeil tests [28]. Decision curve analysis (DCA) was performed to assess clinical usefulness, and the net bene ts of making a decision based on the models were calculated [29,30].
The modi ed Lauren classi cation is an adjusted categorization of the Lauren classi cation, and both classi cations are considered highly relevant. The Cox proportional hazards regression model was employed by incorporating either the Lauren or modi ed Lauren classi cation. Finally, the independent prognostic factors identi ed by multivariate analysis were applied to the nomogram.

Patient characteristics
A total of 2,718 eligible patients with gastric cancer from the SEER program were included. The clinicopathological characteristics are summarized (Table 1). There were 1,588 males (58.4%) and 1,130 were females (41.6%). The median age of all patients was 61 years (range, 18-75 years), and the median follow-up period was 31 months (range, 2-155 months).

Prognostic Factors Of Overall Survival
Univariate analysis identi ed potential prognostic factors, namely age, tumor size, number of retrieved lymph nodes, pT stage, pN stage, tumor differentiation grade, and the modi ed Lauren classi cation (log-rank tests, all P < 0.10). These factors were further applied in multivariate analysis with the Cox proportional hazards regression model. The results indicated that the independent prognostic factors predicting OS were age, tumor size, number of retrieved lymph nodes, pT stage, pN stage, and the modi ed Lauren classi cation (Table 2). However, neither the tumor differentiation grade (P = 0.115) nor the Lauren classi cation (P = 0.163) was found to be an independent predictive factor of OS in further multivariate analysis (Supplementary Table 1).  Fig. 1A and 1B, Fig. 2). The modi ed Lauren classi cation also showed superior model discriminative ability (higher AUC values) and modeltting performance (lower AIC values) in patients that were strati ed by gender, age (female, male), tumor size (< 60 years, ≥ 60 years), number of retrieved lymph nodes (< 16, ≥ 16), pT stage (pT1, pT2-4), and pN stage (pN0, pN1-3) ( Table 4). These results con rmed that the modi ed Lauren classi cation showed the best model discriminative ability and model-tting performance among the three pathological classi cations.

Clinical Utility Of Pathological Classi cations
We conducted decision curve analysis (DCA) to assess the clinical utility of the different pathological classi cations. The results revealed that the modi ed Lauren classi cation had superior net bene ts over the Lauren classi cation and tumor differentiation grade in predicting both 3-and 5-year OS ( Fig. 1C and 1D). Speci cally, the modi ed Lauren classi cation showed superior net bene ts over the tumor differentiation grade between threshold probabilities of 50-65% and 40-80% in predicting 3-and 5-year OS, respectively ( Fig. 1C and 1D). In addition, the modi ed Lauren classi cation also showed superior net bene ts over the Lauren classi cation between threshold probabilities of 30-45% and 40-60% in predicting 3-and 5-year OS, respectively ( Fig. 1C and 1D Fig. 2C-2D, Fig. 3B).
We further conducted DCA to assess the clinical utility of the novel prognostic model and the AJCC 8th Edition TNM classi cation. The novel prognostic model showed superior net bene ts over the AJCC 8th Edition TNM classi cation between threshold probabilities of 40-90% and 50-95% in predicting 3-and 5-year OS, respectively ( Fig. 1E and 1F).

Discussion
Several pathological classi cations of GC are currently in use due to the various morphological characteristics of GC [3,[9][10][11][12][13]. However, it remains controversial as to which classi cation is best. Therefore, we performed a systematic analysis of the three most well-known pathological classi cations and compared prognostic predictive performance with clinical use. In addition to the commonly used Lauren classi cation and tumor differentiation grade, we also compared a new classi cation, the modi ed Lauren classi cation. In our study, pN and pT stages were the most important prognostic factors for survival, thus validating the quality of the participants.
Tumor differentiation grades are commonly used for GC, and the four types of GC are de ned as well differentiated, moderately differentiated, poorly differentiated, and undifferentiated [31]. It has been widely accepted that poorly differentiated tumors usually spread more extensively than well differentiated tumors by the time of surgery, and patients with more differentiated tumors have obvious survival advantages after curative resection [14,15]. However, recent studies have reported that the tumor differentiation grade is not signi cantly associated with the prognosis of patients with GC [16][17][18][19]. In the current study, the tumor differentiation grade was signi cantly associated with the prognosis in log-rank tests; however, it was not an independent prognostic factor of OS. This discrepancy may be due to the mixture of differentiated and undifferentiated GC histologies [18,32]. In addition, it suggests that some welldifferentiated types of GC can change to poorly differentiated types with tumor progression [33,34]. Therefore, further studies are needed to understand the signi cance of the tumor differentiation grade of GC.
The Lauren classi cation of GC is one of the most widely applied histological grading systems in predicting survival [21]. It has been reported that Lauren-classi ed tumor subtypes can respond differently to chemotherapy, thus yielding different survival outcomes [20]. However, the Lauren classi cation has also been demonstrated to have inadequate prognostic discriminative performance, and therefore, its prognostic accuracy remains controversial [23,24]. Speci c pathogenetic and morphologic features of intestinal and diffuse types may underlie their different behaviors [22]. Population-based studies have reported the different epidemiological features of Lauren-classi ed subtypes and cancer of the cardia [35,36]. Epidemiologically, the intestinal type of GC, particularly that of the antrum, is often strongly associated with chronic in ammation as a consequence of chronic infection with H. pylori [37,38]. Anatomically, proximal GC can be classi ed as a third type of GC for which in ammation of a different type may be the driving force for carcinogenesis [39]. Furthermore, the anatomical location of GC is clinically relevant, and proximal third GC is associated with a worse prognosis than middle or distal third GC [40,41].
Therefore, a location-modi ed Lauren classi cation has been proposed. It de nes the subtypes of GC by incorporating epidemiological and histopathological data together with the anatomical location. 3 Several studies have revealed that the modi ed Lauren classi cation has better discriminative ability and monotonicity than the Lauren classi cation [25,26]. The results of the current study demonstrated that the modi ed Lauren classi cation showed superior model discriminative ability, model-tting performance, and net bene ts compared with other classi cations. Similar ndings were also obtained in populations strati ed by gender, age, tumor size, number of retrieved lymph nodes, pT stage, and pN stage. Decision curve analysis con rmed its clinical usefulness over other classi cations.
It remains unclear why the modi ed Lauren classi cation showed a signi cantly better prognostic performance. A previous study has reported that the Kirsten Rat Sarcoma Viral Oncogene Homolog pathway was downregulated in proximal non-diffuse gastric cancer compared with diffuse gastric cancer [42]. In addition, genomic analysis has con rmed that the modi ed Lauren classi cation can achieve a clear molecular distinction [3]. Moreover, HER2 ampli cation or overexpression is not uniform across different GC subtypes; it is most prevalent in proximal GC (a HER2 positivity rate of ~ 30%) and least prevalent in diffuse GC (a HER2 positivity rate of ~ 5%) [43]. Furthermore, whole-genome sequencing of diffuse GC uncovered mutations in RHOA, a gene encoding a well-studied small GTPase, in 15-25% of diffuse tumors but not in non-diffuse tumors [44].
Nomograms are visualization tools for individually predicting survival [45,46] with improved predictive accuracy and comprehensive outcomes for many types of cancers [47][48][49][50][51][52]. Therefore, we developed a novel prognostic nomogram of age, tumor size, number of retrieved lymph nodes, pT stage, pN stage, and the modi ed Lauren classi cation. This novel prognostic model achieved superior model discriminative ability, model-tting performance, and net bene ts over the AJCC 8th Edition TNM classi cation. These ndings support the consideration of more factors spanning different aspects of the disease as the most promising approach to improve the clinical management of GC. However, the ndings of the current study still need to be interpreted with caution because speci c intervention factors of the surgical procedures, chemo-radiotherapeutic regimens, and drug doses were not applied in the current study.

Conclusion
In summary, the modi ed Lauren classi cation provides superior model discriminative ability, model-tting performance, and net bene ts over the tumor differentiation grade and Lauren classi cation. It also shows good       Five-year OS in different prognostic models.