Effect of comorbidity classes on survival of patients with gastrointestinal tract cancer

Abstract

Background

Comorbidities may complicate medical situations and have an impact on the treatment decisions and poor survival of cancer patients. How comorbidities cluster together and ultimately affect patients’ outcomes in gastrointestinal tract cancer (GTC) is a poorly understood area.

Methods

In a multicenter prospective observational study from 2012 to 2021, we grouped the comorbidities of patients with GTC by latent class analysis, obtaining two comorbidity classes. Cox regression models were initially used to predict mortality. LASSO techniques were used to reduce the dimension. The final model included the comorbidity classes and nine more predictors. Additionally, the performance of different simple multimorbidity measures were compared using the Bayesian information criterion (BIC), ROC curves and C-index. Finally, the performance of the final model was analyzed using ROC curves, calibration curves and decision curves. The nomogram was drawn to evaluate the model.

Results

We included 10,019 patients and obtained two comorbidity classes. Class 2 patients have a higher incidence of comorbidities, and a lower survival rate compared to Class 1 (P < 0.001). Compared to models containing the number of comorbidities or only a single comorbidity, the final model with the comorbidity classes has the highest AUC and C-index, as well as the lowest BIC, indicating this model has the best predictive performance.

Conclusion

We identified two classes of comorbidities that were associated with overall survival in patients with GTC. The combination of different comorbidities class plays a vital role in the prognosis of GTC.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-13517-1.

Background

Malignant gastrointestinal tract cancer (GTC) are significant diseases that seriously threaten human health. Esophageal cancer (EC), gastric cancer (GC) and colorectal cancer (CRC) ranked 10th, 6th, and 5th in the global cancer incidence in 2020, with the total number of new cases and deaths accounting for 29% and 33% of the top 10 malignancies [1]. The overall prognosis of patients with these cancers has improved in recent decades due to surgical and medical progress, but overall survival remains poor especially in EC and GC. The 5-year survival rates for EC, GC, CRC were 20% [2], 30% [3] and 65% [2].

The prognostic factors for GTC have been extensively reported. A large number of studies have focused on the impact of a single comorbidity [4] or the numbers of comorbidities [5–7]. It shows that comorbidities are important prognosis factors to gastrointestinal cancer patients [8]. A retrospective study found that almost half of colon cancer patients have three or more comorbidities [8], and those with a higher comorbidity burden are less likely to receive or complete standard anticancer treatments and have shorter overall survival [9–14]. However, the effect of comorbidity may be underestimated if we use simple measures. The pathological mechanisms of different comorbidities will have different effects on the prognosis. Accurately assessing the impact of comorbidities on prognosis should be based on a deep understanding of the types, the numbers, the combinations and interactions of different comorbidities and cancers. Therefore, the comprehensive assessment of comorbidities should be considered in clinical treatment decisions and patients’ quality of life.

The objectives of this study were to identify distinct comorbidity classes and their effect on overall survival in patients from a multi-center, large-sample study, based on the Investigation on Nutrition Status and its Clinical Outcome of Common Cancers (INSCOC) data [15]. We then constructed a clinical prognostic model including comorbidity classes and validated its predictive performance in patients’ survivals.

Materials and methods

Study population

Data were collected from the INSCOC project, spanning the period from January 2012 to May 2021. This project is a multi-center, large-sample, and prospective study conducted in China. The study protocol has been registered with the Chinese Clinical Trial Registry (ChiCTR), with the identifier ChiCTR1800020329 [16]. The inclusion criteria and exclusion criteria were seen in Extension protocol for the INSCOC study. This study ultimately included 10,019 patients from 45 hospitals in 20 provinces of China (Supplementary Table 1). The study design was approved by the local ethics committees of each participating hospital, and the project was conducted in accordance with the updated Declaration of Helsinki. Before participating in the study, all recruited patients signed a written agreement form authorizing the use of their data.

Data acquisition

The baseline information for all cancer patients was collected by one or two well-trained professional researchers at each institution. The information included demographic data, clinical data (cancer type, histological type, cancer stage, number of metastatic organs), laboratory data (including four composite inflammatory markers), Patient interview data (nutrition risk screening 2002 (NRS2002), Karnofsky Performance Status (KPS)), and treatment data (surgery, radiotherapy, chemotherapy, nutrition support). The response variable were the survival status and time of GTC patients.

Follow-up and quality control

Follow-up was performed by each center at the 6th month after enrolling, as well as every year thereafter. All follow-up were carried out by professionally trained staff. The final follow-up was conducted in June 2021. The primary endpoint was the all-cause mortality from enrollment, confirmed by the patients’ family, hospital medical records, and healthcare professionals.

Follow-up and data quality were regularly checked and assessed at multiple centers as part of quality control procedure. Quality control included timely reporting of follow-up progress and regular assessment of follow-up staff. Data quality control included checking for data entry errors, omissions, and outliers, and promptly correcting corresponding any errors.

Statistical analysis

Latent Class Analysis (LCA) is a person-centered analytical approach that partitions a population into subgroups or classes based on a set of observed variables [17]. In this study, we utilized LCA to identify distinct comorbidity classes among patients (Supplementary file 1). The process began by specifying different numbers of classes—ranging from two to a maximum number (designated by researchers as an integer based on the data and research questions). To determine the optimal number of classes, we relied on several model selection criteria. The Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were used, as they penalize model complexity to varying degrees, with lower values suggesting a better balance between model fit and complexity, which helped in avoiding overfitting by preferring simpler models [18]. The Lo-Mendell-Rubin adjusted likelihood ratio test (LMR) was used to compare models with successive numbers of classes. For example, if researchers set the number of classes to four, the LMR assesses whether a model with four classes significantly improves the model fit over a model with three classes. Entropy, a measure of classification precision, was also calculated for each model. High entropy values — ideally above 0.8 — indicate that the classification is reliable, with over 90% of patients being correctly classified into their respective comorbidity classes [19]. Relative Frequency (RF) denotes the proportion of individuals within each latent class relative to the total sample. When the RF of a smallest class falls below 5%, it may suggest that the class is under-represented within the sample, prompting a reevaluation of the model’s rationality [18]. These methods enable the translation of complex comorbidity profiles into comorbidity classes.

The participants were randomly assigned to the training and validation sets. Initially, the Least Absolute Shrinkage and Selection Operator (LASSO) method with 10-fold crossover was used to identify potential predictive factors and construct a final model, a method known for its ability to shrink less important variable coefficients to zero, effectively performing variable selection and enhancing the model’s interpretability. Cox univariate regression analysis was conducted on all variables, and Cox multivariate regression analysis was conducted on the predictive factors selected by LASSO. Secondly, the Comorbidity class variable in the final model was replaced by the number of comorbidities or a single comorbidity to construct four new models (the three most serious common comorbidities that might have a substantive effect of survival were selected for analysis: coronary heart disease (CHD), COPD and diabetes). The methods of model comparison included the use of area under the curve (AUC) of the ROC curve (Receiver Operating Characteristic), BIC and concordance index (C-index). Thirdly, in the final model, a nomogram was constructed in order to predict 1-, 3- and 5-year overall survival, decision curve was generated in order to assess the net benefit of the model, the calibration curve was constructed to evaluate the accuracy of the probabilistic output of the prediction model.

Missing data represent less than 5% of the observations in the full dataset and were imputed using multiple interpolations. For all analysis, p ≤ 0.05 was considered statistically significant. Categorical data were presented as frequencies and proportions. Continuous variable is described using the median, along with the 25th percentile and 75th percentile. The survival time of patients is described using the mean ± standard deviation (SD). Chi-square test and Fisher exact test were used to compare the rates between the two groups. The Kaplan-Meier (K-M) survival curve was used to describe the survival of patients in each group. LCA models were estimated in Mplus Version 8.3 (Muthén & Muthén, 2020) [20]. Other statistical analysis performed were conducted using SPSS 29.0.1 (SPSS Software, Chicago, USA) [21] and R 4.3.2 (R Foundation for Statistical Computing, Vienna, Austria) [22]. The R package used in this study includes glmnet [23], rms [24], VIM [25], survival [26], dcurves [27], autoReg [28], ggplot2 [29], timeROC [30].

Results

Baseline characteristics

A total of 10,019 hospitalized GTC patients (EC n = 2156; GC n = 3057; CRC n = 4806) were included in the analysis (Supplementary Table 2). The average follow-up of patients with esophageal cancer, gastric cancer and colorectal cancer was 31.3 ± 23.7 months, 38.4 ± 24.4 months and 39.0 ± 24.3 months, respectively. The average follow-up time of the total population was 37.2 ± 24.4 months (median 35 months), the longest follow-up was 97 months, and the most times of follow-up visits was seven. There were 2950 (29.4%) patients died during follow-up. The median age at diagnosis was 59 years, and 52.6% were diagnosed at < 60 years; 68.1% were male, and 48.8% underwent surgery. The 1-, 3-, 5-year survival rates of the overall population were 87.4%, 74.7%, and 65.7%.

Comorbidity classes

All models had Entropy > 0.8 and LMR p < 0.001, indicating that the classification accuracy of the models was greater than 90%. Then, we assessed the relative RF of the smallest class for each model and discovered that it was too small in 3-class model (RF = 0.0096) and 4-class model (RF = 0.0026), as detailed in Table 1. It meant that the population was too small to be representative, so we chose the 2-class model.

Table 1.

Latent class Model Fit comparison (N = 10019)

	AIC	BIC	Entropy	LMR p value	RF
1-class model	34,370	34,471	-	-	-
2-class model	32,120	32,329	0.811	< 0.001	0.0868
3-class model	31,866	32,183	0.842	< 0.001	0.0096
4-class model	31,707	32,133	0.869	< 0.001	0.0026

AIC Akaike information criterion, BIC Bayesian information criterion, LMR Lo-Mendell-Rubin adjusted likelihood ratio test, RF Relative frequency for smallest class

The median age was not different between the two classes. Class 1 included relatively healthy patients or with fewer comorbidities, accounting for 91.3% of the total population. Among Class 1 patients, 75.2% had no comorbidities, 23.1% had only one comorbidity, and 1.7% had two comorbidities. None of the patients in this group had stroke, myocardial infarction, chronic pancreatitis, osteoporosis, or renal disease. Class 2, although accounting for 8.7% of the total population (870 individuals), was mainly characterized by having multiple comorbidities (Supplementary Table 3 and Supplementary Table 4). Among Class 2 patients, 3.0% had only one comorbidity, 53.7% had two comorbidities, and 22.6% had three comorbidities, and 20.7% had four or more comorbidities. In addition, except for Tuberculosis and Hepatic disease, which were not discriminatory between the two categories, the incidence of comorbidities was much higher in Class 2 than in Class 1. Of these, the incidence of inflammatory bowel disease in Class 2 was 5.2 times higher than that in Class 1, COPD was 8.6 times, diabetes was 10.9 times, and coronary heart disease (CHD) was 23.8 times. KM survival curves showed that the survival of patients in Class 2 was significantly shorter than those in Class 1 (P < 0.001) (Supplementary Fig. 1). The 1-, 3-, 5-year survival rates in Class 1 were 87.6%, 75.0%, 66.6%, and in Class 2 were 85.1%, 71.3%, 55.7%.

To explore the potential differences between comorbidity classes, we conducted univariate analyses targeting other factors (Supplementary Table 5). We found that Class 1 had a significantly higher proportion of males, smokers, and alcohol drinkers, as well as patients with only a compulsory education, with a normal weight, without a family history, and residing in rural areas, compared to Class 2 (all P < 0.001). Class 1 tended to include younger patients, with more than half of the patients being under 60 years old, and only 13.9% were 70 years old or older. In contrast, Class 2 had 33.3% of patients under 60 years old, and a notably higher proportion of patients aged 70 years and older, accounting for 27.9% of the class. And the proportion of retired patients in Class 2 (45.7%) was significantly higher than that in Class 1 (25.6%), which could be explained by the higher proportion of older patients in Class 2. In addition, we found no significant difference in cancer staging between Class 1 and Class 2. And there were more patients with colorectal cancer in Class 2 (P = 0.002).

Training set and validation set

The proportions of training and validation sets were 70.0% (n = 7013) and 30.0% (n = 3006) in the patients. The mean follow-up for the training and validation sets were 37.4 ± 24.3 months and 36.5 ± 24.4 months. There were no significant differences in most baseline characteristics between two cohorts (P > 0.05), suggesting an equilibrium distribution for the two cohorts (Supplementary Table 2). The KM survival curves showed no significant difference in overall survival between the training and validation sets (P = 0.19) (Supplementary Fig. 2), indicating the even distribution between two sets and suitable for internal verification. In the training set, the 1-, 3- and 5-year survival rates for Class 1 were 87.7%, 75.4% and 67.1%, and for Class 2 were 85.3%, 71.4% and 55.3%. In the validation set, the 1-, 3- and 5-year survival rates for Class 1 were 87.4%, 74.1% and 65.4%, and for Class 2 were 84.6%, 71.2% and 56.8%.

Variable selection

The Lasso method (Supplementary Fig. 3 and Supplementary Fig. 4) was initially employed to reduce the number of candidate variables, resulting in the selection of 10 from the original 25, including age, smoking, NRS2002, KPS, stage, cancer type, histological type, the number of metastatic organs, surgery, and comorbidity classes. The results of the Cox regression analysis of these factors, as illustrated in Table 2, indicated that comorbidity class was an independent risk factor for predicting patient outcomes.

Table 2.

Univariate and multivariate analysis

Variable	Value	Univariate		multivariate
		HR (95% CI)	P	HR (95% CI)	P
Sex	Female	0.824 (0.749, 0.907)	< 0.001
Age	50–60	1.206 (1.056, 1.377)	0.006	1.168 (1.020, 1.336)	0.024
Age	60–70	1.411 (1.242, 1.603)	< 0.001	1.302 (1.142, 1.484)	< 0.001
Age	≥ 70	1.917 (1.665, 2.207)	< 0.001	1.563 (1.349, 1.810)	< 0.001
Smoking	Ever	1.235 (1.132, 1.347)	< 0.001	1.140 (1.042, 1.248)	0.004
Drinking	Ever	1.098 (0.996, 1.212)	0.060
FH	Yes	1.001 (0.884, 1.133)	0.988
Place	Rural areas	1.023 (0.938, 1.116)	0.604
Education	Above compulsory education	0.878 (0.800, 0.963)	0.006
Occupation	Physical work	0.899 (0.808, 1.001)	0.052
Occupation	Mental work	0.774 (0.677, 0.883)	< 0.001
Occupation	Other work	1.005 (0.882, 1.146)	0.936
BMI	18.5–24	0.680 (0.611, 0.757)	< 0.001
BMI	≥ 24	0.618 (0.542, 0.705)	< 0.001
NRS2002		1.172 (1.142, 1.202)	< 0.001	1.111 (1.081, 1.141)	< 0.001
KPS		0.987 (0.985, 0.990)	< 0.001	0.990 (0.987, 0.993)	< 0.001
NLR		1.001 (0.999, 1.004)	0.234
PLR		1.000 (1.000, 1.000)	0.508
SII		1.000 (1.000, 1.000)	0.046
PNI		0.997 (0.995, 1.000)	0.020
Hospital stays		1.002 (0.999, 1.006)	0.190
Stage	Stage III + Stage IV	2.030 (1.798, 2.292)	< 0.001	1.590 (1.404, 1.801)	< 0.001
Cancer type	GC	0.761 (0.680, 0.850)	< 0.001	0.582 (0.448, 0.756)	< 0.001
Cancer type	CRC	0.535 (0.479, 0.596)	< 0.001	0.432 (0.331, 0.563)	< 0.001
Histological type	Squamous cell carcinoma	1.546 (1.401, 1.708)	< 0.001	0.752 (0.580, 0.974)	0.031
Histological type	Signet-ring cell carcinoma	1.756 (1.412, 2.183)	< 0.001	1.720 (1.374, 2.153)	< 0.001
MON	1	2.149 (1.920, 2.406)	< 0.001	2.074 (1.849, 2.327)	< 0.001
MON	≥ 2	3.194 (2.834, 3.599)	< 0.001	2.956 (2.615, 3.342)	< 0.001
Surgery	Yes	0.766 (0.702, 0.836)	< 0.001	0.885 (0.808, 0.968)	0.008
Radiotherapy	Yes	1.124 (0.978, 1.290)	0.099
Chemotherapy	Yes	1.298 (1.190, 1.416)	< 0.001
Nutrition support	Yes	0.986 (0.897, 1.084)	0.770
Comorbidity class	Class 2	1.376 (1.200, 1.578)	< 0.001	1.324 (1.151, 1.523)	< 0.001

NRS 2002 nutrition risk screening 2002; KPS Karnofsky Performance Scale; NLR neutrophil-to-lymphocyte ratio; PNI the prognostic nutritional index; MON Metastatic organ number; EC Esophageal cancer; GC Gastric cancer; CRC Colorectal cancer; AC adenocarcinoma; SCC Squamous cell carcinoma; SRCC Signet-ring cell carcinoma

In the multivariate analysis, age, smoking status, NRS2002, KPS, cancer stage, cancer type, histological type, metastatic organ number, surgery status, and comorbidity class were found to be significant predictors for patient outcomes. The adjusted hazard ratios (HR) and their 95% confidence intervals (CI) are presented, with P-values indicating the level of statistical significance. A P-value of less than 0.05 was considered statistically significant

Comparison of comorbidities measures

Whether in the training or validation set, the final model has the highest AUC of the ROC curve, (Fig. 1), the lowest BIC, and the highest C-index (Table 3). These results suggest that the comorbidity classes offer a better performance than the basic measures.

Fig. 1.

ROC for training set (A) and validation set (B)

Table 3.

Comparison of different comorbidity measures

	Comorbidity classes model	Comorbidity count model	CHD model	COPD model	Diabetes model
Train set
AUC-1year	0.746 (0.728–0.763)	0.622 (0.602–0.643)	0.621 (0.601–0.642)	0.621 (0.601–0.642)	0.624 (0.604–0.644)
AUC-3year	0.732 (0.717–0.747)	0.601 (0.584–0.618)	0.598 (0.580–0.615)	0.598 (0.580–0.615)	0.601 (0.584–0.618)
AUC-5year	0.720 (0.703–0.738)	0.628 (0.609–0.646)	0.619 (0.600–0.638)	0.616 (0.597–0.635)	0.622 (0.603–0.641)
BIC	33,359	33,935	33,936	33,943	33,928
C-index	0.704	0.601	0.598	0.597	0.599
Validation set
AUC-1year	0.749 (0.724–0.774)	0.630 (0.599–0.661)	0.622 (0.591–0.654)	0.622 (0.591–0.654)	0.629 (0.597–0.660)
AUC-3year	0.740 (0.717–0.762)	0.614 (0.589–0.640)	0.603 (0.578–0.629)	0.606 (0.580–0.632)	0.612 (0.586–0.638)
AUC-5year	0.722 (0.695–0.748)	0.626 (0.596–0.655)	0.619 (0.589–0.648)	0.620 (0.590–0.649)	0.621 (0.592–0.650)

AUC area under the curve; BIC bayesian information criterion; C-index concordance index; ROC Receiver Operating Characteristic

We have renamed the final model with the comorbidity classes as the ‘Comorbidity classes model’ in this table to better distinguish it from the other four models

Charasteristics of final model with the comorbidity classes

The nomogram (Fig. 2) provides an integrated perspective to evaluate the impact of different factors on the survival of cancer patients, helping to guide clinical decision making and patient management. The points on the calibration plot generally fall along the 45-degree line, indicating that the predicted probability of survival closely matches the actual probability of survival, which suggests that the model predicts correctly (Fig. 3. A-C). The decision curve (Fig. 3. D-F) showed different net benefits of the model at different threshold probabilities at 1, 3 and 5 years, which helped clinicians to make decisions accordingly.

Fig. 2.

Nomogram based on final model with the comorbidity classes Note: The steps on how to read and calculate the total integral of a nomogram: 1) Determine the specific value of each prognostic factor 2) Draw a vertical line for each prognostic factor, find the corresponding score on the Points line at the top of the Fig. 3) Add all corresponding scores to obtain the total score 4) The Total Points line at the bottom of the nomogram represents the relationship between total points and survival probability. Map the total score to this line, and obtain predicted survival probabilities for time points such as 1 year, 3 years, and 5 years

Fig. 3.

Calibration curves for 1-year(A), 3-year(B), 5-year(C); Decision curves for 1-year(D), 3-year(E), 5-year(F)

Discussion

Of the 4.8 million new cases of gastrointestinal cancers and 3.4 million related deaths worldwide in 2018, it is estimated that 38% of the cases and 41% of the deaths occurred in China [31]. Complications are important factors affecting the prognosis of patients. Numerous studies have confirmed the effect of a single comorbidity on gastrointestinal cancers [32–35]. However, it is a critically understudied area to determine how comorbidities cluster together and affect prognosis.

In this study, we first used the LCA method and its parameters (AIC, BIC, Entropy, LMR p value, RF) to group the comorbidities of EC, GC, and CRC patients and identified two classes of comorbidities. Among them, Class 2 patients have a high incidence of comorbidities and a short survival period. Secondly, we compared the models containing the Comorbidity class with those containing only a single comorbidity or the number of comorbidities, demonstrating that the combination of comorbidities has a significant impact on the prognosis of gastrointestinal cancer patients. Therefore, our study emphasizes that when evaluating the impact of comorbidities on patient prognosis, individual comorbidities should not be examined in isolation, but their combined effects on the overall health status and survival of patients should be comprehensively considered.

Among the various comorbidities we collected, COPD, diabetes and CHD are dominant. Numerous findings have demonstrated an increased risk of death due to COPD in cancer patients and highlighted the importance of respiratory care throughout survival [36–38]. However, the relationship between cancer and COPD mortality remains unclear [39]. Studies have reported that risks of death and progression were greater for patients with diabetes than those without diabetes in gastrointestinal cancer patients [33, 40]. A consensus report from the American Diabetes Association and American Cancer Society points that the association between diabetes and cancer is not well established. Although a number of potential mechanisms have been proposed, such as insulin resistance and hyperglycaemia cancer promotion. The biological mechanisms by which diabetes increases cancer risk remain undefined [34]. The impact of CHD on the prognosis of patients with gastrointestinal cancers may come from obesity [41]. Obesity is a significant risk factor for CHD. The major systematic changes associated with obesity include the alteration of insulin, leptin, cytokines, and steroid hormones. These factors alter the nutritional environment, which may create conditions that trigger and sustain tumor growth [42].

The other predictive factors included in our model have been supported by relevant literature in their prognostic importance. In the overall population, signet ring cell carcinoma (SRCC) had the highest risk, which is consistent with previous studies. Signet ring cell carcinoma is a histological type based on microscopic characteristics that occurs in gastrointestinal cancer. In early-stage cancer, the prognosis of SRCC is similar [32, 35] or better than other adenocarcinoma [43, 44]. Among advanced gastric cancer patients, SRC patients had a worse prognosis than other cell types [45–47]. In our study, SRCC was an independent predictor of poor prognosis, consistent with the study by Guillaume Piessen et al. [48].

The final model with the comorbidity classes did not include sex, which was traditionally considered important. However, we found sex differences in different classes, so we speculate that when using LCA method to construct the comorbidity class variable, the potential impact of sex was already included. Therefore, in Lasso screening, comorbidity class was selected and sex was removed.

Our study has certain limitations. First, This study was a retrospective analysis of prospectively collected trial data, which may be subject to selection bias, as well as information bias caused by incomplete or inaccurate data recording. Second, the study may not cover all types of comorbidities, and we were only aware if a patient had a certain type of comorbidity, without knowing the severity of the comorbidity (such as stage 1 hypertension or stage 2 hypertension, well-controlled diabetes or poorly-controlled diabetes). If we can collect more detailed information about comorbidities in the future, it will help improve the accuracy of the model. Third, although there is no external validation, this study still has specific value based on multi-center cooperation. This study included patient from 45 hospitals in 20 provinces, which is representative in exploring the clustering rules of comorbidities and fitting high-performance prognosis prediction models. In the future, to enhance the reliability and reproducibility of our research, we plan to actively pursue external validation and conduct further validation studies using public datasets. Additionally, we aim to develop a user-friendly web application that will allow doctors to quickly determine the classification of patients’ comorbidities by simply inputting relevant information, thereby providing more effective support for clinical practice.

Conclusion

We identified two classes of comorbidities in patients with GTC that were associated with overall survival: the first class with no or few comorbidities, and the second class with a high presence and combination of comorbidities. We found that the combination of different comorbidities plays a vital role in the prognosis of these cancers.

Abbreviations

EC

Esophagus cancer

GC

Gastric cancer

CRC

Colorectal cancer

INSCOC

The Investigation on Nutritional Status and its Clinical Outcomes of Common Cancers in China

ICD-10

The International Classification of Diseases, Tenth Revision

NRS2002

Nutritional risk screening 2002

KPS

Karnofsky Performance Scale

BMI

Body Mass Index

NLR

The neutrophil-lymphocyte ratio

PLR

The platelet-lymphocyte ratio

SII

The systemic immunoinflammatory index

PNI

The prognostic nutritional index

LCA

Latent Class Analysis

AIC

Akaike information criterion

LMRT

Lo-Mendell-Rubin adjusted likelihood ratio test

RF

Relative frequency for smallest class

LASSO

Least Absolute Shrinkage and Selection Operator

ROC

Receiver operating characteristic curve

Authors’ contributions

LG: conceptualization, methodology, writing (original draft preparation, review & editing); TY, SG: conceptualization, supervision, writing (review & editing); JC, WL, ZG, HX, MW, SL, QY, WH, LZ, JChen, XW, QZ, HL: data curation; HS, YB, HH: conceptualization, supervision, methodology, writing (review & editing).All authors read and approved the final manuscript.

Funding

This work was supported by the National Key Research and Development Program (2022YFC2009600, 2022YFC2009601), the Shanxi Provincial Basic Research Program (202303021211202), and Doctoral Startup Fund by Shanxi Medical University (XD2104).

Data availability

The datasets used and analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The data was extracted from the Investigation on Nutritional Status and its Clinical Outcomes of Common Cancers (INSCOC) project in China (chictr.org.cn: ChiCTR1800020329). The study design was approved by the local ethics committees of each participating hospital, and the project was conducted in accordance with the updated Declaration of Helsinki. Before participating in the study, all recruited patients signed a written agreement form authorizing the use of their data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.