
The diagnosis rate of early gastric cancer has increased remarkably, and radical gastrectomy is the most effective treatment for gastric cancer [1]. Although there are patients with gastric cancer who complain of significant weight loss due to the gastric cancer before surgery, the anatomical changes of the stomach and small bowel after gastrectomy create a higher possibility of weight loss in the short- or long-term post-surgically [2-4]. While mild weight loss is associated only with decreased quality of life (QoL) [5,6], postoperative unintentional severe weight loss may lead to shorter survival [7-9].
As minimally invasive surgery, surgeons’ skill levels and pre- and post-operative nutritional supplementation management have improved, the rate of significant weight loss after surgery for gastric cancer is gradually decreasing. However, in 10%~20% of patients, significant unintentional weight loss after gastrectomy has been reported in recent studies [10,11]. The risk of unintended post-surgical weight loss is a common pre-surgical concern of gastric cancer patients, and gastric cancer surgeons have become increasingly aware of the importance of postoperative nutritional management.
Postoperative nutritional supplementation to support malnourished patients after gastrectomy has been a topic of research [12,13], and we sought to develop a tool to allow for screening of patients who may become malnourished after gastrectomy to ensure active nutritional support before and after surgery. We recently studied the pattern of body mass index (BMI) loss over time after gastrectomy using a group-based trajectory model and developed a prediction model for screening patients at risk of malnutrition six months post-gastrectomy [11,14]. During this process, we identified the need for a tool for predicting the extent of weight loss in each patient after gastrectomy. A few studies that identify risk factors for weight loss after gastrectomy in patients with gastric cancer have been conducted [15-17]. However, externally verified models for predicting the magnitude of weight loss at specific time points after gastrectomy based on these independent risk factors are lacking.
The purposes of this study were to develop a predictive model for percentage (%) weight loss after gastrectomy and to perform external validation of this model using multicenter clinical data. Through the external validation process, our goal was to develop a clinically applicable model.
This was a retrospective multicenter study in Korea. All participating institutions have received research approval for this study from their Institutional Review Boards: Seoul National University Hospital (SNUH, H-2101-007-1184), Chung-Ang University Hospital (2110-036-19390), The Catholic University of Korea St. Vincent’s Hospital (XC21RCDI0003V), Soonchunhyang University Bucheon Hospital (2021-02-027-001), Pusan National University Yangsan Hospital (05-2021-055), Dankook University Hospital (2020-12-026), CHA Bundang Medical Center (2020-12-052), Keimyung University Dongsan Medical Center (2021-01-053), Yonsei University Wonju Severance Christian Hospital (CR320169), The Catholic University of Korea Yeouido St. Mary’s Hospital (020-4467-0001), Pusan National University Hospital (2102-001-099), Kyung Hee University Hospital at Gangdong (2021-06-003), Gyeongsang National University Hospital (2021-01-015-001), and Ulsan University Hospital (2021-01-033). The requirement for informed consent was waived because of the retrospective study design. All procedures followed were in accordance with the institutional and national ethical standards of the responsible committee on human experimentation and with the Helsinki Declaration of 1964 and later versions.
Clinical data from patients with stage I~III gastric cancer who underwent curative gastrectomy (R0 resection with either D1+ or D2 lymph node dissection) between January 2014 and January 2019 at 14 different gastric cancer centers and who met study inclusion criteria were retrospectively collected. These inclusion criteria were patients with data on measurement of preoperative height and body weight and patients with more than one body weight measurement at six, 12, 24, and 36 months after gastrectomy (Fig. 1).
The same definition of malnutrition used in our previous study was applied in this study: BMI <18.5 kg/m2 [11]. This definition meets the criteria for being ‘at nutritional risk’ in the Malnutrition Universal Screening Tool [18] and the criteria for malnutrition in the European Society of Clinical Nutrition and Metabolism guidelines [19].
Our study group has used a group-based trajectory model to evaluate weight loss pattern after gastric cancer surgery [11]. The factors that were highly associated with post-gastrectomy weight loss in our previous study were the factors we primarily considered in the development of our prediction model. These factors were age, sex, preoperative BMI, preoperative nutrition state, cancer stage (I~III), surgical approach (open, laparoscopic or robotic), operation type (total gastrectomy [TG], proximal gastrectomy [PG], distal gastrectomy [DG], or pylorus-preserving gastrectomy [PPG]), reconstruction (Roux-en-Y esophago-jejunostomy, esophago-gastrostomy, double tract, Billroth I, Billroth II, Roux-en-Y gastro-jejunostomy or gastro-gastrostomy), use of adjuvant chemotherapy and postoperative complications (Clavien–Dindo grade). Based on clinician input, nutritional laboratory indicators, such as preoperative hemoglobin, protein, albumin, and cholesterol levels, were also included as predictor candidates.
Data from SNUH, a tertiary referral center in Korea, were used as a derivation set to develop the post-gastrectomy weight loss prediction model. The same SNUH data were used for internal validation (homogeneous setting) using tenfold cross-validation. The internally validated prediction model developed was externally validated with data collected from 13 different gastric cancer centers (a heterogeneous setting).
Characteristics of subjects are summarized and were compared using mean (standard deviation) and independent t-test for continuous variables and frequency (proportion) and chi-square test for categorical variables. A prediction model that accounts for individual patient % weight loss over time after gastrectomy was developed using a linear mixed effect model. The fixed effects were the predictors, time and interaction between time and predictors; the random effect was the subject. The estimated effect of each predictor was adjusted by time period after gastrectomy, and the statistically significant predictors were used for multivariable analysis. The prediction model was determined by the stepwise variable selection method. The adjusted mean difference per unit change in each predictor and its 95% confidence interval (CI) were calculated. The prediction model is presented using a prediction scoring table and nomogram. The prediction model was internally validated using tenfold cross-validation. The model performance was evaluated internally and externally using R2, the distribution of the difference between observed and predicted % weight loss and calibration slope. All statistical analyses were conducted by professional statisticians using SAS 9.4 (SAS Institute Inc., Cary, NC, USA).
Clinical data from 2,649 patients who met all inclusion criteria were collected from 14 participating institutions. Data from 1,420 SNUH patients (derivative set) were used for development and internal validation of the weight loss prediction model, and data from 1,229 patients collected from the remaining 13 hospitals (validation set) were used for external validation. Fig. 1 details the number of patients from each institution.
As shown in Table 1, the predominant characteristics at the baseline for each variable in both sets were similar; but there was a statistically significant difference in the proportion of some variables. In both sets, approximately 60-year-old male patients with stage I gastric cancer who underwent laparoscopic DG with Billroth I or II reconstruction were predominant. However, significant differences between the derivative set and validation set were demonstrated for these characteristics: age (60.1±11.3 years vs. 62.4±11.5 years, P<0.0001), sex (61.83% males vs. 70.63% males, P<0.0001), use of robotic surgery (8.03% vs. 3.91%, P<0.0001), use of PPG (18.38% vs. 1.14%, P<0.0001), use of Billroth I and gastro-gastrostomy reconstruction (Billroth I, 39.37% vs. 16.11%; gastro-gastrostomy, 18.38% vs. 1.14%; P<0.0001 for both) and complication rate (25.42% vs. 12.21%, P<0.0001). All preoperative nutritional laboratory indicators were in normal range in both sets.
In the derivation set (n=1,420), nine candidate predictors were sufficiently associated with % weight loss over time after gastrectomy: age, sex, preoperative BMI, presence of preoperative malnutrition, surgical approach, operation type, reconstruction procedure, cancer stage and use of adjuvant chemotherapy (Table 2). In the multivariable analysis by stepwise selection, postoperative duration (six months after gastrectomy vs. 12 months, 24 months or 36 months, P=0.0004); age (P=0.0053); sex (male vs. female, P<0.0001); preoperative BMI (P<0.0001); operation type (PPG vs. TG, PG, or DG, P<0.0001); and cancer stage (I vs. II or III, P=0.0159) were independently associated with postoperative weight loss. The prediction model for % weight loss at six, 12, 24, and 36 months after gastrectomy is shown in Table 3: weight loss (%) = –14.8+(0.412×12 months after gastrectomy)+(0.077×24 months after gastrectomy)+(–0.087×36 months after gastrectomy)+(0.04×age)+(2.069×female)+(0.782×preoperative BMI)+(5.275×TG)+(2.599×PG)+(0.755×DG)+(1.374×cancer stage II)+(0.663×cancer stage III).
A nomogram and scoring table were developed based on the equation (Fig. 2).
The developed prediction model for the % weight loss at six, 12, 24, and 36 months after gastrectomy was internally validated in the same derivation set using tenfold cross-validation (n=1,420): R2 (95% CI)=0.20 (0.20~0.21), 0.21 (0.20~0.21), 0.17 (0.17~0.17), and 0.18 (0.18~0.18), respectively, and calibration slope (95% CI)=0.95 (0.85~1.05) (Fig. 3A).
The model was also validated externally in the validation set (n=1,229): R2 (95% CI)=0.20 (0.20~0.20), 0.22 (0.22~0.23), 0.18 (0.18~0.18), and 0.18 (0.18~0.19), respectively, and calibration slope (95% CI)=1.00 (0.89~1.12). Even though the calibration slope and plot of external validation tended to slightly overestimate the % of weight loss after gastrectomy in the interval of 4%, 8%, and 13%, the slope showed acceptable calibration overall (Fig. 3B).
The area in which research on predictive models of weight loss has been most active is bariatric surgery [20-23]. Predicting postoperative weight loss is also an important subject in other fields of gastrointestinal surgery, but only a few relevant studies have been conducted [24,25]. Postoperative nutrition status is closely related to postoperative recovery and postoperative QoL improvement [5,16]. Because unintentional rapid weight loss after gastrectomy may lead to postoperative malnutrition [26,27], a tool to predict the extent of postoperative weight loss is necessary for both gastrointestinal surgeons and patients. In this study, we have developed a prediction model for % weight loss after gastrectomy and validated its accuracy and utility.
In our previous study, the BMI-loss trajectory model of patients with gastric cancer showed that severe BMI loss (approximately 21.5% reduction from preoperative BMI) at six months after gastrectomy was significantly associated with being elderly, being female, having higher preoperative BMI, being at an advanced cancer stage, having had open surgery, having had a TG performed, having had a Roux-en-Y reconstruction, having had chemotherapy and having had postoperative complications [11]. Among those factors, specific time point after gastrectomy, age, sex, preoperative BMI, surgical approach, and cancer stage were determined to be predictors of % weight loss after gastrectomy in this study. Older age, being female, having a higher preoperative BMI, having had a TG performed, being at cancer stage II and being at one year after gastrectomy increased % of post-gastrectomy weight loss significantly. Risk factors independently associated with weight loss after gastrectomy for gastric cancer in other studies were not entirely consistent with our results, but higher operative BMI and TG were risk factors identified in many studies [15,17].
Higher preoperative BMI was associated with greater postoperative BMI loss in the prediction model for % weight loss. Davis et al. [17] conducted a similar study in the Western population and reported that preoperative BMI and procedure type were the independent post-gastrectomy risk factors. In their study, even though patients with higher preoperative BMI had significantly greater weight loss than patients with lower preoperative BMI, the higher preoperative BMI patients maintained normal or overweight BMI throughout [17]. A similar Asian study also identified higher BMI as the most influential factor affecting postoperative weight loss. Moreover, the patients with baseline obesity (BMI >25 kg/m2, according to Asia BMI criteria) [28] exhibited the largest postoperative weight loss compared to normal BMI and underweight patients [15]. In our previous study, the BMI-loss trajectory model demonstrated a significant association between high preoperative BMI and severe weight loss after gastrectomy. However, low, not high, preoperative BMI was one of the risk factors of post-gastrectomy malnutrition [11,29]. As we discussed in the previous study, patients with high preoperative BMI may lose more weight; but those with lower preoperative BMI are at greater risk for malnutrition after gastrectomy.
Consistent with our previous BMI-loss trajectory model and other studies, TG surgeries were associated with greater weight loss than the other operation types (PPG, DG and PG) [25]. Davis et al. [17] also reported that patients who underwent TG lost significantly more weight across all time points than patients who underwent subtotal gastrectomy regardless of reconstruction procedure (P<0.01). One possible reason is the loss of storage volume after TG. In the patients who undergo DG, a portion of the stomach is retained and is usable as a reservoir; but patients who undergo TG have no remaining stomach. A second possible reason for the TG-associated body weight loss is the loss of ghrelin, which works as the promotion of the appetite signal in the hypothalamus and the stimulation of gastrointestinal activity. Previous studies demonstrated that persistent decline of serum ghrelin and body weight was observed commonly after TG, and short-term administration of synthetic ghrelin successfully lessened postoperative body weight loss after TG [30,31].
External validation procedures showed that our model was only able to predict weight loss at each follow-up mark with an accuracy of R2=0.20 at six months, 0.22 at 12 months, 0.18 at 24 months and 0.18 at 36 months (Fig. 3B). Postoperative weight loss prediction models have been developed for bariatric surgeries and have been able to predict weight loss with an accuracy of R2 ranging from 0.24 to 0.75 [20-23]. Even though the R2 values in these studies were higher than ours, R2 in these studies was still low. Livingston et al. [32] concluded that accurate prediction of weight loss depends on the initial fat and lean compartment mass since weight is lost from these at different rates and to different extents. Development of a predictive model with high accuracy is difficult even in weight loss predictive models after bariatric surgery calculated only using factors such as preoperative BMI, age, and sex. For patients with gastric cancer, development of a prediction model with high accuracy is more difficult because there are more factors, such as cancer stage, operation type, reconstruction method and use of chemotherapy for which to account.
This is a limitation of our study. Ideally, in order to develop a prediction model, data representative of all gastric cancer patients, a development set, should be used. However, predictive models are usually developed using data from one institution or one country; and we have also developed the predictive models using data from a single institution. Therefore, in our prediction model, there were some differences in baseline characteristics of the derivative and validation sets. SNUH used robotic surgery, PPG, Billroth I and gastro-gastrostomy reconstruction more frequently and had a higher complication rate than the other institutions. These differences may be due to the location of the center, case volume, surgeons’ technical preference and available surgical equipment. Differences in the complication rates may have been affected because each institution had different methods of collecting data on complications. However, there was no “within 30-day postoperative mortality” in the entire cohort. To overcome this limitation, we evaluated the predictive power of the model through external validation in a slightly heterogeneous group of gastric cancer patients. Another limitation is that we could not include other variables which may affect weight loss after gastrectomy such as modification in physical activity or dietary behavior, psychological profiles, genetic background and support group participation. Because our research design was that of a multi-center and retrospective study, missing data was an issue; collecting and standardizing such data from all participating centers was difficult. These variables should be included in a future prospective multicenter study.
While we developed a model for predicting weight loss after gastrectomy, its prediction accuracy was only about 20%. This did not meet our expectations. However, since similar predicted values were shown in internal and external validation, we expect that the model can be used as a reliable reference material in actual clinical practice. The % weight loss prediction model may be helpful to facilitate appropriate preoperative consultation with patients regarding individualized postoperative weight changes, nutritional supplementation strategies and follow-up.
Conceptualization: JHP, HJL. Data curation: JHP, JWK, KBP, IC, SHH, DWK, SMK, SWR, SCG, PYJ, HR, SGK, CIC, DHK, SIC, JHP, DJP, GYK, SHK, DJP, HKY, YC, HJL. Formal analysis: JHP, YC, HJL. Funding acquisition: JHP, HJL. Investigation: JHP, HJL. Methodology: JHP, YC, HJL. Project administration: JHP, HJL. Resources: JHP, JWK, KBP, IC, SHH, DWK, SMK, SWR, SCG, PYJ, HR, SGK, CIC, DHK, SIC, JHP, DJP, GYK, SHK, DJP, HKY, YC, HJL. Software: JHP, YC, HJL. Supervision: YC, HJL. Validation: JHP, YC, HJL. Visualization: JHP, YC. Writing – original draft: JHP. Writing – review & editing: JHP, JWK, KBP, IC, SHH, DWK, SMK, SWR, SCG, PYJ, HR, SGK, CIC, DHK, SIC, JHP, DJP, GYK, SHK, DJP, HKY, YC, HJL.
The authors of this manuscript have no conflicts of interest to disclose.
This study was supported by Seoul National University College of Medicine and by a research grant (no. 0431-20200021) from Korean Society of Surgical Metabolism and Nutrition.
![]() |
![]() |