An ensemble machine learning approach for predicting colorectal polyps using novel inflammatory markers

Abstract

Objectives

This study aimed to investigate the associations of two novel inflammatory-nutritional markers, the C-reactive Protein-Albumin-Lymphocyte (CALLY) index and C-reactive protein-to-lymphocyte ratio (CLR), with colorectal polyps, and to develop machine learning models for improved risk prediction.

Methods

The study conducted a retrospective analysis of 946 patients who underwent colonoscopy at the Second Qilu Hospital of Shandong University from 2023 to 2024. Clinical, laboratory, and endoscopic data were retrieved. Four machine learning models were developed: Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). Additionally, a Stacking ensemble model was constructed, integrating RF, XGBoost, and SVM with logistic regression as the meta-learner. Model performance was evaluated using metrics including the Area Under the Curve (AUC), accuracy, precision, recall, F1-score, and Kappa coefficient.

Results

Both the CALLY index and CLR were significantly associated with the presence of colorectal polyps. The Stacking ensemble model achieved the highest predictive performance (AUC: 0.96), outperforming all individual base models. The RF, SVM, and XGBoost models also attained robust predictive capability (AUC: 0.75–0.76). By comparison, LightGBM exhibited suboptimal performance (AUC = 0.72). Feature importance analysis revealed CA19-9 as the top predictive factor.

Conclusion

Elevated CLR and reduced CALLY index, together with traditional risk factors, synergistically contribute to colorectal polyp formation. The Stacking ensemble model serves as a highly accurate tool for identifying individuals at high risk of colorectal polyps, thereby facilitating targeted screening and timely early intervention. These findings support the incorporation of multi-dimensional inflammatory-nutritional biomarkers and machine learning strategies into clinical practice to optimize Colorectal Cancer (CRC) prevention.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12876-026-05106-8.

Introduction

CRC ranks as the third most commonly diagnosed and second most lethal malignancy globally. In 2020, it accounted for approximately 1.9 million new cases and nearly 1 million deaths worldwide [1]. China bears the greatest global burden of CRC, accounting for more new cases and deaths than any other country [2]. Colorectal polyps are well-recognized precursor lesions of CRC, with over 90% of CRC cases known to develop progressively from these lesions; the progression period from adenomatous polyps to invasive carcinoma ranges from 5 to 15 years. Early detection and intervention during this premalignant window can reduce CRC incidence by more than 50% [3, 4]. Accordingly, the accurate identification of individuals at high risk of colorectal polyps constitutes a critical component of mitigating the global disease burden of CRC.

The pathogenesis of colorectal polyps is a complex, multifactorial, and multistage process. Beyond traditional risk factors—including age, gender, dietary patterns, and genetic predisposition [5, 6]—the role of chronic inflammation in driving polyp development has been well documented [7]. Chronic inflammation facilitates polyp initiation and malignant transformation through the release of pro-inflammatory cytokines (e.g., IL-6, TNF-α), activation of oxidative stress responses, and disruption of the intestinal mucosal barrier [8, 9]. Currently used clinical inflammatory markers, such as C-reactive protein (CRP), only reflect a single aspect of inflammation, limiting their ability to comprehensively assess systemic immune-inflammatory-nutritional homeostasis [10].

The CALLY index—a novel integrated marker of inflammation, immunity, and nutrition—is calculated using the formula: (albumin × lymphocyte count)/CRP. This index simultaneously captures the body’s nutritional reserve, immune function, and inflammatory status, and has demonstrated prognostic value in malignancies including hepatocellular carcinoma and CRC [11, 12]. Furthermore, the neutrophil-to-lymphocyte ratio (NLR)—a simplified inflammatory marker—is increasingly employed for prognostic evaluation in gastrointestinal malignancies [13, 14]. However, few studies have investigated the utility of inflammatory markers, particularly the CALLY index, for risk stratification of colorectal polyps, and their clinical value in early polyp screening remains undefined.

Traditional risk prediction models for colorectal polyps have primarily relied on statistical methodologies such as logistic regression. These models have inherent limitations, including limited variable selection capacity, poor adaptability to variations in data scale and dimensionality, and increased risk of overfitting and reduced model stability. Furthermore, they struggle to strike a balance between interpretability and predictive performance, and fail to intuitively quantify the contribution of individual indicators to colorectal polyp occurrence. In contrast, machine learning exhibits robust capability in high-dimensional data processing and demonstrates strong adaptability to large-scale clinical datasets [15–17]. Notably, ensemble learning frameworks—such as RF and XGBoost—further enhance predictive accuracy via multi-model integration, while mitigating the inherent biases of single-model approaches. Moreover, many machine learning algorithms enable variable importance visualization, thereby facilitating a clear delineation of the predictive weights of various risk factors [18]. To address this research gap, our study performed a retrospective analysis of clinical data from a large cohort of colonoscopy participants, aiming to systematically investigate the associations of inflammatory markers (CRP, CLR, CALLY index) and traditional risk factors with colorectal polyp occurrence. Additionally, four predictive models—RF, SVM, XGBoost, and LightGBM—were developed, with the goal of providing novel tools and strategies for the targeted screening of individuals at high risk of colorectal polyps.

Materials and methods

Study population

This study design is illustrated in Fig. 1. A retrospective study was conducted, enrolling adult patients who underwent colonoscopy at the Second Qilu Hospital of Shandong University between 2023 and 2024. This study was approved by the Research Ethics Committee of the Second Qilu Hospital of Shandong University (approval number: KYLL202510750). Patient data were extracted from the hospital’s electronic health record (EHR) system, which contained comprehensive information on demographic characteristics, laboratory results, and colonoscopic findings. The inclusion criteria were as follows: (1) Age ≥ 18 years; (2) Complete colonoscopy data, with clear documentation on the presence or absence of polyps, as well as the pathological types of any polyps; (3) Completion of laboratory tests (including blood routine, biochemical indicators, and tumor markers) within one week before colonoscopy; (4) Comprehensive documentation of clinical data, including medical history and lifestyle factors. Patients were excluded if they met any of the following criteria: (1) Use of anti-inflammatory agents (e.g., glucocorticoids, nonsteroidal anti-inflammatory drugs) or immunomodulatory therapies within one month before colonoscopy; (2) History of diagnosed CRC or other malignancies (e.g., gastric cancer, lung cancer); (3) Presence of severe immunological disorders (e.g., systemic lupus erythematosus), inflammatory bowel disease, hematologic diseases (e.g., leukemia, aplastic anemia), severe liver dysfunction (e.g., Child-Pugh class C), or renal insufficiency (e.g., requiring dialysis); (4) Pregnancy, planned pregnancy, or lactation.

Fig. 1.

Schematic flowchart of the study design

Data collection

Data were collected from the EHR system and comprised the following categories:

Baseline data

Gender; age; smoking history (defined as cumulative smoking of ≥ 100 cigarettes over ≥ 6 months); drinking history (defined as alcohol consumption ≥ 3 times/week over ≥ 6 months); history of hypertension (systolic blood pressure ≥ 140 mmHg and/or diastolic blood pressure ≥ 90 mmHg, or prior diagnosis with ongoing antihypertensive therapy); history of type 2 diabetes mellitus (fasting plasma glucose ≥ 7.0 mmol/L, glycated hemoglobin ≥ 6.5%, or prior diagnosis with ongoing medication); family history of CRC in first-degree relatives; constipation (defined as ≤ 2 bowel movements/week over the past 3 months or persistent difficult defecation for ≥ 1 month); and body mass index (BMI), calculated as weight/height² (kg/m²).

Laboratory data

Triglycerides (TG); total cholesterol (TC); low-density lipoprotein cholesterol (LDL-C); albumin (ALB); white blood cell count (WBC); lymphocyte count (LYM); monocyte count (MONO); neutrophil count (NEUT); platelet count (PLT); high-sensitivity C-reactive protein (hs-CRP); carcinoembryonic antigen (CEA); and carbohydrate antigen 19 − 9 (CA19-9). The following derived indices were calculated [19–25]:

Endoscopic data

Documentation of polyp presence or absence as well as the pathological types of identified polyps.

Feature selection

To assess potential multicollinearity, we first calculated the correlation coefficients among all explanatory variables and plotted the correlation matrix [26]. To mitigate multicollinearity and reduce model overfitting risk, we then employed least absolute shrinkage and selection operator (LASSO) regression for initial variable selection. The dataset was randomly split at a ratio of 7:3 into a training set and a test set. Stratified sampling was adopted to ensure balanced sample distribution between the two subsets. We only conducted correlation analysis and LASSO-based feature selection on the training set. No feature screening procedures were applied to the test set, so as to rule out potential data leakage and ensure the reliability of model validation results.

LASSO regression incorporates an L1 regularization term, governed by a penalty parameter λ. This regularization technique promotes sparsity in the regression coefficients, thereby reducing model complexity [27]. Consequently, LASSO regression effectively identifies the most informative independent variables for constructing an optimal model. Ten-fold cross-validation was used to determine the optimal λ value, and the features selected by this model were retained for subsequent model development.

Encoding of variables

Categorical variables (e.g., gender, smoking status) were subjected to one-hot encoding, while numerical variables (e.g., TG, WBC count) were directly included in the model.

Although age was recorded as a continuous variable, it was stratified into four clinically meaningful groups (≤ 40, 41–50, 51–60, and > 60 years) for model inclusion. This categorization was adopted for two key reasons: first, the relationship between age and colorectal polyp risk is non-linear, with a pronounced increase in prevalence observed after the fifth decade of life; second, discretization enhances model interpretability and clinical applicability. However, it should be acknowledged that while this approach improves clinical utility, the discretization of continuous variables may reduce statistical granularity.

Notably, models such as LightGBM can natively process categorical variables, thus obviating the need for one-hot encoding.

Model construction and evaluation

Four machine learning models were developed to predict colorectal polyp presence: RF, SVM, XGBoost, and LightGBM. Prior to model training, all models were built and validated using R 4.3.1.

RF is an ensemble learning algorithm that uses multiple decision trees, demonstrating robust performance on large-scale datasets [28]. It is particularly adept at quantifying the relative importance of different variables during the classification process [29]. During model development, the ntree parameter typically stabilizes after 50 trees; therefore, ntree = 100 was selected as the final parameter value. For the mtry parameter, the default setting was used, corresponding to the square root of the total number of explanatory variables.

SVM is a widely employed supervised learning algorithm for both classification and regression tasks [30]. Its core objective is to identify an optimal hyperplane in the feature space that not only separates data points of different classes but also maximizes the margin between classification boundaries [31]. SVM is particularly proficient at processing high-dimensional data and accommodates both linear and nonlinear classification problems [32].

XGBoost is another ensemble learning algorithm built on decision trees [33]. Its core principle lies in an iterative process in which decision trees are trained sequentially, with each new tree focusing on correcting errors made by its predecessor [34]. This iterative refinement yields a model that becomes increasingly robust and accurate.

Similarly, LightGBM is an optimized version of the Gradient Boosting Decision Tree (GBDT) algorithm, characterized by faster training speeds and lower memory consumption compared to XGBoost [35]. A key advantage of LightGBM is its ability to directly process categorical variables, thereby mitigating inefficiencies and information loss associated with one-hot encoding [36]. For both XGBoost and LightGBM, the number of iterations was set to 500, with the learning rate fixed at 0.1.

Hyperparameter settings for the machine learning models can be found in Supplementary Table 2.

The constructed models were validated on the test set, and confusion matrices were derived using the predict() function implemented in R 4.3.1. Each confusion matrix provides counts of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) for its respective model.

Using these metrics, receiver operating characteristic (ROC) curves were generated, and key performance metrics were computed to evaluate model performance:

• Accuracy: . This metric assesses the model’s overall predictive accuracy by quantifying the proportion of accurate predictions it generates [37].

• Precision: . Precision quantifies the proportion of true positives among all predicted positive cases, reflecting the model’s accuracy in identifying positive instances.

• Recall: . Recall (or sensitivity) quantifies the proportion of true positives among all actual positive cases, reflecting the model’s capacity to correctly identify positive instances [38].

• F1-Score: . The F1-Score is the harmonic mean of precision and recall, yielding a unified metric that balances the trade-off between these two performance measures [39].

• Cohen's Kappa Coefficient: Cohen’s Kappa coefficient measures the agreement between predicted and observed classifications, adjusting for chance agreement.

In addition to discrimination metrics (AUC), model calibration was evaluated using the Hosmer-Lemeshow goodness-of-fit test. Calibration refers to the agreement between predicted probabilities and observed event frequencies. A non-significant Hosmer-Lemeshow test (P > 0.05) indicates satisfactory calibration.

All performance metrics — including accuracy, precision, recall, F1-score, and Cohen’s Kappa — were derived directly from the confusion matrices generated on the independent test set.

Ensemble learning

Ensemble learning refers to a technique that combines predictions from multiple models—commonly termed base-learners—to improve predictive performance [40]. Its core principle is to integrate the advantages of individual models, thereby yielding more accurate and robust predictions [41]. Theoretically, when base classifiers have independent error distributions, the ensemble’s misclassification rate decreases exponentially with an increasing number of base-learners, approaching zero under ideal conditions [42]. Three conventional ensemble learning strategies are commonly employed: Boosting, Bagging, and Stacking.

• Boosting operates sequentially, where each subsequent base-learner is trained to correct the errors of previous learners, leading to strong interdependencies among base-learners. A typical example is XGBoost [43].

• Bagging, by contrast, trains base-learners in parallel and independently. The final prediction is obtained via majority voting or averaging [44], as exemplified by RF.

Notably, both Boosting and Bagging typically rely on homogeneous base-learners—such as decision trees, as used in RF and XGBoost.

• Stacking, however, enables the integration of heterogeneous models. In this framework, predictions from multiple base-learners serve as input features for a meta-learner, which is trained to optimally aggregate these predictions. Logistic regression is commonly employed as the meta-learner for classification tasks [45].

In brief, the proposed Stacking architecture consists of two hierarchical layers: a base-learner layer comprising three heterogeneous classifiers (RF, XGBoost, and SVM) that independently generate predicted probabilities, and a meta-learner layer (logistic regression) that integrates these outputs to produce the final risk estimate.

In the present study, a Stacking ensemble model was developed to integrate predictions from RF, XGBoost, and SVM, with logistic regression utilized as the meta-learner to improve predictive accuracy and generalization ability.

To prevent data leakage and obtain unbiased estimates for the meta-learner, predictions from the base-learners (RF, XGBoost, and SVM) were generated using a 5-fold cross-validation strategy applied strictly to the training cohort. Specifically, the training data were randomly partitioned into five equal folds. For each iteration, four folds were used to train the base-learners, and predictions were obtained for the held-out fold (out-of-fold predictions). This process was repeated until all samples in the training set had an associated out-of-fold prediction from each base-learner. These out-of-fold predictions served as the input features for training the logistic regression meta-learner. The final model performance was evaluated on an independent test set that was completely excluded from both base-learner training and meta-learner fitting.

Feature explanation

The R packages “shapviz” and “DALEX” were employed for model interpretation. For XGBoost and LightGBM models, the SHAP method in the “shapviz” package was implemented. SHAP improves the interpretability of tree-based models via a game-theoretic approach to quantify the marginal contribution of each feature, thereby aggregating individual local explanations into a consistent global interpretation [46]. This method is widely recognized as superior to most other global approximation methods.

For RF and SVM models, interpretations were generated using the DALEX package. DALEX provides a model-agnostic interpretability framework, supporting multiple interpretation methods (e.g., SHAP) [47].To balance interpretive accuracy with computational efficiency, the type parameter in the model_parts function was set to “variable_importance” to interpret the RF and SVM models [48].

Notably, both interpretation approaches facilitate global interpretation—i.e., revealing the overall importance of each variable to the model—as well as local interpretation, i.e., clarifying the prediction mechanism for individual samples.

Results

Data statistics

A total of 946 patients were enrolled in this study, including 507 with colorectal polyps and 439 polyp-free controls. Baseline characteristics of the two groups are summarized in Table 1. Notably, several composite indices, including the CALLY index (OR = 0.959) and CLR (OR = 0.882), differed significantly between the two groups (both P < 0.05), consistent with findings from previous studies. Statistically significant differences were observed between the two groups for age, gender, smoking status, alcohol consumption, diabetes mellitus, constipation, TG, LDL-C, CA19-9 and NLR (all P < 0.05). No significant differences were found for BMI, hypertension, family history of CRC, hs-CRP, WBC, CEA, ALB, LYM, MONO, NEUT, PLR, LMR, PIV, SIRI, or SII (all P > 0.05). Given the ability of machine learning models to handle high-dimensional data, these features were not excluded; instead, LASSO regression was employed for variable selection.

Table 1.

Baseline characteristics of the 946 subjects according to colorectal polyp status

	polyp group (n = 507)	control group (n = 439)	P	OR(95% CI)
Age	55 (45, 63)	53 (41, 59)	0.009	1.019(1.005–1.033)
Gender (male)	313 (61.7%)	176 (40.1%)	< 0.001	2.441 (1.663–3.584)
BMI	25.0 (22.7, 27.6)	24.3 (21.9, 26.8)	0.643	1.011(0.967–1.056)
Smoking	128 (25.2%)	70 (15.9%)	0.018	1.565(1.352–1.905)
Alcohol drinking	75 (14.8%)	55 (12.5%)	0.016	1.939(1.134–3.315)
Hypertension	184 (36.3%)	141 (32.1%)	0.640	0.926(0.669–1.280)
Diabetes mellitus	139 (27.4)	83 (18.9%)	0.033	1.675(1.470–1.968)
Family CRC history	12 (2.4%)	11 (2.5%)	0.790	0.875(0.327–2.340)
Constipation	206 (40.6%)	119 (27.1%)	< 0.001	1.501(1.366–1.688)
TG	1.43 (0.98, 2.05)	0.99 (0.76, 1.41)	< 0.001	1.380(1.153–1.651)
ALB	45.6 (43.5, 47.7)	45.2 (43.2, 47.2)	0.284	1.028(0.977–1.083)
TC	4.70 (4.10, 5.30)	4.38 (3.71, 5.11)	0.525	0.921(0.715–1.187)
LDL-C	2.95 (2.47, 3.26)	2.63 (2.14, 3.10)	0.003	1.678(1.190–2.367)
CA19-9	11.5 (7.93, 15.8)	7.83 (5.98, 10.6)	< 0.001	1.076(1.048–1.105)
CEA	1.86 (1.10, 3.00)	1.48 (0.99, 2.41)	0.084	1.065(0.992–1.144)
hs-CRP	2.00 (1.00, 3.30)	1.40 (1.00, 2.10)	0.118	1.173(0.960–1.432)
WBC	5.69 (4.77, 6.63)	5.51 (4.58, 6.41)	0.101	3.013(0.806–11.263)
LYM	1.73 (1.38, 2.06)	1.72 (1.38, 2.09)	0.384	0.481(0.093–2.493)
MONO	0.39 (0.31, 0.48)	0.39 (0.31, 0.48)	0.114	0.019(0.000-2.609)
NEUT	3.33 (2.61, 4.11)	3.12 (2.51, 3.79)	0.206	0.414(0.106–1.623)
PLR	228 (189, 267)	232 (195, 268)	0.546	1.003(0.993–1.014)
NLR	1.92 (1.44, 2.51)	1.81 (1.35, 2.33)	0.049	1.117(0.982–1.134)
LMR	4.45(3.48, 5.67)	4.46 (3.57, 5.63)	0.694	1.060(0.792–1.419)
CLR	1.16(0.68, 1.95)	0.87 (0.60, 1.32)	0.015	1.091(0.715–1.134)
PIV	166.5(110.3, 247.1)	162.3 (104.9, 236.9)	0.614	0.999(0.994–1.004)
SIRI	0.74(0.50, 1.05)	0.70 (0.48, 0.98)	0.367	1.966(0.452–8.550)
CALLY	3.93(2.31, 6.57)	5.21 (3.36, 7.35)	0.035	0.933(0.875–0.995)
SII	427.3(313.0, 586.9)	418.0 (285.1, 569.6)	0.771	1.001(0.997–1.005)
Total	n = 946 (662 in training set and 284 in test set)

Feature selection

Correlation analysis of the features is presented in Fig. 2A. Several features showed significant correlations (|r| > 0.8), with hs-CRP and CLR exhibiting the strongest collinearity (r = 0.93, P < 0.001). To address multicollinearity and reduce overfitting, LASSO regression was employed for variable selection. Consequently, BMI, TC, LYM, NEUT, PLR, PIV, SIRI, and SII were excluded from further analyses as they did not meet the LASSO selection criteria (Fig. 2B-C).

Fig. 2.

Correlation analysis of variables and LASSO regression for feature selection

Model performance

Table 2 and Fig. 3 summarize the performance of the machine learning models in predicting colorectal polyps. Details are shown in Supplementary Table 3. When evaluated using the AUC as the primary metric, the XGBoost model achieved the highest AUC of 0.76 among the base-learners. Both the RF and SVM models attained an AUC of 0.75, while LightGBM exhibited relatively lower performance with an AUC of 0.72. Regarding other key evaluation metrics, different models displayed distinct predictive profiles: RF and SVM demonstrated high specificity, whereas XGBoost showed a notably high negative predictive value (NPV). The perfect training performance observed in the RF and XGBoost models may indicate partial overfitting; however, this issue was mitigated through the Stacking ensemble framework and independent test-set validation.The Stacking ensemble model effectively integrated the strengths of individual base-learners, achieving a substantially higher AUC of 0.96. Furthermore, it outperformed all other models across multiple key performance measures, including overall accuracy. The Stacking ensemble model also demonstrated excellent calibration, with a Hosmer-Lemeshow test P-value of 0.50 (χ² = 7.34, df = 8), indicating no significant deviation between predicted and observed probabilities. Decision curve analysis (DCA) confirmed the clinical value of the Stacking model, which demonstrated a positive net benefit across threshold probabilities from 0% to 50% (Supplementary Fig. 1). Calibration results for all other models are provided in Supplementary Table 3.

Table 2.

Comparison of predictive performance among five models

	Model	Accuracy	Sensitivity	Specificity	Kappa coefficient	NPV	AUC
Training set	RF	1.0000	1.0000	1.0000	1.0000	1.0000	1.00
	SVM	0.8205	0.8019	0.8366	0.6390	0.8296	0.88
	XGBoost	1.0000	1.0000	1.0000	1.0000	1.0000	1.00
	LightGBM	0.8808	0.9000	0.8640	0.7615	0.9077	0.87
	Stacking	1.0000	1.0000	1.0000	1.0000	1.0000	0.95
Test set	RF	0.6855	0.6794	0.6908	0.3692	0.7143	0.75
	SVM	0.6855	0.6183	0.7434	0.3639	0.6933	0.75
	XGBoost	0.6749	0.6565	0.6908	0.3469	0.7000	0.76
	LightGBM	0.6855	0.6047	0.7532	0.3608	0.6946	0.72
	Stacking	0.9046	0.8947	0.9133	0.8024	0.9073	0.96

Fig. 3.

ROC curves of the five machine learning models

Supplementary Table 1 summarizes the confusion matrices for all models evaluated on the training and test sets. This comparison reveals that the Stacking ensemble learning model demonstrates superior generalizability, as indicated by the minimal performance gap between its training and test set performance, thereby effectively alleviating overfitting.

Feature importance

Feature importance across all models was evaluated, with results summarized in Fig. 4. The SHAP summary plots (Fig. 4A-B) depict the top 15 predictive features — ranked by importance — for the XGBoost and LightGBM models. For the RF, SVM, and Stacking models, feature importance was assessed using the R package DALEX (Fig. 4C-D). Across all five models, CA19-9 was consistently identified as the most important predictive factor. Other notable features included TG, LDL-C, hs-CRP, smoking history, gender, constipation, LMR, and the CALLY index, all of which contributed substantially to predictive performance across multiple models.

Fig. 4.

SHAP analysis and feature importance assessment of machine learning models for colorectal polyp prediction

To interpret individual model predictions, the model’s decision-making process was visualized using force plots and waterfall plots generated with the R package shapviz. As illustrated in Fig. 5, the prediction mechanism of the LightGBM model for a single patient instance is depicted. In this case, features including constipation, elevated hs-CRP, and decreased CALLY index increased the predicted probability of colorectal polyps, whereas low values of CA19-9, TG, and LMR decreased this probability. This plot further quantifies the magnitude of each feature’s contribution to the final prediction.

Fig. 5.

SHAP force plot and waterfall plot of machine learning models for individual colorectal polyp prediction

Discussion

In this retrospective analysis of 946 patients who underwent colonoscopy, we integrated the CLR and CALLY index into a risk stratification framework for colorectal polyps. While previous studies have applied machine learning to polyp risk stratification, they have largely relied on conventional demographic and laboratory predictors or single-algorithm approaches [49]. We confirmed that male gender, advanced age, elevated TG, elevated LDL-C, diabetes mellitus, and constipation are independent risk factors for colorectal polyps—findings that are largely consistent with existing literature. Males accounted for 61.7% of all polyp cases. When calculated by gender across the entire cohort, the polyp incidence was 64.0% for males and 42.5% for females, consistent with cross-ethnic evidence from the study by Corley et al. [50] in a large U.S. population and that by Erdem et al. [51] in a Turkish population. Additionally, five robust machine learning models were developed and evaluated for predicting colorectal polyps.

This study identified elevated CLR and reduced CALLY index as independent risk factors for colorectal polyp formation. Chronic inflammation is a key driver in the initiation and progression of colorectal polyps and CRC, promoting aberrant proliferation of the intestinal epithelium through multiple pathophysiological pathways. On one hand, pro-inflammatory cytokines (e.g., IL-6, TNF-α) released within the inflammatory microenvironment activate key signaling cascades, including STAT3 and NF-κB. This activation subsequently upregulates the expression of oncogenes (e.g., Cyclin D1, c-Myc) and accelerates cell cycle progression. On the other hand, reactive oxygen species (ROS) generated by oxidative stress induce DNA double-strand breaks, increasing the mutation probability of tumor suppressor genes including KRAS and APC [7, 8, 52].

Elevated CLR indicates a systemic dysregulation characterized by enhanced inflammatory activity and relatively impaired cellular immune function. Elevated CRP directly reflects the severity of intestinal mucosal inflammation, while reduced lymphocyte count denotes impairment of local intestinal immune surveillance—failing to efficiently eliminate cells with early-stage aberrant proliferation. Notably, CLR has the unique advantage of integrating both inflammatory and immune information, which confers superior performance in risk stratification for colorectal polyps. As a novel comprehensive indicator, the CALLY index allows simultaneous evaluation of the body’s nutritional reserves, immune function, and inflammatory status. In the present study, the CALLY index was significantly lower in the colorectal polyp group than in the polyp-free group, which corroborates findings from the study by Iida et al. [11] (in patients with liver cancer) and that by Takeda et al. [12] (in patients with CRC). Both studies demonstrated that a decreased CALLY index is linked to poor prognosis of gastrointestinal tract malignancies, and the current study further extends its early warning value to the initial stage of colorectal polyp formation. The underlying mechanism may involve the following: In the context of chronic inflammation, pro-inflammatory cytokines (e.g., IL-6) not only inhibit hepatic albumin synthesis but also accelerate protein catabolism, resulting in diminished nutritional reserves; additionally, enhanced lymphocyte apoptosis induced by inflammatory mediators further impairs immune clearance capacity [53].

Among the identified risk factors, constipation is one of the significant risk factors (OR = 1.501), with its pathogenic mechanism involving multiple pathways—including intestinal microecology, physicochemical environment, and mucosal irritation. Constipation prolongs colonic transit time (typically > 72 h), leading to fecal stasis in the intestinal lumen. Consequently, carcinogens such as secondary bile acid metabolites (e.g., deoxycholic acid) and nitrosamines have substantially prolonged contact with the intestinal mucosa. Additionally, reduced intestinal pH facilitates the activation of these carcinogens [54–56]. Patients with constipation frequently exhibit intestinal dysbiosis, characterized by increased Bacteroidetes abundance and decreased Bifidobacterium levels. Diminished production of short-chain fatty acids (SCFAs) by the gut microbiota impairs their anti-inflammatory effects, while increased endotoxin release triggers activation of the TLR4/NF-κB pathway, which in turn amplifies inflammatory responses [57].

In this study, we integrated clinical data and inflammatory markers to develop multiple robust machine learning models, which generally demonstrated superior performance relative to conventional regression methods. This advantage mainly arises from the models’ capacity to more effectively capture complex nonlinear relationships in the data [58]. Many features included in this study—such as inflammatory markers and tumor biomarkers—exhibit distinct threshold effects. Data fluctuations around these critical thresholds have significant clinical relevance, whereas variations well below or above these thresholds tend to be of limited significance. Traditional linear models (e.g., logistic regression) often fail to capture such patterns, thus limiting their predictive performance.

Several of the developed models consistently identified carbohydrate antigen 19 − 9 (CA19-9) as the most influential predictive feature. While CA19-9 is widely acknowledged as a tumor marker for pancreatic cancer, previous studies have reported its elevation in various malignancies, benign conditions (e.g., hepatitis and cholecystitis), and even healthy individuals [59]. CA19-9 is not a tumor-specific antigen; it is commonly expressed in normal glandular epithelial cells with secretory functions [60]. In the context of inflammation, increased vascular permeability enables CA19-9 secreted by glandular tissues to enter the bloodstream, resulting in elevated serum concentrations [61]. Among all laboratory parameters included in this study, CA19-9 is minimally affected by external confounding factors. It exhibits low variability and excellent stability, and its serum concentrations are hardly altered by sampling time, dietary status or conventional anti-inflammatory drugs. In this regard, CA19-9 outperforms TG and hs-CRP, which are susceptible to external interference. During the feature learning process of machine learning algorithms, variables with high stability and low data noise are assigned higher weights. This is because such indicators can generate more reliable predictive results and are less vulnerable to random errors. These superior data characteristics enable CA19-9 to be identified as the core predictive factor across all established models.

Another notable finding is the markedly superior predictive performance of the Stacking ensemble model compared to other models, in terms of both overall accuracy and AUC. This improvement stems from two key factors. First, the Stacking framework integrates predictions from multiple base-learners, thereby harnessing their unique advantages and minimizing cumulative error. Second, the base-learners employed in our Stacking model—including RF (based on Bagging), XGBoost (gradient-boosted decision trees), and SVM (which separates data via hyperplanes)—exhibit substantial differences in their underlying algorithmic frameworks. This algorithmic diversity enables mutual complementarity, collectively enhancing predictive performance [62]. Furthermore, logistic regression was employed as the meta-learner to aggregate predictions from the base models; this design choice aids in alleviating overfitting [63]. An additional advantage of the Stacking approach is its potential for enhanced generalization with an increasing number of base-learners.In an era marked by the growing application of machine learning in clinical practice, our approach—characterized by the effective integration of diverse machine learning models—holds substantial promise for future clinical applications [64, 65].

While the Stacking model incorporates both inflammatory indices and routine laboratory markers, the practical feasibility of implementing such a model in real-world settings warrants consideration. The CALLY index and CLR are derived from readily available clinical parameters—albumin, lymphocyte count, and high-sensitivity CRP—which are routinely measured in most healthcare facilities at minimal additional cost. Similarly, CA19-9 is a standard tumor marker frequently ordered during colorectal cancer screening, though its cost and availability may vary across institutions. Notably, our feature importance analysis revealed that even when CA19-9 is not available, the combination of inflammatory markers and demographic variables still provides substantial predictive value. Future prospective studies should evaluate the cost-effectiveness and incremental benefit of incorporating these biomarkers into existing screening workflows.

The comparatively lower AUC observed for the LightGBM model (0.72) relative to the other base-learners (RF, SVM, XGBoost) may be attributed to its inherent sensitivity to hyperparameter settings and the relatively modest size of our training cohort. LightGBM employs a leaf-wise growth strategy and is particularly optimized for large-scale datasets; in smaller samples, it may be more prone to overfitting or may require extensive tuning to achieve optimal performance.

All analyses were conducted in R version 4.3.1 on a standard desktop workstation (Intel Core i7-12700 processor, 32 GB RAM, Windows 11 operating system). The base-learners required training times ranging from approximately 3.8 s (LightGBM) to 7.2 s (SVM), while the Stacking ensemble required approximately 60 s due to the 5-fold cross-validation procedure. Inference on the test set (n = 284) was completed in under 0.3 s for all models, and peak memory usage remained below 600 MB. No specialized or proprietary hardware or software is required for model implementation; the open-source R environment and freely available packages (e.g., randomForest, e1071, xgboost, lightgbm) ensure that the proposed approach is accessible and cost-effective for routine clinical informatics applications. Several limitations of this study warrant mention. First, this is a single-center study, lacking multicenter external cohort validation, which limits the generalizability of our findings. Future research should conduct multicenter external validation incorporating datasets from geographically diverse medical institutions to further confirm the stability and general applicability of the model. Furthermore, prospective real-world validation is essential prior to routine clinical implementation, providing robust evidence to support the clinical translation and widespread adoption of this predictive tool. Second, as a retrospective analysis, the accuracy of data collection is inherently constrained, and potential selection or information biases cannot be excluded. Finally, although this study explored the associations between various inflammatory markers and colorectal polyps, the underlying molecular mechanisms remain incompletely elucidated and require in-depth investigation.

Conclusion

Inflammatory markers (elevated CLR and reduced CALLY index) and traditional risk factors (e.g., gender, age, TG, LDL-C, and constipation) are collectively associated with colorectal polyp formation. The Stacking ensemble model serves as a highly accurate tool for identifying high-risk individuals, facilitating targeted clinical screening and early intervention. These findings highlight the clinical significance of integrating multi-dimensional inflammatory-nutritional markers and machine learning approaches into routine clinical practice to optimize CRC prevention strategies. Future prospective multicenter studies with external validation are warranted to confirm the clinical utility and generalizability of the proposed model.

Abbreviations

CRC

Colorectal cancer

BMI

Body mass index

EHR

Electronic health records

RF

Random forest

SVM

Support vector machine

XGBoost

eXtreme Gradient Boosting

LightGBM

Light Gradient Boosting Machine

LASSO

Least Absolute Shrinkage and Selection Operator

AUC

Area under the curve

CI

Confidence Interval

ROC

Receiver operating characteristic

GBDT

Gradient Boosting Decision Tree

TG

Triglycerides

TC

Total cholesterol

LDL-C

Low-density lipoprotein cholesterol

hs-CRP

High-sensitivity C-reactive protein

CEA

Carcinoembryonic antigen

CA19-9

Carbohydrate antigen 19 − 9

LYM

Lymphocyte Count

PLR

Platelet-to-lymphocyte ratio

CLR

C-reactive protein-to-lymphocyte ratio

NLR

Neutrophil-to-lymphocyte ratio

SII

Systemic immune-inflammation index

SIRI

Systemic inflammation response index

PIV

Pan-immune inflammation value

LMR

Lymphocyte-to-monocyte ratio

NPV

Negative predictive value

TP

True positives

FP

False positives

TN

True negatives

FN

False negatives

Authors’ contributions

LXZ, YFW and WHY designed the study. LXZ, YFW, QJ, YW and YXW contributed data collection and analysis. HQ and XNZ were involved in visualization. LXZ and YFW wrote the original draft. WHY was responsible for review and editing the draft. All authors critically reviewed the manuscript and contributed important intellectual content. All authors have read and approved the final manuscript as submitted.

Funding

No funding was received for this study.

Data availability

The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request, with permission from the Research Ethics Committee of The Second Qilu Hospital of Shandong University.

Declarations

Ethics approval and consent to participate

This retrospective and non-interventional study was reviewed and approved by the Research Ethics Committee of The Second Qilu Hospital of Shandong University (approval number: KYLL202510750). The need for written informed consent was waived by the committee due to the retrospective, non-interventional nature of the study. All procedures were conducted in accordance with the ethical principles of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.