Risk factors of tumor invasion and node metastasis in early gastric cancer with undifferentiated component: a multicenter retrospective study on biopsy specimens and clinical data.

BACKGROUND: Early gastric cancer (EGC) with undifferentiated component (UDC) is a more aggressive entity, where the significance of preoperative data to tumor invasion and lymph node metastasis (LNM) remains unclarified.
METHODS: A total of 5,020 GC patients undergoing radical gastrectomy in three centers were reviewed, of which, EGC with UDC in preoperative biopsy specimens were enrolled. The histology of biopsy and surgical specimens was graded according to the proportion of UDC and signet ring cells (SRCs). Risk factors of tumor invasion and LNM were evaluated with histological, clinical and demographic data.
RESULTS: Lower body mass index (BMI), melena and larger tumor size were the independent preoperative risk factors of both LNM and LVI, while ulcerative lesion (UL) and the lower third stomach were only correlated with LNM. No relevance was found between the histological features of biopsy specimens and LNM, but SRC or >50% UDC lowered the risk of lymphovascular invasion (LVI) and/or submucosal (SM) invasion. When surgical data (depth of invasion and LVI included) were added, lower BMI, melena and the lower third stomach were still the independent preoperative risk factors of LNM, and LVI, SRC and SM invasion also showed relevance to LNM. The performance of predictive models using pre- or postoperative histological data was comparable.
CONCLUSIONS: The preoperative data were significantly relevant to tumor invasion and LNM, showing comparable risk strength with surgical specimens in histology. 2020 Annals of Translational Medicine. All rights reserved.

Introduction

Gastric cancer (GC) is the fifth most common cancer and the third leading cause of cancer death worldwide, responsible for over 1,000,000 new cases in 2018 and 1 in every 12 deaths globally (1). Although the high mortality GC causes, the early diagnosis and intervention improve the 5-year survival rate from 20–30% in patients at late stage (2) to 70–90% in early gastric cancer (EGC) (3,4). Recently, endoscopic submucosal dissection (ESD) plays an increasingly important role in EGC therapy. However, in the absolute indication for ESD established by the Japanese Gastric Cancer Association, only non-ulcerated, well-differentiated and mucosal ≤2 cm lesions were included, and the expanded criteria, including non-lymphovascular-infiltrated, larger, ulcerated, and undifferentiated mucosal lesions, or differentiated lesions with slight submucosal (SM) invasion, remain controversial (5). How to identify patients with high risk of lymph node metastasis (LNM) accurately is the core problem so far.

Many scholars have tried to identify the risk factors of LNM in previous studies. It was reported that the rates of LNM in intramucosal or EGC varied greatly, ranging from 0.4% to 24% (2,3,6,7). According to histological manifestation, EGC is divided into differentiated and undifferentiated type, the former covers papillary adenocarcinoma, well and moderately-differentiated tumor and the latter includes poorly-differentiated tumor and signet ring cell carcinoma (SRCC) (8). For total or differentiated EGC, various demographic and clinicopathological features were picked out as the risk factors of LNM from original research, meta-analysis or data of Surveillance, Epidemiology and End Results program (SEER), such as patients’ gender (6), age (9), tumor location (3,6), gross pattern (10), tumor size (10), ulcerative lesion (UL) (2), depth of invasion (11), histological differentiation (10), and lymphovascular invasion (LVI) (11). For pure undifferentiated EGC, gender, age, tumor location, tumor size, gross pattern, LVI, and depth of invasion were also identified as the risk factors of LNM in previous studies (12-14). However, evidence is still lacking concerning differentiated EGC accompanied with undifferentiated components (UDC) (5), as the presence of UDC showed significantly different biological behaviors when their proportion varied (2,11,15-17).

More importantly, nearly all the previous studies were based on the clinicopathological features of ESD/surgical specimens, which means all the risk factors mentioned above were postoperative. In this scenario, patients with high risk of LNM could only be identified at the post-ESD stage, and they have to undergo the additional gastrectomy and lymph node dissection after an unnecessary but innocent ESD. Therefore, in the present study, we focused on the relationship between the histological features of biopsy specimens and clinical data, and found the significant association between preoperative data and tumor behaviors.

Methods

Patient selection

This was a multi-center retrospective study. The pathological reports of 5,020 patients who underwent radical gastrectomy and lymph node dissection between 2013–2018 in Second Affiliated Hospital of Zhejiang University, Union Hospital of Fujian Medical University and Second Affiliated Hospital of Fujian Medical University were reviewed. Cases met the following inclusion criteria were enrolled in this study: patients were pathologically diagnosed as early gastric adenocarcinoma (pT1 tumor); detailed pathological reports of radical gastrectomy (lymph node dissection included) and hematoxylin & eosin (H&E) slides of matched preoperative biopsy specimens, as well as their clinical data (endoscopic data included) were available; UDC was present in preoperative biopsy specimens (). The study was approved by the Institutional Review Board of Second Affiliated Hospital of Zhejiang University (2019-ERR-446), Union Hospital of Fujian Medical University (2020KY005) and Second Affiliated Hospital of Fujian Medical University (2019-SAHFMER-208). Patient consent was waived by the institutional review boards, as this study was retrospective and patients’ information was protected by blind method.

Figure 1

Flowchart of case selection. +, presence; −, absence. UDC, undifferentiated component; LNM, lymph node metastasis.

Clinicopathological analysis

The histology of biopsy and surgical specimens was graded on the proportion of UDC first. Tumors composed of <50% UDC were classified into U1, those composed of ≥50% but ≤90% UDC were classified into U2, and those composed of >90% UDC were classified into U3. Then, tumors containing SRC were classified into S1, and those without SRC were classified into S0. The information of depth of invasion, LVI and LNM was retrieved from the pathological reports of surgical specimens.

Demographic and clinical data collection

Age, sex, body mass index [BMI, BMI = weight (kg)/height (m)2], and the presence of weight loss, melena or hematemesis was collected from Electronic Medical Record System. Tumor location (cardia and fundus, body, or, angle and antrum were classified into the upper, middle or lower third stomach, retrospectively), tumor size (the maximal diameter), the presence of UL, and main gross patterns evaluated based on the Paris endoscopic classification of superficial neoplastic lesions (18) were retrieved from original endoscopic reports.

Statistical analysis

The χ2 test or Fisher’s exact test was used to analyze the difference between demographic and clinicopathological ratios, and analysis of variance (ANOVA) was used to compare the quantitative data. Correlation between factors was evaluated by Pearson or Spearman test. Risk analysis was performed using binary logistic regression, and the model performance was evaluated by receiver operating characteristic curve (ROC) curves.

A P value less than 0.05 was considered statistically significant. Statistical analyses were performed using SPSS 22.0 (SPSS Inc., Chicago, IL, USA).

Results

Demographic and clinical characteristics

A total of 323 patients were enrolled in this study, of which 247 (76.5%), 42 (13.0%) and 34 (10.5%) were from Second Affiliated hospital of Zhejiang University, Union Hospital and Second Affiliated Hospital of Fujian Medical University, respectively. Patients aged from 24 to 85, with a slight male predominance. Two hundred and seventeen (67.2%) patients were in normal BMI range (18.5–24 kg/m2) with other 21 (6.5%) and 85 (26.3%) showing low (<18.5 kg/m2) and high (≥24 kg/m2) BMI. Eighty-six (26.6%) patients presented with at least one of the specific symptoms (weight loss, melena and/or hematemesis). Tumors were relatively small and located at the lower third stomach mostly, followed by the middle and upper third stomach. Most of them showed the gross pattern of IIc, and UL was seen in nearly 40% of the patients ().

Table 1

Demographic and clinical characteristics (%)

Factors	Total	Node metastasis		P value
		Negative	Positive
Age (year)
Mean ± SD	56.2±11.6	55.8±11.8	57.4±10.9	0.309
Median	57	56.9	57.9	–
Sex
Male	174 (53.9)	138 (54.3)	36 (52.2)	0.786
Female	149 (46.1)	116 (45.7)	33 (47.8)
BMI
Mean ± SD	22.5±3.0	22.6±3.0	22.3±2.9	0.562
Median	22.5	22.5	22.2	–
Weight loss
−	261 (80.8)	210 (82.7)	51 (73.9)	0.120
+	62 (19.2)	44 (17.3)	18 (26.1)
Melena
−	290 (89.8)	237 (93.3)	53 (76.8)	<0.001
+	33 (10.2)	17 (6.7)	16 (23.2)
Hematemesis
−	312 (96.6)	247 (97.2)	65 (94.2)	0.257
+	11 (3.4)	7 (2.8)	4 (5.8)
Tumor location
Upper third	80 (24.8)	69 (27.2)	11 (15.9)	0.033
Middle third	108 (33.4)	88 (34.6)	20 (29.0)
Lower third	135 (41.8)	97 (38.2)	38 (55.1)
Tumor size (cm)
Mean ± SD	2.5±1.5	2.3±1.4	3.1±1.8	0.002
Median	2.0	2.0	2.6	–
Main gross pattern
IIa	24 (7.4)	19 (7.5)	5 (7.2)	0.570
IIb	64 (19.8)	54 (21.3)	10 (14.5)
IIc	231 (71.5)	178 (70.1)	53 (76.8)
III	4 (1.2)	3 (1.2)	1 (1.4)
UL
−	195 (60.4)	163 (64.2)	32 (46.4)	0.008
+	128 (39.6)	91 (35.8)	37 (53.6)

The italic numbers are P values with statistical significance (<0.05). −, absence; +, presence. BMI, body mass index; UL, ulcerative lesion.

LNM was found in 69 (21.4%) patients who presented with melena at a higher rate. Tumors with LNM were located at the lower third stomach more frequently, accompanied with UL and larger size ().

Clinicopathological characteristics

For the histology of either biopsy or surgical specimens, most cases were classified into grade U3, and nearly half of the cases harbored SRC. No significant histological difference was found between different LNM status. However, tumors with LNM showed deeper invasion and higher LVI rate in surgical specimens. To evaluate the representativeness of biopsy specimens, the correlation of histology between biopsy and corresponding surgical specimens was analyzed, from which significant positive relevance of UDC and SRC grades was found ().

Table 2

Pathological characteristics (%)

Data source	Total	Node metastasis		P value
		Negative	Positive
Biopsy
UDC grade*
U1	44 (13.6)	35 (13.8)	9 (13.0)	0.518
U2	43 (13.3)	31 (12.2)	12 (17.4)
U3	236 (73.1)	188 (74.0)	48 (69.6)
SRC grade**
S0	177 (54.8)	136 (53.5)	41 (59.4)	0.415
S1	146 (45.2)	118 (46.5)	28 (40.6)
Surgery
UDC grade*
U1	24 (7.4)	22 (8.7)	2 (2.9)	0.280
U2	92 (28.5)	72 (28.3)	20 (29.0)
U3	207 (64.1)	160 (63.0)	47 (68.1)
SRC grade**
S0	169 (52.3)	136 (53.5)	33 (47.8)	0.418
S1	154 (47.7)	118 (46.5)	36 (52.2)
Depth of invasion
LP	113 (35.0)	103 (40.6)	10 (14.5)	<0.001
MM	52 (16.1)	43 (16.9)	9 (13.0)
SM	158 (48.9)	108 (42.5)	50 (72.5)
LVI
−	284 (87.9)	238 (93.7)	46 (66.7)	<0.001
+	39 (12.1)	16 (6.3)	23 (33.3)

*, r=0.523, P<0.001; **, r=0.491, P<0.001. The italic numbers are P values with statistical significance (<0.05). −, absence; +, presence. SRC, signet ring cell; LP, lamina propria; MM, muscularis mucosa; SM, submucosa; LVI, lymphovascular invasion.

Risk analysis

Preoperative parameters

In univariate analyses, including demographic, clinical and histological features of biopsy specimens: lower BMI, the presence of melena, larger tumor size and the lower third stomach were the risk factors of LNM (); lower BMI, the presence of melena, larger tumor size and the lower third stomach were the risk factors of LVI (); larger tumor size and the presence of UL were the risk factors of SM invasion (). While, the presence of SRC (S1) was the protective factor of LVI (); the SRC grade of S1 and the UDC grade of U3 were the protective factors of SM invasion (); in T1a patients, the gender of female and the middle third stomach were the protective factors of muscularis mucosa (MM) invasion (). The detailed data of the univariate analyses could be found in .

Figure 2

Univariate analysis of pre- and postoperative data. (A) The univariate analysis for LNM; (B) the univariate analysis for LVI; (C) the univariate analysis for SM invasion; (D) the univariate analysis for MM invasion in T1a patients; (E) the univariate analysis for LNM; (F) the univariate analysis for LNM with data including LVI and depth of invasion. *, statistical significance. BMI, body mass index; UL, ulcerative lesion; LVI, lymphovascular invasion; Ref., reference; HR, hazard ratio.

Table S1

Multivariate analysis for LNM with preoperative data

Factors	Univariate analysis
	P	HR	95% CI
			Lower	Upper
Sex
Male	Ref.	1	–
Female	0.684	1.131	0.626	2.041
Age
CV	0.177	0.985	0.963	1.007
BMI
CV	0.014	0.919	0.859	0.983
Weight loss
−	Ref.	1	–
+	0.511	1.264	0.629	2.538
Hematemesis
−	Ref.	1	–
+	0.806	1.214	0.259	5.694
Melena
−	Ref.	1	–
+	0.002	4.535	1.778	11.570
Location
Upper third	Ref.	1	–
Middle third	0.717	1.170	0.500	2.735
Lower third	0.007	3.063	1.360	6.897
Size
CV	0.002	1.357	1.123	1.639
Gross pattern
IIa	Ref.	1	–
IIb	0.166	0.398	0.109	1.464
IIc	0.653	0.774	0.254	2.363
III	0.972	0.950	0.054	16.623
UL
−	Ref.	1	–
+	0.088	1.797	0.916	3.526
UDC grade
U1	Ref.	1	–
U2	0.655	1.288	0.425	3.909
U3	0.823	0.903	0.369	2.209
SRC grade
S0	Ref.	1	–
S1	0.419	0.765	0.400	1.464

−, absence; +, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion; UDC, undifferentiated component; SRC, signet ring cell.

Table S2

Multivariate analysis for LVI with preoperative data

Factors	Univariate analysis
	P	HR	95% CI
			Lower	Upper
Sex
Male	Ref.	1	–
Female	0.426	0.725	0.328	1.601
Age
CV	0.791	0.996	0.968	1.025
BMI
CV	0.001	0.865	0.792	0.944
Weight loss			–
−	Ref.	1	–
+	0.598	0.774	0.300	2.001
Hematemesis
−	Ref.	1	–
+	0.864	1.183	0.172	8.155
Melena
−	Ref.	1	–
+	0.012	4.284	1.381	13.294
Location
Upper third	Ref.	1	–
Middle third	0.481	1.519	0.474	4.867
Lower third	0.042	3.224	1.045	9.948
Size
CV	0.001	1.503	1.190	1.898
Gross pattern
IIa	Ref.	1	–
IIb	0.558	0.612	0.119	3.161
IIc	0.894	1.101	0.267	4.535
III	0.843	1.401	0.050	39.111
UL
−	Ref.	1	–
+	0.569	0.774	0.319	1.873
UDC grade
U1	Ref.	1	–
U2	0.099	2.845	0.823	9.834
U3	0.834	0.892	0.305	2.607
SRC grade
S0	Ref.	1	–
S1	0.009	0.290	0.115	0.732

−, absence; +, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion; UDC, undifferentiated component; SRC, signet ring cell.

Table S3

Multivariate analysis for SM with preoperative data

Factors	Univariate analysis
	P	HR	95% CI
			Lower	Upper
Sex
Male	Ref.	1	–
Female	0.429	0.819	0.499	1.343
Age
CV	0.221	1.012	0.993	1.031
BMI
CV	0.804	1.007	0.951	1.067
Weight loss			–
−	Ref.	1	–
+	0.260	0.701	0.377	1.301
Hematemesis
−	Ref.	1	–
+	0.458	1.781	0.388	8.179
Melena
−	Ref.	1	–
+	0.791	1.128	0.463	2.748
Location
Upper third	Ref.	1	–
Middle third	0.101	0.585	0.309	1.110
Lower third	0.719	0.891	0.475	1.671
Size
CV	0.000	1.364	1.146	1.623
Gross pattern
IIa	Ref.	1	–
IIb	0.523	0.698	0.231	2.104
IIc	0.796	0.877	0.326	2.359
III	0.512	0.442	0.038	5.083
UL
−	Ref.	1	–
+	0.005	2.287	1.291	4.050
UDC grade
U1	Ref.	1	–
U2	0.055	0.361	0.128	1.021
U3	0.001	0.241	0.102	0.567
SRC grade
S0	Ref.	1	–
S1	0.041	0.574	0.337	0.978

−, absence; +, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion; UDC, undifferentiated component; SRC, signet ring cell.

Table S4

Multivariate analysis for MM with preoperative data of T1a patients

Factors	Univariate analysis
	P	HR	95% CI
			Lower	Upper
Sex
Male	Ref.	1	–
Female	0.000	0.204	0.089	0.469
Age
CV	0.325	0.985	0.956	1.015
BMI
CV	0.099	1.081	0.985	1.187
Weight loss
−	Ref.	1	–
+	0.057	2.497	0.972	6.417
Hematemesis
−	Ref.	1	–
+	0.800	0.673	0.031	14.390
Melena
−	Ref.	1	-
+	0.572	1.643	0.293	9.212
Location
Upper third	Ref.	1	–
Middle third	0.006	0.223	0.077	0.646
Lower third	0.178	0.507	0.189	1.362
Size
CV	0.081	1.298	0.968	1.740
Gross pattern
IIa	Ref.	1	–
IIb	0.392	0.457	0.076	2.741
IIc	0.620	0.664	0.132	3.339
III	0.970	0.938	0.032	27.767
UL
−	Ref.	1	–
+	0.336	1.570	0.627	3.936
UDC grade
U1	Ref.	1	–
U2	0.121	0.166	0.017	1.602
U3	0.345	0.457	0.090	2.325
SRC grade
S0	Ref.	1	–
S1	0.747	1.147	0.498	2.641

−, absence; +, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion; UDC, undifferentiated component; SRC, signet ring cell.

In corresponding multivariate analyses: lower BMI, the presence of melena/UL, larger tumor size and the lower third stomach were the independent risk factors of LNM; lower BMI, the presence of melena and larger tumor size were the independent risk factors of LVI; larger tumor size and the presence of UL were the independent risk factors of SM invasion. However, the SRC grade of S1 were the independent protective factors of LVI; the UDC grade of U2/U3 and the SRC grade of S1 were the independent protective factors of SM invasion; in T1a patients, the gender of female was the independent protective factor of MM invasion ().

Table 3

Preoperative independent risk factors of tumor invasion, LVI and LNM

Clinicopathological characteristics	Factors	Variables	P value	HR	95% CI
					Lower	Upper
LNM	BMI	CV	<0.001	0.878	0.846	0.912
	Melena	+	<0.001	4.669	2.052	10.621
	Tumor size	CV	0.007	1.259	1.065	1.488
	Tumor location	Upper third	Ref.	1	–
		Middle third	0.759	1.137	0.502	0.759
		Lower third	0.012	2.666	1.238	0.012
	UL	+	0.020	1.970	1.114	3.483
LVI	BMI	CV	<0.001	0.882	0.853	0.911
	Melena	+	0.010	3.263	1.327	8.023
	Tumor size	CV	<0.001	1.488	1.220	1.816
	SRC	S1	0.001	0.247	0.108	0.565
SM invasion	Tumor size	CV	<0.001	1.389	1.197	1.613
	UL	+	<0.001	2.556	1.574	4.152
	UDC grade	U1	Ref.	1	–
		U2	0.033	0.424	0.193	0.033
		U3	<0.001	0.291	0.163	<0.001
	SRC grade	S1	0.030	0.565	0.336	0.948
MM invasion*	Sex	Female	<0.001	0.232	0.135	0.399

*, in T1a patients. The italic numbers are P values with statistical significance (<0.05). +, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion; LVI, lymphovascular invasion; SM, submucosa; MM, muscularis mucosa; UDC, undifferentiated component; SRC, signet ring cell.

Postoperative parameters

In univariate analysis, including the same demographic, clinical but histological features of surgical specimens: lower BMI, the presence of melena/UL, larger tumor size and the lower third stomach were the risk factors of LNM (); in multivariate analysis, all the risk factors mentioned above showed independent significance to LNM ().

Table 4

Postoperative independent risk factors of LNM

Condition	Factors	Variables	P value	HR	95% CI
					Lower	Upper
Without LVI & T data	BMI	CV	<0.001	0.878	0.846	0.912
	Melena	+	<0.001	4.669	2.052	10.621
	Tumor location	Upper third	Ref.	1	–
		Middle third	0.759	1.137	0.502	2.577
		Lower third	0.012	2.666	1.238	5.741
	Tumor size	CV	0.007	1.259	1.065	1.488
	UL	+	0.020	1.970	1.114	3.483
With LVI & T data	BMI	CV	<0.001	0.862	0.824	0.902
	Melena	+	0.001	4.244	1.751	10.283
	Tumor location	Upper third	Ref.	1	–
		Middle third	0.643	1.224	0.521	2.877
		Lower third	0.016	2.659	1.196	5.911
	SRC	S1	0.038	1.912	1.038	3.522
	Depth of invasion	LP	0.011	1	–
		MM	0.186	1.945	0.725	5.216
		SM	0.003	3.294	1.508	7.197
	LVI	+	0.001	3.832	1.694	8.667

The italic numbers are P values with statistical significance (<0.05). +, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion; LP, lamina propria; MM, muscularis mucosa; SM, submucosa; SRC, signet ring cell; LVI, lymphovascular invasion.

When depth of invasion and LVI were added into the covariates, lower BMI, the presence of melena/LVI, the lower third stomach and SM invasion were risk factors of LNM in univariate analysis (), and lower BMI, the presence of melena/LVI, the lower third stomach, the SRC grade of S1 and SM invasion showed independent significance in multivariate analysis (). The detailed data of the univariate analyses could be found in .

Table S5

Multivariate analysis for LNM with postoperative data (depth of invasion and LVI excluded)

Factors	Univariate analysis
	P	HR	95% CI
			Lower	Upper
Sex
Male	Ref.	1	–
Female	0.960	1.015	0.561	1.839
Age
CV	0.301	0.988	0.966	1.011
BMI
CV	0.008	0.909	0.848	0.976
Weight loss
−	Ref.	1	–
+	0.626	1.192	0.588	2.415
Hematemesis
−	Ref.	1	–
+	0.807	1.214	0.257	5.734
Melena
−	Ref.	1	–
+	0.001	4.815	1.900	12.199
Location
Upper third	Ref.	1	–
Middle third	0.702	1.180	0.506	2.752
Lower third	0.009	3.020	1.324	6.887
Size
CV	0.003	1.334	1.104	1.612
Gross pattern
IIa	Ref.	1	–
IIb	0.086	0.328	0.092	1.173
IIc	0.348	0.588	0.194	1.782
III	0.960	1.073	0.066	17.328
UL
−	Ref.	1	–
+	0.034	2.076	1.057	4.077
UDC grade
U1	Ref.	1	–
U2	0.994	.995	0.287	3.447
U3	0.965	1.028	0.303	3.484
SRC grade
S0	Ref.	1	–
S1	0.396	1.331	0.688	2.573

−, absence; +, presence. CV, continuous variable; Ref., reference; BMI, body mass index, UL, ulcerative lesion; UDC, undifferentiated component; SRC, signet ring cell.

Table S6

Multivariate analysis for LNM with postoperative data (containing depth of invasion and LVI)

Factors	Univariate analysis
	P	HR	95% CI
			Lower	Upper
Sex
Male	Ref.	1	–
Female	0.613	1.182	0.618	2.260
Age
CV	0.106	0.980	0.956	1.004
BMI
CV	0.003	0.889	0.822	0.962
Weight loss			–
−	Ref.	1	–
+	0.556	1.256	0.588	2.684
Hematemesis
−	Ref.	1	–
+	0.903	1.106	0.220	5.560
Melena
−	Ref.	1	–
+	0.007	4.043	1.455	11.237
Location
Upper third	Ref.	1	–
Middle third	0.631	1.247	0.506	3.075
Lower third	0.013	3.053	1.266	7.360
Size
CV	0.094	1.192	0.970	1.464
Gross pattern
IIa	Ref.	1	–
IIb	0.068	0.281	0.072	1.100
IIc	0.225	0.484	0.150	1.562
III	0.813	0.710	0.041	12.215
UL
−	Ref.	1	–
+	0.069	1.957	0.949	4.034
UDC grade
U1	Ref.	1	–
U2	0.813	1.165	0.327	4.149
U3	0.608	1.388	0.396	4.857
SRC grade
S0	Ref.	1	–
S1	0.110	1.801	0.876	3.702
LVI
−	Ref.	1	–
+	0.002	4.087	1.709	9.769
Depth of invasion
Lamina propria	Ref.	1	–
Muscularis mucosa	0.217	1.932	0.679	5.495
Submucosa	0.006	3.240	1.401	7.493

−, absence; +, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion; UDC, undifferentiated component; SRC, signet ring cell; LVI, lymphovascular invasion.

Comparison of the models with pre- or postoperative risk factors

The performance of the predictive models was evaluated by ROC curves. In those with the preoperative independent risk factors, the model for LNM () showed the second best performance, exceeded by that for LVI () and followed by those for SM () and MM () invasion. In models with the postoperative independent risk factors excluding depth of invasion and LVI, the model for LNM () showed the same performance as that with preoperative data, and when depth of invasion (), LVI (), or both of them () were added into the covariates, the performance of predictive models increased mildly.

Figure 3

The ROC curves of predictive models. (A) The ROC curve of the model for LNM with the preoperative independent risk factors; (B) the ROC curve of the model for LVI with the preoperative independent risk factors; (C) the ROC curve of the model for SM invasion with the preoperative independent risk factors; (D) the ROC curve of the model for MM invasion with the preoperative independent risk factors; (E) the ROC curve of the model for LNM with the postoperative independent risk factors screened out from the variates without depth of invasion and LVI; (F) the ROC curve of the model for LNM with the postoperative independent risk factors screened out from the variates containing depth of invasion; (G) the ROC curve of the model for LNM with the postoperative risk factors screened out from the variates containing LVI; (H) the ROC curve of the model for LNM with the postoperative risk factors screened out from the variates containing depth of invasion and LVI. AUC, areas under the ROC curve; LNM, lymph node metastasis; LVI, lymphovascular invasion; MM, muscularis mucosa.

Discussion

EGC-UDC is a special entity of EGC, of which biological behaviors were proved to be more malignant than that of differentiated EGC (19). To the best of our knowledge, this is the first study to explore the risk factors of tumor invasion and LNM by combing clinical data and the histology of biopsy specimens, revealing that the preoperative information was significantly associated with tumor invasion and/or LNM, the predictive models based on which showed comparable performance with those using postoperative data.

Due to the limited representativeness of biopsy specimens, the significance of their histological features to tumor behaviors attracted less attention in the past. Ryu (20) and Takao et al. (21) found a relatively high discrepancy rate between biopsy and surgical specimens, and believed that the diagnosis from biopsy specimens was easy to be affected by the accuracy of sampling. However, in Park’s (22) and Park (23) work, biopsy specimens showed an overall accuracy of 87.2% in the pathological diagnosis of total EGC, and a specificity of 100% in detecting EGC-UDC. In the present study, using our grading system, biopsy and surgical specimens showed a significant histological correlation, which provided us a foundation for evaluating the significance of biopsy specimens, and the same result of Logistic regression or the same performance of predictive models also demonstrated the comparable risk strength of biopsy and surgical specimens in histology.

In the present study, although two grading systems, UDC and SRC, were used to grab the histological features of biopsy specimens, no significant correlation was found between the histology and LNM. This was consistent with Kang’s finding (24) that the different histological types of biopsy specimens were not correlated with LNM. But in Shim’s work (25), moderately differentiated histology in biopsy specimens was reported to be relevant to LNM in “mixed histology predominantly of differentiated type EGC”. Besides, the irrelevance between histology and LNM was also observed in surgical specimens here, which differed from some previous findings. Imamura (16) and Kim (26) revealed that the mixed-SRC histological type were independent predictors of LNM in EGC-UDC, and the presence of SRC was an independent favorable prognostic factor. Horiuchi et al. (17) showed that the curative resection rate was significantly higher in pure SRCC than in mixed poorly differentiated GC. Jin et al. (27) also found that patients with SRCC had the lowest LNM rate than those with poorly differentiated or even moderately differentiated GC. Back to our study, the irrelevance between histology and LNM might be caused by adding clinical variables into the models, such as BMI and specific symptoms, which showed strong and persistent risk strength in the models for LNM, LVI, and/or SM invasion.

Nevertheless, it is worth noting that the presence of SRC in biopsy specimens significantly lower the risk of LVI in EGC-UDC independently. In numerous previous studies, as well as ours, LVI was closely correlated with node metastasis (19,28,29) or early distant metastasis (30), and functioned as an strong risk factor of LNM in EGC (odds ratio: 3.1–290.7) (10,16,24,26,31). However, rare studies had focused on the prediction of LVI with preoperative data in the past. As far as we have known, Kang et al. (24) were the first team to analyze the relationship between LVI and the histological types of biopsy specimens, and reported that, among poorly differentiated, SRCC (SRC >50%), poorly differentiated with SRCC (SRC ≤50%), mixed type, medullary type and poorly cohesive carcinoma (PCC), PCC and poorly differentiated/mixed differentiated histological type increased the risk of LVI independently. However, in our study, only the presence of SRC (grade S1) was identified as an independent protective factor of LVI. The difference between Kang and our results might be due to the different classifications of histological features: Kang et al. (24) defined a histological type of a tumor according to the tumor component exceeding 50%, while, in our study, UDC were graded in more detail and SRC was graded into two tiers (because of the relatively small sample size, the proportion of SRC was not subgrouped). The protective effect of SRC might be partially explained by the low microvessel density (MVD) in EGC with SRC, which was observed by imaging approach before (32). Intriguing, in the risk analysis using postoperative UDC and SRC grades, as well as the depth of invasion and LVI, the presence of SRC turned to a risk factor of LNM. Given the fact that more patients were found with SRC in surgical specimens than in biopsy specimens, SRC might function in a location-dependent manner, as only the adequate sampling of surgical specimens could reveal the SRC located in deep layers where SRC behaved differently from those in superficial layers (32). In addition, another explanation for the opposite roles of SRC in LVI and LNM might be the way SRC increased the risk of LNM, which might be not by increasing LVI.

Another dependent variable correlated with the histology of biopsy specimens was SM invasion, which is significant to the curability of ESD for tumors of expanded indication, including EGC-UDC (8). However, in practice nowadays, the depth of invasion is only concluded from endoscopic observation, CT images or lifting sign, causing SM invasion could be found in nearly 8% of ESD specimens (33). Our finding based on the histological features of biopsy specimens might provide a new direction for the evaluation of the depth of invasion, although, the performance of the predictive model was not excellent so far, and identifying more preoperative risk factors might be an effective way to improve the model performance in the future. Further, both UDC and SRC played protective roles in SM invasion here, combing the fact that the histological features of either biopsy or surgical specimens were not risk factors of LNM in our study, the traditional opinion on the equivalence between malignancy and undifferentiation was challenged. This situation pushes us to seek novel approaches to predict tumor behaviors with higher reliability. Integrated molecular classification of EGC might be a promising way, since some immunophenotypes (27), mi-RNA (34) and long noncoding RNAs (35), et al., rather than the morphological differentiation alone, had been found relevant to LNM in EGC.

The role of BMI in EGC seems to be controversial so far. It was demonstrated that high BMI (≥25 kg/m2) increased the risk of EGC or mucosal dysplasia (36), with insulin resistance playing an important role (37). But other studies showed that lower BMI was associated with metachronous GC (38) and worse survival of EGC patients (39). Our analysis identified BMI as a protective factor of both LNM and LVI in EGC-UDC, which was consistent with previous findings that BMI was inversely associated with tumor size, tumor depth, LNM and tumor stage in early or advanced GC (40,41). To explore whether the controversy mentioned above is due to patient grouping, we evaluated the role of BMI in underweight (<18.5 kg/m2), normal (18.5–24 kg/m2) or overweight (≥24 kg/m2) patients separately, and found that the protective effect were persistent in all the three groups (). These results indicated that the different influences of BMI might be decided by different phases (before or after the occurrence of EGC) rather than the different amount of lipid reserve, and once a tumor starts, with excessive energy consumption, sufficient energy storage might be essential for anti-tumor response and a better outcome.

Table S7

The significant risk factors of LNM in patients with different BMI

BMI (kg/m2)	Factors	Variables	P value	HR	95% CI
					Lower	Upper
<18.5	BMI	CV	0.026	0.938	0.886	0.992
18.5–24	BMI	CV	<0.001	0.832	0.760	0.911
	Melena	+	<0.001	8.635	3.217	23.173
	Location	Upper third	Ref.	1	–
		Middle third	0.531	1.396	0.491	3.968
		Lower third	0.004	4.591	1.625	12.971
	Gross pattern	IIa	Ref.	1	–
		IIb	0.538	1.807	0.275	11.872
		IIc	0.068	4.926	0.888	27.339
		III	0.042	24.918	1.124	552.444
≥24	BMI	CV	<0.001	0.866	0.812	0.924
	Size	CV	0.004	1.936	1.228	3.053
	UL	+	0.003	6.935	1.973	24.376

+, presence. CV, continuous variable; Ref., reference; BMI, body mass index; UL, ulcerative lesion.

To explore the significance of clinical manifestation, weight loss, melena and hematemesis were selected as representative symptoms in this study. To our surprise, the presence of melena was firstly found increasing the risk of LNM and LVI by 4.898 and 4.606 times in EGC-UDC respectively. It was found that overt bleeding (hematemesis, melena or hematochezia) was an index symptom of malignancy (42), and occurred in 17.4% of EGC (43). Some scholars tried to screen EGC by detecting upper gastrointestinal tract bleeding (44,45), but the relatively low specificity confines its application. However, in our study, the importance of melena opens a new gate for stomach bleeding detection, not only to screen EGC, but also to evaluate the risk of LNM and LVI. Further, as a risk factor of LVI and a parallel risk factor of LNM with LVI, melena might play a multifaced role in the progression of EGC-UDC. Firstly, it was relevant to the peri-tumoral vascular injury caused by tumor invasion in LVI; secondly, it might imply the nature of the tumors, since melena could be independent from LVI to increase the risk of LNM. In previous studies, the bleeding of GC was found closely associated with the vascular composition of tumors, according to which, GC could be categorized into hypo-vascular or hyper-vascular types (43), and the latter was proved to significantly correlates with LNM (46). So, given the parallel relationship between melena and LVI in the multivariate analysis for LNM, patients with melena might harbored hyper-vascular tumors, and the bleeding was not only due to the peri-tumoral vascular injury, but also related to the hypervascularity of these tumors. Thirdly, it is also possible that melena was an accompanied symptom which was not led by tumor invasion directly but induced through other unknown mechanisms, such as a weaker gastric mucosal barrier associated with some specific microbiota found in the past (47). However, more studies are needed to clarify the speculation above.

There were several limitations in our study. Firstly, due to the requirement of both surgical and matched preoperative biopsy data, although cases from three centers have been enrolled, the eligible cohort size was relatively small, which limited the statistical analysis, especially for EGC with different proportion of SRC. Secondly, the ability of the three symptoms selected to represent the clinical manifestation was relatively weak, and more attention should be paid to the importance of clinical data, since melena was identified as a strong independent risk factor of LNM and LVI. Thirdly, the role of gastric microbe was not involved, which is worth exploring in the future due to its role in tumorigenesis and progression of GC.

In summary, the preoperative data, including the histological features of biopsy specimens and clinical data were significantly associated with tumor invasion and/or LNM, which deserves more attention. The predictive models with pre- or postoperative histological features showed comparable performance. Further studies based on a larger sample size will be needed to verify or even expand the significance of preoperative risk factors and build more efficient models to identify EGC-UDC patients eligible for ESD treatment.