Interpretable Machine Learning Models Based on Blood Cell-Derived Inflammatory Indices for Identifying Colorectal Neoplasia: A Retrospective Study

Xinya Zeng; Zongshou Li; Lulu Cai; Shengmei Liang; Siyan Wang; Kai Li; Ni Chen; Xiaomei Yang; Shiquan Liu

doi:10.2147/JIR.S623892

. 2026 Sep 23;19:623892. doi: 10.2147/JIR.S623892

Interpretable Machine Learning Models Based on Blood Cell-Derived Inflammatory Indices for Identifying Colorectal Neoplasia: A Retrospective Study

Xinya Zeng ^1,^*, Zongshou Li ^1,^*, Lulu Cai ^1,^*, Shengmei Liang ¹, Siyan Wang ¹, Kai Li ¹, Ni Chen ¹, Xiaomei Yang ², Shiquan Liu ^1,^✉

PMCID: PMC13617003 PMID: 42807415

Abstract

Purpose

Noninvasive approaches for the risk assessment of colorectal neoplasia remain limited, particularly in terms of interpretability and clinical applicability. This study aimed to develop and validate interpretable machine learning models using blood cell-derived inflammatory indices to identify colorectal adenoma (CRA) and colorectal cancer (CRC).

Patients and Methods

This retrospective study included 1301 participants from the Second Affiliated Hospital of Guangxi Medical University between January 2017 and April 2022, comprising 487 healthy controls (HC), 327 patients with CRA, and 487 patients with CRC. Restricted cubic spline (RCS) analysis was performed to evaluate the associations between inflammatory indices and colorectal neoplasia. Least absolute shrinkage and selection operator regression and the Boruta algorithm were used for feature selection. Five machine learning algorithms were applied and compared: logistic regression, support vector machine, k-nearest neighbors, random forest, and extreme gradient boosting (XGB). Model performance was evaluated using receiver operating characteristic analysis, calibration curves, and decision curve analysis. SHapley Additive exPlanations (SHAP) was applied to enhance interpretability.

Results

Eight blood cell–derived inflammatory indices differed significantly across groups and showed stage-related alterations. RCS analysis revealed dose-response relationships between inflammatory indices and colorectal neoplasia. Among the five algorithms, XGB achieved the best performance in the validation cohort, with areas under the curve of 0.928 (95% CI, 0.900–0.957) for CRC and 0.829 (95% CI, 0.777–0.880) for CRA. The models also showed good calibration and favorable clinical net benefit. SHAP analysis identified the platelet-to-lymphocyte ratio (PLR) as an important shared predictor, whereas the platelet-to-neutrophil ratio (PNR) and aggregate index of systemic inflammation (AISI) contributed more prominently to the prediction of CRC and CRA, respectively.

Conclusion

XGB models incorporating blood cell-derived inflammatory indices demonstrated promising discriminatory ability and interpretability, and may serve as a noninvasive adjunct for colorectal neoplasia risk stratification pending further validation in prospective and external cohorts.

Keywords: colorectal cancer, inflammatory indices, machine learning, prediction model, SHAP

Introduction

Colorectal cancer (CRC) is a major contributor to the global cancer burden and a leading cause of cancer-related mortality worldwide.¹ The prognosis of CRC is strongly stage-dependent, and the 5-year survival rate for advanced disease remains approximately 14%.² Most CRCs arise through the well-established adenoma-carcinoma sequence, in which colorectal adenoma (CRA) serves as a critical precancerous lesion.³ Therefore, early detection of CRA and CRC is essential for timely intervention and improved clinical outcomes.⁴^,⁵ Although colonoscopy remains the gold standard for colorectal neoplasia screening, its invasiveness limits broad implementation.⁶^,⁷ The fecal immunochemical test is more acceptable but demonstrates less than 40% sensitivity for precursor lesions and early-stage disease.⁸ In addition, tumor markers such as carcinoembryonic antigen (CEA) and carbohydrate antigen 19–9 (CA19-9) are commonly used for treatment evaluation and disease monitoring in CRC, although their value in screening is limited by suboptimal sensitivity and specificity.⁹ Accordingly, simple, accessible, and noninvasive approaches are still needed to complement existing screening strategies and facilitate the clinical detection of colorectal neoplasia.

Inflammation is increasingly recognized as a central driver of colorectal carcinogenesis and tumor progression.¹⁰^,¹¹ In the colorectal tumor microenvironment, inflammatory cells and mediators can promote genomic instability, angiogenesis, immune evasion, and metastasis, thereby linking immune dysregulation to malignant transformation.¹²^,¹³ In recent years, inflammatory biomarkers derived from routine peripheral blood parameters, including the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), systemic immune-inflammation index (SII), and systemic inflammation response index (SIRI), have attracted increasing attention as inexpensive and readily available indicators of cancer-related systemic inflammation.¹⁴ Compared with single hematologic parameters, these composite indices may better reflect the complex interplay among host immunity, systemic inflammation, and tumor biology. However, existing studies have mainly focused on the prognostic value of these indices in patients with established CRC. For example, elevated NLR, PLR, and SII, along with reduced LMR, have been associated with adverse pathological features and poor prognosis.^15–17 In contrast, evidence regarding the diagnostic value of these indices for colorectal neoplasia remains limited.¹⁸ Moreover, the diagnostic performance of individual markers is suboptimal, with reported AUCs of only 0.622–0.647 for NLR, LMR, and their combination in CRC detection.¹⁹

Traditional statistical approaches may be insufficient to fully capture the complex relationships between blood cell-derived inflammatory indices and colorectal neoplasia. Machine learning (ML) is a subfield of artificial intelligence that uses mathematical algorithms to identify patterns in data and generate accurate predictions, essentially mimicking human pattern recognition capabilities.²⁰ ML provides a flexible framework for integrating multidimensional clinical data and has shown considerable potential in cancer diagnosis and risk assessment; however, its clinical implementation is often hindered by limited interpretability.²¹^,²² Explainable artificial intelligence methods, such as SHapley Additive exPlanations (SHAP), can enhance model transparency by quantifying the contribution of individual predictors to model output, thereby facilitating the interpretation of model decisions.²³^,²⁴ Therefore, this study aimed to develop and evaluate interpretable machine learning models based on blood cell-derived inflammatory markers for identifying CRC and CRA, with the goal of providing a practical reference for clinical evaluation and risk stratification.

Methods

Study Population

This retrospective study consecutively enrolled individuals who underwent clinical evaluation at the Second Affiliated Hospital of Guangxi Medical University between January 2017 and April 2022. Healthy controls (HC) were defined as individuals who underwent routine health examinations and had no prior or current diagnosis of colorectal neoplasia or any other malignancy according to available medical records. CRC and CRA were diagnosed on the basis of histopathological examination. Eligible participants were required to be aged 18 years or older and to have available baseline demographic and laboratory data for analysis. Participants were excluded if they had concurrent malignancies at other sites, hematologic disorders, autoimmune diseases, acute infectious or inflammatory conditions, previous antitumor treatment, or substantial missing data. The study was reviewed and approved by the medical ethics committee of the Second Affiliated Hospital of Guangxi Medical University (2023-KY-0684) and was conducted in accordance with the Declaration of Helsinki. Given the retrospective design and the use of de-identified clinical data, the requirement for informed consent was waived in accordance with institutional policy. This study was reported in accordance with the REporting of studies Conducted using Observational Routinely-collected Data (RECORD) reporting guideline. The overall study workflow is shown in Figure 1.

A flowchart of study design from population selection to model interpretability using various steps and methods. The flowchart outlines the study design process in six steps. Step 1, Study Population & Data Collection, involves selecting participants from the Second Affiliated Hospital of Guangxi Medical University, categorized as HC, CRA and CRC, with data on demographics, tumor markers and blood cell-derived inflammatory indices. Step 2, Data Preprocessing, includes matching using Propensity Score Matching (PSM) with N=1301 and data splitting into training and validation sets at a 7:3 ratio. Step 3, Feature Selection, utilizes LASSO regression and the Boruta algorithm. Step 4, Model Construction, involves various models: LR, SVM, KNN, RF and XGB. Step 5, Model Evaluation, displays graphs comparing model performance. Step 6, Model Interpretability, uses Shapley Additive exPlanations to interpret model results, shown in graphical form. Each step is connected sequentially, illustrating the workflow from data collection to model analysis. — Flowchart of participant selection and study design.

**Abbreviations**: HC, healthy controls; CRA, colorectal adenoma; CRC, colorectal cancer; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbors; RF, random forest; XGB, extreme gradient boosting.

Data Collection and Definitions

Clinical data were extracted from the electronic medical record system. The collected variables included demographic characteristics (age, sex, and body mass index [BMI]), routine blood parameters (white blood cell count, platelet count, absolute neutrophil count, absolute lymphocyte count, and absolute monocyte count), and tumor markers (CEA and CA19-9). For patients with CRC or CRA, clinicopathological characteristics were recorded, including tumor location, TNM stage, depth of invasion (T stage), nodal status (N stage), distant metastasis (M stage), and advanced adenoma (AA) status, where applicable. AA was defined as an adenoma measuring at least 1 cm in diameter, a villous adenoma containing at least 25% villous histology, or an adenoma with high-grade dysplasia.

The following inflammatory indices were derived: NLR, neutrophil-to-monocyte ratio (NMR), PLR, platelet-to-neutrophil ratio (PNR), LMR, SII, SIRI, and aggregate index of systemic inflammation (AISI). These indices were calculated as follows: NLR = neutrophils/lymphocytes; NMR = neutrophils/monocytes; PLR = platelets/lymphocytes; PNR = platelets/neutrophils; LMR = lymphocytes/monocytes; SII = (platelets × neutrophils)/lymphocytes; SIRI = (neutrophils × monocytes)/lymphocytes; and AISI = (platelets × neutrophils × monocytes)/lymphocytes.

Data Preprocessing

Data preprocessing included propensity score matching (PSM), missing data handling, and dataset splitting. Because of imbalances in sample size, age, and sex between the colorectal neoplasia and control groups, PSM was performed for the CRC-versus-control analysis using age and sex as matching variables and a 1:1 nearest-neighbor matching strategy. After PSM, missing data were present only for BMI in four participants, whereas all other variables were complete. Missing BMI values were handled using multiple imputation to preserve the matched cohort structure. The preprocessed dataset was subsequently randomly divided into a training set and an internal validation set at a ratio of 7:3 using stratified sampling to preserve class proportions.

Variable Selection

Given the potential collinearity among blood cell counts and derived inflammatory indices, a two-step feature selection strategy was applied within the training set to identify stable and clinically meaningful predictors. First, least absolute shrinkage and selection operator (LASSO) regression was performed with 10-fold cross-validation to determine the optimal penalty parameter, and variables with nonzero coefficients were retained. Second, the Boruta algorithm was applied as a random forest-based wrapper approach to identify variables with robust predictive importance. The final predictor set was determined by integrating the results of LASSO and Boruta.

Model Development and Performance

Five ML algorithms were applied, including logistic regression (LR), support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), and extreme gradient boosting (XGB). LR is a classical linear classification model that estimates the probability of a binary outcome by modeling the relationship between one or more predictors and the log odds of the outcome.²⁵ SVM constructs a maximum-margin hyperplane based on the principle of structural risk minimization to separate different classes.²⁶ KNN is an instance-based learning algorithm that identifies the k nearest neighbors of a new observation in the feature space based on distance calculations and makes predictions using majority voting or weighted averaging.²⁷ RF and XGB are ensemble learning methods that combine multiple decision trees, using bagging and sequential boosting, respectively, to improve predictive performance and generalizability.²⁸ Model training and hyperparameter tuning were performed in the training set using grid search with 10-fold cross-validation. The optimized models were subsequently evaluated in the validation set. Model performance was assessed in terms of discrimination, calibration, and clinical utility. Discrimination was primarily evaluated using the area under the receiver operating characteristic curve (AUC). Additional performance metrics included accuracy, sensitivity, specificity, precision, the F1 score, and the Brier score. Calibration was assessed using calibration plots. Decision curve analysis (DCA) was performed to evaluate the clinical usefulness of each model across a range of threshold probabilities.

Model Interpretability

To enhance transparency and facilitate clinical interpretation, SHAP analyses were applied to the best-performing model for each classification task. SHAP provides both global and local interpretability by quantifying the contribution of each predictor to the model output. Global interpretability was presented using SHAP summary plots and feature importance rankings, which illustrate the overall impact and direction of individual predictors. Local interpretability was examined using waterfall plots for representative cases to show how specific variables influenced the predicted outcome for individual participants.

Statistical Analysis

The normality of continuous variables was assessed using the Shapiro–Wilk test. Normally distributed variables were expressed as mean ± standard deviation (SD) and compared using Student’s t-test or one-way analysis of variance (ANOVA), as appropriate. Non-normally distributed variables were expressed as median (interquartile range [IQR]) and compared using the Mann–Whitney U-test or Kruskal–Wallis test, as appropriate. Categorical variables were expressed as percentages and compared using the chi-square test or Fisher’s exact test, as appropriate. Restricted cubic spline (RCS) analysis was performed to assess the potential nonlinear associations between inflammatory indices and colorectal neoplasia. All analyses were performed using R software (version 4.2.2). A two-sided p-value < 0.05 was considered statistically significant.

Results

Baseline Characteristics

The baseline characteristics of the study population are summarized in Table 1. After eligibility screening and PSM, a total of 1301 participants were included in the final analysis, comprising 487 HC, 327 patients with CRA, and 487 patients with CRC. Overall, the CRC group showed clinical characteristics consistent with a higher disease burden, including lower BMI, higher CEA levels, and a more pronounced systemic inflammatory profile, as reflected by reduced lymphocyte counts and elevated inflammatory indices such as NLR, PLR, and SII (all P < 0.001). The study population was randomly divided into a training cohort (N = 911) and a validation cohort (N = 390) at a ratio of 7:3. The training cohort consisted of 341 HC, 229 CRA, and 341 CRC cases, whereas the validation cohort comprised 146 HC, 98 CRA, and 146 CRC cases. No significant differences were observed between the training and validation cohorts in the major demographic variables, tumor markers, or most inflammatory indices (P > 0.05). These findings support the comparability of the training and validation cohorts and their suitability for subsequent model development and validation. Detailed clinical characteristics of the training and validation cohorts are provided in Table 2.

Table 1.

Baseline Characteristics of the Study Population

Variable	Healthy (N=487)	CRA (N=327)	CRC (N=487)	P Value
Age (median [IQR])	61.00 [53.00, 69.00]	55.00 [47.00, 63.00]	61.00 [53.00, 69.00]	<0.001
Gender (%)				>0.999
Male	296 (60.78%)	199 (60.86%)	296 (60.78%)
Female	191 (39.22%)	128 (39.14%)	191 (39.22%)
BMI (median [IQR])	23.69 [21.47, 25.80]	23.44 [21.33, 25.51]	22.23 [20.03, 24.45]	<0.001
Location (%)				<0.001
Colon	–	252 (77.06%)	319 (65.50%)
Rectum	–	39 (11.93%)	168 (34.50%)
Both	–	36 (11.01%)	0 (0.00%)
Advanced adenoma (%)
YES	–	80 (24.46%)	–
NO	–	247 (75.54%)	–
AJCC T (%)
1	–	–	63 (12.94%)
2	–	–	200 (41.07%)
3	–	–	170 (34.91%)
4	–	–	54 (11.09%)
AJCC N (%)
0	–	–	275 (56.47%)
1	–	–	207 (42.51%)
2	–	–	5 (1.03%)
AJCC M (%)
0	–	–	405 (83.16%)
1	–	–	82 (16.84%)
AJCC TNM (%)
1	–	–	91 (18.69%)
2	–	–	164 (33.68%)
3	–	–	150 (30.80%)
4	–	–	82 (16.84%)
CEA (median [IQR])	1.48 [0.89, 2.42]	1.91 [1.33, 2.94]	4.14 [2.13, 11.51]	<0.001
CA19-9 (median [IQR])	13.45 [6.94, 20.93]	4.81 [2.35, 10.34]	8.13 [3.06, 21.13]	<0.001
WBC (median [IQR])	6.46 [5.52, 7.51]	5.94 [5.11, 7.23]	6.40 [5.26, 7.73]	0.002
Platelets (median [IQR])	231.40 [196.50, 267.40]	238.00 [203.00, 282.00]	276.00 [226.00, 336.00]	<0.001
Neutrophils (median [IQR])	3.60 [2.91, 4.42]	3.41 [2.74, 4.34]	3.74 [2.92, 4.96]	<0.001
Lymphocytes (median [IQR])	2.07 [1.68, 2.48]	1.98 [1.54, 2.28]	1.67 [1.39, 2.04]	<0.001
Monocytes (median [IQR])	0.46 [0.37, 0.56]	0.42 [0.33, 0.50]	0.49 [0.39, 0.63]	<0.001
NLR (median [IQR])	1.76 [1.34, 2.25]	1.79 [1.41, 2.45]	2.19 [1.60, 3.04]	<0.001
NMR (median [IQR])	7.78 [6.38, 9.88]	8.35 [6.77, 10.79]	7.71 [6.10, 9.55]	<0.001
PLR (median [IQR])	112.74 [88.39, 143.79]	123.49 [99.60, 154.38]	164.79 [124.11, 217.45]	<0.001
PNR (median [IQR])	64.77 [51.49, 80.96]	67.61 [54.80, 87.52]	71.54 [56.09, 96.84]	<0.001
LMR (median [IQR])	4.54 [3.63, 5.55]	4.82 [3.60, 6.06]	3.46 [2.59, 4.56]	<0.001
SII (median [IQR])	403.98 [293.69, 554.77]	429.66 [314.36, 600.33]	589.88 [418.05, 960.73]	<0.001
SIRI (median [IQR])	0.80 [0.57, 1.11]	0.75 [0.50, 1.06]	1.06 [0.68, 1.71]	<0.001
AISI (median [IQR])	185.67 [119.96, 271.50]	176.91 [110.14, 270.72]	298.60 [185.71, 533.26]	<0.001

Open in a new tab

Abbreviations: CRA, colorectal adenoma; CRC, colorectal cancer; BMI, body mass index; AJCC, American Joint Committee on Cancer; CEA, carcinoembryonic antigen; CA19-9, carbohydrate antigen 19–9; WBC, white blood cell count; NLR, neutrophil-to-lymphocyte ratio; NMR, neutrophil-to-monocyte ratio; PLR, platelet-to-lymphocyte ratio; PNR, platelet-to-neutrophil ratio; LMR, lymphocyte-to-monocyte ratio; SII, systemic immune-inflammation index; SIRI, systemic inflammation response index; AISI, aggregate index of systemic inflammation.

Table 2.

The Clinical Characteristics of the Study Populations for the Training and Validation Sets

Variable	Training Cohort (N=911)				Validation Cohort (N=390)				P Value^c
Variable	Healthy (N=341)	CRA (N=229)	CRC (N=341)	P value^a	Healthy (N=146)	CRA (N=98)	CRC (N=146)	P value^b	P Value^c
Age (median [IQR])	60.00 [53.00, 69.00]	56.00 [48.00, 63.00]	61.00 [53.00, 69.00]	<0.001	61.00 [54.00, 68.00]	54.00 [46.00, 59.00]	60.00 [53.00, 67.75]	<0.001	0.401
Gender (%)				0.732				0.475	0.239
Male	198 (58.06%)	140 (61.14%)	206 (60.41%)		98 (67.12%)	59 (60.20%)	90 (61.64%)
Female	143 (41.94%)	89 (38.86%)	135 (39.59%)		48 (32.88%)	39 (39.80%)	56 (38.36%)
BMI (median [IQR])	23.74 [21.45, 25.80]	23.32 [21.36, 25.45]	22.39 [20.12, 24.51]	<0.001	23.39 [21.54, 25.76]	24.06 [21.28, 25.84]	22.04 [20.00, 23.92]	<0.001	0.548
Location (%)				<0.001				<0.001	0.163
Colon	–	176 (76.86%)	231 (67.74%)		–	76 (77.55%)	88 (60.27%)
Rectum	–	25 (10.92%)	110 (32.26%)		–	14 (14.29%)	58 (39.73%)
Both	–	28 (12.23%)	0 (0.00%)		–	8 (8.16%)	0 (0.00%)
Advanced adenoma (%)									0.483
YES	–	59 (25.76%)	–		–	21 (21.43%)	–
NO	–	170 (74.24%)	–		–	77 (78.57%)	–
AJCC T (%)									0.927
1	–	–	42 (12.32%)		–	–	21 (14.38%)
2	–	–	142 (41.64%)		–	–	58 (39.73%)
3	–	–	119 (34.90%)		–	–	51 (34.93%)
4	–	–	38 (11.14%)		–	–	16 (10.96%)
AJCC N (%)									0.442
0	–	–	190 (55.72%)		–	–	85 (58.22%)
1	–	–	146 (42.82%)		–	–	61 (41.78%)
2	–	–	5 (1.47%)		–	–	0 (0.00%)
AJCC M (%)									0.512
0	–	–	286 (83.87%)		–	–	119 (81.51%)
1	–	–	55 (16.13%)		–	–	27 (18.49%)
AJCC TNM (%)									0.588
1	–	–	61 (17.89%)		–	–	30 (20.55%)
2	–	–	114 (33.43%)		–	–	50 (34.25%)
3	–	–	111 (32.55%)		–	–	39 (26.71%)
4	–	–	55 (16.13%)		–	–	27 (18.49%)
CEA (median [IQR])	1.44 [0.87, 2.30]	1.96 [1.33, 2.89]	4.07 [2.26, 10.82]	<0.001	1.59 [0.99, 2.54]	1.83 [1.36, 3.02]	4.28 [1.87, 12.38]	<0.001	0.639
CA19-9 (median [IQR])	12.61 [6.29, 20.41]	4.60 [2.35, 9.24]	9.16 [3.43, 22.72]	<0.001	15.85 [9.72, 21.86]	5.47 [2.39, 12.04]	6.66 [2.26, 16.20]	<0.001	0.472
WBC (median [IQR])	6.43 [5.51, 7.47]	5.89 [5.14, 7.20]	6.33 [5.22, 7.68]	0.026	6.58 [5.59, 7.57]	6.14 [5.09, 7.28]	6.57 [5.31, 7.86]	0.046	0.275
Platelets (median [IQR])	230.60 [196.80, 270.90]	235.00 [201.00, 283.00]	276.00 [227.00, 334.00]	<0.001	233.30 [195.50, 262.65]	248.05 [211.40, 281.78]	271.10 [224.93, 340.10]	<0.001	0.787
Neutrophils (median [IQR])	3.60 [2.91, 4.39]	3.38 [2.84, 4.28]	3.72 [2.91, 4.84]	0.027	3.60 [2.98, 4.52]	3.46 [2.64, 4.35]	3.98 [3.01, 5.04]	0.023	0.615
Lymphocytes (median [IQR])	2.07 [1.67, 2.46]	1.97 [1.50, 2.24]	1.64 [1.36, 2.02]	<0.001	2.13 [1.70, 2.58]	1.99 [1.60, 2.40]	1.73 [1.45, 2.08]	<0.001	0.021
Monocytes (median [IQR])	0.46 [0.38, 0.56]	0.42 [0.33, 0.49]	0.50 [0.40, 0.62]	<0.001	0.46 [0.37, 0.55]	0.42 [0.34, 0.54]	0.48 [0.37, 0.65]	0.005	0.994
NLR (median [IQR])	1.76 [1.35, 2.29]	1.85 [1.41, 2.46]	2.19 [1.59, 3.09]	<0.001	1.76 [1.33, 2.20]	1.63 [1.35, 2.43]	2.19 [1.62, 2.98]	<0.001	0.272
NMR (median [IQR])	7.71 [6.33, 9.72]	8.45 [7.10, 10.96]	7.63 [5.99, 9.52]	<0.001	7.90 [6.44, 10.09]	7.88 [6.51, 10.21]	7.92 [6.44, 9.77]	0.904	0.678
PLR (median [IQR])	114.81 [88.65, 143.79]	126.12 [100.00, 160.00]	168.10 [125.00, 221.47]	<0.001	106.78 [88.46, 142.38]	118.18 [98.06, 143.34]	154.61 [123.80, 213.77]	<0.001	0.106
PNR (median [IQR])	65.08 [50.43, 81.56]	66.30 [55.57, 86.63]	71.83 [56.36, 98.76]	<0.001	63.89 [53.24, 76.85]	71.24 [52.03, 90.65]	69.50 [55.74, 90.61]	0.018	0.751
LMR (median [IQR])	4.54 [3.59, 5.55]	4.72 [3.50, 6.05]	3.41 [2.57, 4.42]	<0.001	4.53 [3.75, 5.52]	5.00 [3.84, 6.04]	3.67 [2.70, 4.78]	<0.001	0.071
SII (median [IQR])	409.94 [294.28, 557.08]	437.08 [317.15, 609.60]	585.42 [418.05, 1014.33]	<0.001	396.08 [293.68, 528.95]	402.87 [306.71, 553.30]	601.32 [428.03, 874.89]	<0.001	0.441
SIRI (median [IQR])	0.79 [0.58, 1.16]	0.75 [0.52, 1.06]	1.06 [0.70, 1.63]	<0.001	0.80 [0.57, 1.00]	0.72 [0.49, 1.06]	1.06 [0.68, 1.90]	<0.001	0.508
AISI (median [IQR])	184.64 [118.39, 291.32]	176.91 [112.31, 270.72]	289.75 [187.12, 533.26]	<0.001	188.49 [124.03, 251.33]	176.94 [105.87, 268.82]	308.50 [182.01, 505.25]	<0.001	0.562

Open in a new tab

Notes: Categorical variables are presented as n (%), and continuous variables are presented as median [IQR]. Percentages indicate the proportion of patients within the corresponding cohort. ^aP values indicate comparisons among the Healthy, CRA, and CRC groups in the training cohort. ^bP values indicate comparisons among the Healthy, CRA, and CRC groups in the validation cohort. ^cP values indicate comparisons between the training and validation cohorts.

Abbreviations: CRA, colorectal adenoma; CRC, colorectal cancer; BMI, body mass index; AJCC, American Joint Committee on Cancer; CEA, carcinoembryonic antigen; CA19-9, carbohydrate antigen 19-9; WBC, white blood cell count; NLR, neutrophil-to-lymphocyte ratio; NMR, neutrophil-to-monocyte ratio; PLR, platelet-to-lymphocyte ratio; PNR, platelet-to-neutrophil ratio; LMR, lymphocyte-to-monocyte ratio; SII, systemic immune-inflammation index; SIRI, systemic inflammation response index; AISI, aggregate index of systemic inflammation.

Comparison of Inflammatory Indices Among HC, CRA, and CRC Groups

Significant differences were observed in the levels of inflammatory indices among the HC, CRA, and CRC groups (Figure 2). PLR and PNR increased progressively from HC to CRA and then to CRC, with statistically significant differences across all pairwise comparisons (all P < 0.05). In contrast, AISI, SII, SIRI, and NLR were significantly elevated in the CRC group compared with both the HC and CRA groups (all P < 0.0001), whereas no significant differences were detected between the HC and CRA groups. Conversely, LMR exhibited an opposite pattern, with significantly lower levels in the CRC group than in the HC and CRA groups. Further stratification by TNM stage showed that PLR, SII, SIRI, AISI, and NLR generally increased with tumor progression, particularly in advanced-stage CRC, whereas LMR showed a declining trend with increasing stage (Figure 3). These findings indicate that certain inflammatory alterations are already present at the adenoma stage, whereas more extensive and pronounced systemic inflammatory dysregulation is primarily associated with CRC.

Eight box-and-scatter plots comparing NLR, NMR, PLR, PNR, LMR, SII, SIRI and AISI across HC, CRA and CRC groups. Eight plots compare inflammatory indices among HC, CRA and CRC groups. Significance is marked as ns (not significant), * (p<0.05), ** (p<0.01), *** (p<0.001) and **** (p<0.0001). Each plot includes a boxplot with whiskers and scattered data points. Plot A: NLR (0-20), HC and CRA show ns; both differ from CRC (****), with CRC outliers near 20. Plot B: NMR (0-100), HC vs CRA is ns; CRA vs CRC (****), HC vs CRC (***). Plot C: PLR (0-1500), HC vs CRA (***); CRC outliers >1000. Plot D: PNR (0-500), HC vs CRA (**); CRC outliers near 500. Plot E: LMR (0-250), CRC lower than HC and CRA (****); HC vs CRA is ns. Plot F: SII (0-8000), CRC outliers >6000; HC vs CRA is ns. Plot G: SIRI (0-80), CRC outliers near 70; HC vs CRA is ns. Plot H: AISI (0-20000), CRC outliers near 17000; HC vs CRA is ns. Most indices rise towards CRC, except LMR, which decreases. — Comparison of blood cell-derived inflammatory indices among HC, CRA, and CRC groups. (A) NLR, (B) NMR, (C) PLR, (D) PNR, (E) LMR, (F) SII, (G) SIRI, (H) AISI. Kruskal–Wallis test followed by Wilcoxon rank-sum tests for pairwise comparisons. Significance is indicated as ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

Eight box and scatter plots comparing inflammatory indices across TNM stages I to IV. Panel A: NLR (0-20) shows median increase from Stage I to IV, significant between I and IV. Panel B: NMR (0-100) shows no stage differences. Panel C: PLR (0-1500) increases from Stage I to IV, significant between I and III. Panel D: PNR (0-600) peaks at Stage IV, significant between I and IV. Panel E: LMR (0-10) decreases from Stage I to IV, significant between I and IV. Panel F: SII (0-7500) increases across stages, significant between I and IV. Panel G: SIRI (0-100) rises from Stage I to IV, significant between I and IV. Panel H: AISI (0-20000) increases across stages, significant between I and IV. Color-coded points and boxes enhance visual differentiation by stage. — Comparison of blood cell-derived inflammatory indices according to TNM stage in CRC. (A) NLR, (B) NMR, (C) PLR, (D) PNR, (E) LMR, (F) SII, (G) SIRI, (H) AISI. Kruskal–Wallis test followed by Wilcoxon rank-sum tests for pairwise comparisons. Significance is indicated as ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001.

RCS Analysis of the Associations of Inflammatory Indices with CRC and CRA

RCS analysis was further performed to characterize the dose-response relationships between blood cell-derived inflammatory indices and the risks of CRC and CRA. After adjustment for confounders, NLR, PLR, and SII exhibited positive linear associations with CRC risk (P for overall < 0.05, P for nonlinear > 0.05), whereas NMR, PNR, LMR, SIRI, and AISI were nonlinearly associated with CRC risk (P for overall < 0.05, P for nonlinear < 0.05) (Figure 4). More specifically, lower levels of PNR, SIRI, and AISI were related to a lower risk of CRC, with the risk gradually increasing as the levels of these markers rose. Conversely, higher LMR was associated with a lower risk of CRC. For CRA, NMR, PLR, and PNR were linearly associated with CRA risk (P for overall < 0.05, P for nonlinear > 0.05), while LMR and SIRI showed nonlinear associations with CRA risk (P for overall < 0.05, P for nonlinear < 0.05). No significant associations were identified for NLR, SII, or AISI (P for overall > 0.05) (Figure 5).

Eight RCS curves show adjusted associations of NLR, NMR, PLR, PNR, LMR, SII, SIRI, and AISI with CRC. Image A: Line graph shows Odds Ratio rising from near 1 at low NLR to ~15 at NLR 9 (P overall <0.001, P nonlinear 0.339). Image B: Odds Ratio starts ~2.2 at NMR 4-5, drops to ~1 at NMR 8-9, rises to ~2.2 at NMR 20 (P overall 0.014, P nonlinear 0.005). Image C: Odds Ratio near 1, slight rise at PLR 500 (P overall <0.001, P nonlinear 0.483). Image D: Odds Ratio near 1 at low PNR, rises to ~12-15 at PNR 180 (P overall <0.001, P nonlinear 0.048). Image E: Odds Ratio starts ~12 at LMR 1, drops to ~1 by LMR 4-5 (P overall <0.001, P nonlinear <0.001). Image F: Odds Ratio near 1 at low SII, rises to ~150-200 at SII 3500 (P overall <0.001, P nonlinear 0.424). Image G: Odds Ratio near 1 up to SIRI 1, rises to ~8-9 at SIRI 6.5 (P overall <0.001, P nonlinear 0.003). Image H: Odds Ratio near 1 at low AISI, rises to ~40-50 at AISI 2500 (P overall <0.001, P nonlinear <0.001). — RCS analysis of associations between blood cell-derived inflammatory indices and CRC. RCS curves show the adjusted associations of (A) NLR, (B) NMR, (C) PLR, (D) PNR, (E) LMR, (F) SII, (G) SIRI, and (H) AISI with CRC. The blue line represents the fitted RCS curve, and the blue shaded area represents the 95% confidence interval. The dashed horizontal line indicates the reference line at an odds ratio of 1. P for overall indicates the overall association between the index and CRC, whereas P for nonlinear indicates whether the association deviates from linearity.

Eight line graphs showing odds ratio versus NLR, NMR, PLR, PNR, LMR, SII, SIRI and AISI. Image A: The line graph shows that the odds ratio is near 1 at low NLR and rises to approximately 1.6 at NLR 5 (P overall = 0.223, P nonlinear = 0.665). Image B: the odds ratio is below 1 at NMR 5, crosses 1 at NMR 9–10, and rises to approximately 2.5 at NMR 18 (P overall = 0.013, P nonlinear = 0.549). Image C: the odds ratio starts below 1 at PLR 50, crosses 1 at PLR 110–120, and reaches approximately 3 at PLR 250 (P overall < 0.001, P nonlinear = 0.373). Image D: the odds ratio is below 1 at PNR 25, crosses 1 at PNR 60–70, and rises to approximately 2.2 at PNR 130 (P overall = 0.042, P nonlinear = 0.823). Image E: the odds ratio is approximately 2.3 at LMR 2, drops to approximately 1 at LMR 4.5–5, and then rises to approximately 2.4 at LMR 9 (P overall = 0.010, P nonlinear = 0.008). Image F: the odds ratio is near 1 at SII 400 and gradually rises to approximately 2.1 at SII 1400 (P overall = 0.139, P nonlinear = 0.868). Image G: the odds ratio is above 2 at SIRI 0, drops to approximately 1 at SIRI 0.6–0.8, and slowly rises to approximately 1.4 at SIRI 3 (P overall = 0.029, P nonlinear = 0.011). Image H: the odds ratio is near 2 at AISI 0, drops to approximately 1 at AISI 150–200, and rises to approximately 1.6 at AISI 1000 (P overall = 0.122, P nonlinear = 0.063). — RCS analysis of associations between blood cell-derived inflammatory indices and CRA. RCS curves show the adjusted associations of (A) NLR, (B) NMR, (C) PLR, (D) PNR, (E) LMR, (F) SII, (G) SIRI, and (H) AISI with CRA. The blue line represents the fitted RCS curve, and the blue shaded area represents the 95% confidence interval. The dashed horizontal line indicates the reference line at an odds ratio of 1. P for overall indicates the overall association between the index and CRA, whereas P for nonlinear indicates whether the association deviates from linearity.

Feature Selection

To reduce feature redundancy and identify robust features for subsequent model development, candidate variables were screened using LASSO regression and the Boruta algorithm. In the comparison between HC and CRC, LASSO selected 14 features and Boruta identified 16 features, with 12 overlapping features retained by both methods, namely BMI, CA19-9, CEA, LMR, lymphocytes, monocytes, NLR, PLR, PNR, platelets, SII, and WBC (Figure 6A–C). In the comparison between HC and CRA, LASSO selected 9 features and Boruta identified 13 features, of which 7 overlapped, namely AISI, age, CA19-9, CEA, LMR, monocytes, and PLR (Figure 6D–F). Features retained by both methods were subsequently included in model development.

A composite of 2 line graphs, 2 boxplot charts and 2 Venn diagrams for LASSO and Boruta features. Image A: Line graph with ′Log(lambda)′ on x-axis (-10 to -2) and ′Binomial Deviance′ on y-axis (0.9 to 1.4). Red line dips near -7, rises to 1.4 near -2. Vertical lines at -6, -5. Top numbers: 18, 16, 16, 16, 17, 16, 14, 14, 11, 10, 9, 9, 8, 8, 7, 7, 5, 3, 2, 1. Image B: Boxplot with ′Attributes′ on x-axis and ′Importance′ on y-axis (0 to 50). Attributes rise from near 0 to CA19-9 at 45-50, CEA at 20-25, platelets at 10-15. Image C: Venn diagram ′LASSO′ and ′Boruta′: 2 in LASSO only, 12 overlap, 4 in Boruta only. Image D: Line graph with ′Log(lambda)′ on x-axis (-10 to -2) and ′Binomial Deviance′ on y-axis (1.05 to 1.35). Red line declines from 1.15 to 1.10, rises to 1.35 near -2. Vertical lines at -5, -3. Top numbers: 18, 18, 18, 16, 15, 16, 14, 12, 10, 9, 9, 9, 9, 10, 8, 6, 6, 4, 1, 0. Image E: Boxplot with ′Attributes′ on x-axis and ′Importance′ on y-axis (0 to 35). CA19-9 highest at 32-35, CEA at 12-18. Image F: Venn diagram ′LASSO′ and ′Boruta′: 2 in LASSO only, 7 overlap, 6 in Boruta only. — Feature selection using LASSO regression and the Boruta algorithm. For CRC, (A) shows the LASSO cross-validation deviance curve, (B) the Boruta feature importance plot, and (C) the overlap of features retained by the two methods. For CRA, (D) shows the LASSO cross-validation deviance curve, (E) the Boruta feature importance plot, and (F) the overlap of features retained by the two methods.

Model Development and Validation

Based on the features jointly selected by LASSO regression and the Boruta algorithm, five ML algorithms, including LR, SVM, KNN, RF, and XGB, were used to develop models for distinguishing HC from patients with CRC and CRA. In the training set, all models showed discriminative ability, although their performance varied substantially across algorithms. For CRC identification, the AUCs of the LR, SVM, KNN, RF, and XGB models were 0.889 (95% CI, 0.865–0.914), 0.922 (95% CI, 0.903–0.942), 0.831 (95% CI, 0.801–0.861), 1.000 (95% CI, 1.000–1.000), and 0.980 (95% CI, 0.972–0.988), respectively (Figure 7A). For CRA identification, the corresponding AUCs were 0.794 (95% CI, 0.758–0.831), 0.819 (95% CI, 0.785–0.853), 0.806 (95% CI, 0.770–0.842), 0.999 (95% CI, 0.997–1.000), and 0.967 (95% CI, 0.954–0.980), respectively (Figure 7B). Overall, RF and XGB showed the best performance in the training set, with RF demonstrating near-perfect classification in both tasks.

Four receiver operating characteristic plots comparing five models for colorectal cancer and colorectal adenoma. Receiver operating characteristic (ROC) curves for colorectal cancer and adenoma are presented. The horizontal axis is 1 minus specificity and the vertical axis is sensitivity, both ranging from 0.0 to 1.0. Five curves are shown with a diagonal reference line. For colorectal cancer training: - Logistic Regression (LR) AUC: 0.889 - Support Vector Machine (SVM) AUC: 0.922 - K-Nearest Neighbors (KNN) AUC: 0.831 - Random Forest (RF) AUC: 1.000 - XGBoost (XGB) AUC: 0.980 For colorectal adenoma training: - LR AUC: 0.794 - SVM AUC: 0.819 - KNN AUC: 0.806 - RF AUC: 0.999 - XGB AUC: 0.967 For colorectal cancer validation: - LR AUC: 0.837 - SVM AUC: 0.823 - KNN AUC: 0.784 - RF AUC: 0.909 - XGB AUC: 0.928 For colorectal adenoma validation: - LR AUC: 0.764 - SVM AUC: 0.775 - KNN AUC: 0.766 - RF AUC: 0.825 - XGB AUC: 0.829. — ROC curves of five machine learning models in the training and validation cohorts. (A) CRC in the training cohort; (B) CRA in the training cohort; (C) CRC in the validation cohort; and (D) CRA in the validation cohort.

In the validation set, the discriminatory performance of all models decreased to varying degrees compared with that in the training set, but remained acceptable overall. For CRC classification, XGB achieved the highest AUC of 0.928 (95% CI, 0.900–0.957), exceeding that of RF (0.909, 95% CI, 0.877–0.942) and those of the remaining models (Figure 7C). XGB also maintained a favorable balance between sensitivity and specificity, with values of 0.808 and 0.897, respectively, together with an accuracy of 0.853 and an F1 score of 0.846 (Table 3). For CRA classification, XGB again ranked first, with an AUC of 0.829 (95% CI, 0.777–0.880), slightly higher than that of RF (0.825, 95% CI, 0.772–0.877), and also showed a favorable overall balance across classification metrics, including an accuracy of 0.766, a sensitivity of 0.592, and an F1 score of 0.671 (Figure 7D and Table 4). These findings indicate that XGB maintained more stable predictive performance across different datasets, suggesting better robustness and potential generalizability for the identification of CRC and CRA. In contrast, although RF performed exceptionally well in the training set, its performance was less stable on unseen data.

Table 3.

Performance Metrics of ML Models for CRC Classification in the Training and Validation Cohorts

Model	Dataset	AUC	Accuracy	Sensitivity	Specificity	Precision	F1	Brier Score
LR	Training	0.889 (0.865–0.914)	0.804 (0.773–0.831)	0.742 (0.692–0.786)	0.865 (0.827–0.903)	0.846 (0.809–0.883)	0.791 (0.754–0.824)	0.132 (0.116–0.148)
SVM	Training	0.922 (0.903–0.942)	0.843 (0.817–0.871)	0.751 (0.704–0.795)	0.935 (0.909–0.959)	0.921 (0.892–0.948)	0.827 (0.794–0.860)	0.117 (0.105–0.129)
KNN	Training	0.831 (0.801–0.861)	0.739 (0.710–0.770)	0.601 (0.548–0.654)	0.877 (0.842–0.909)	0.830 (0.792–0.869)	0.697 (0.657–0.739)	0.173 (0.161–0.185)
RF	Training	1.000 (1.000–1.000)	1.000 (1.000–1.000)	1.000 (1.000–1.000)	1.000 (1.000–1.000)	1.000 (1.000–1.000)	1.000 (1.000–1.000)	0.017 (0.015–0.019)
XGB	Training	0.980 (0.972–0.988)	0.937 (0.918–0.956)	0.930 (0.903–0.956)	0.944 (0.918–0.968)	0.943 (0.920–0.966)	0.936 (0.918–0.955)	0.060 (0.051–0.071)
LR	Validation	0.837 (0.790–0.883)	0.771 (0.719–0.812)	0.692 (0.616–0.760)	0.849 (0.788–0.904)	0.821 (0.763–0.876)	0.751 (0.691–0.800)	0.163 (0.138–0.190)
SVM	Validation	0.823 (0.775–0.872)	0.753 (0.705–0.798)	0.623 (0.541–0.699)	0.884 (0.829–0.938)	0.843 (0.778–0.907)	0.717 (0.651–0.774)	0.166 (0.144–0.191)
KNN	Validation	0.784 (0.732–0.836)	0.702 (0.651–0.750)	0.527 (0.445–0.603)	0.877 (0.822–0.925)	0.811 (0.735–0.879)	0.639 (0.560–0.703)	0.192 (0.174–0.211)
RF	Validation	0.909 (0.877–0.942)	0.822 (0.777–0.863)	0.788 (0.726–0.849)	0.856 (0.795–0.904)	0.846 (0.794–0.893)	0.816 (0.769–0.861)	0.123 (0.101–0.146)
XGB	Validation	0.928 (0.900–0.957)	0.853 (0.812–0.890)	0.808 (0.740–0.870)	0.897 (0.842–0.938)	0.887 (0.837–0.932)	0.846 (0.801–0.885)	0.108 (0.086–0.132)

Open in a new tab

Abbreviations: ML, machine learning; CRC, colorectal cancer; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbors; RF, random forest; XGB, extreme gradient boosting; AUC, area under the receiver operating characteristic curve; F1, F1 score.

Table 4.

Performance Metrics of ML Models for CRA Classification in the Training and Validation Cohorts

Model	Dataset	AUC	Accuracy	Sensitivity	Specificity	Precision	F1	Brier Score
LR	Training	0.794 (0.758–0.831)	0.698 (0.660–0.735)	0.598 (0.533–0.664)	0.765 (0.718–0.809)	0.631 (0.580–0.683)	0.614 (0.565–0.660)	0.180 (0.167–0.195)
SVM	Training	0.819 (0.785–0.853)	0.730 (0.695–0.765)	0.642 (0.581–0.699)	0.789 (0.745–0.830)	0.671 (0.623–0.723)	0.656 (0.611–0.700)	0.170 (0.158–0.184)
KNN	Training	0.806 (0.770–0.842)	0.754 (0.719–0.788)	0.694 (0.638–0.751)	0.795 (0.751–0.836)	0.694 (0.644–0.740)	0.694 (0.650–0.738)	0.180 (0.168–0.192)
RF	Training	0.999 (0.997–1.000)	0.975 (0.963–0.988)	0.952 (0.921–0.978)	0.991 (0.979–1.000)	0.986 (0.969–1.000)	0.969 (0.953–0.985)	0.048 (0.043–0.054)
XGB	Training	0.967 (0.954–0.980)	0.902 (0.875–0.925)	0.878 (0.834–0.917)	0.918 (0.886–0.944)	0.878 (0.839–0.913)	0.878 (0.845–0.907)	0.119 (0.112–0.126)
LR	Validation	0.764 (0.702–0.826)	0.734 (0.680–0.787)	0.541 (0.439–0.633)	0.863 (0.808–0.911)	0.726 (0.640–0.815)	0.620 (0.526–0.699)	0.194 (0.168–0.222)
SVM	Validation	0.775 (0.715–0.834)	0.721 (0.672–0.779)	0.541 (0.449–0.643)	0.842 (0.781–0.897)	0.697 (0.613–0.792)	0.609 (0.530–0.690)	0.188 (0.161–0.215)
KNN	Validation	0.766 (0.706–0.826)	0.717 (0.664–0.770)	0.531 (0.439–0.633)	0.842 (0.781–0.897)	0.693 (0.610–0.784)	0.601 (0.520–0.685)	0.193 (0.171–0.214)
RF	Validation	0.825 (0.772–0.877)	0.762 (0.713–0.811)	0.551 (0.449–0.653)	0.904 (0.856–0.952)	0.794 (0.704–0.881)	0.651 (0.562–0.733)	0.166 (0.142–0.192)
XGB	Validation	0.829 (0.777–0.880)	0.766 (0.717–0.816)	0.592 (0.500–0.694)	0.884 (0.829–0.932)	0.773 (0.688–0.855)	0.671 (0.587–0.746)	0.172 (0.156–0.191)

Open in a new tab

Abbreviations: ML, machine learning; CRA, colorectal adenoma; LR, logistic regression; SVM, support vector machine; KNN, k-nearest neighbors; RF, random forest; XGB, extreme gradient boosting; AUC, area under the receiver operating characteristic curve; F1, F1 score.

Calibration curves showed good overall agreement between predicted and observed probabilities in both the training and validation sets, with XGB and RF demonstrating better calibration than LR, SVM, and KNN (Figure 8A–D). Decision curve analysis further demonstrated that, for both CRC and CRA classification, RF and XGB provided greater net benefit across a wide range of threshold probabilities, whereas LR, SVM, and KNN offered relatively limited clinical utility (Figure 8E–H). Notably, when discrimination, calibration, and clinical net benefit were considered together in the validation set, XGB showed the most favorable balance of overall performance. Therefore, although RF showed the strongest fitting performance in the training set, XGB was ultimately selected as the final model.

Graphs of calibration plots & decision curves for CRC/CRA in training & validation sets. Eight line graphs compare five machine learning models labeled LR, SVM, KNN, RF and XGB across calibration plots and decision curve analyses for CRC and CRA. Panel A, CRC Training calibration: x-axis is predicted probability, y-axis is observed probability, both ranging 0.0 to 1.0. RF and XGB track closest to the dashed diagonal. Panel B, CRC Validation calibration: same axes; curves show more scatter, with XGB and RF remaining nearest the diagonal across the 0.2 to 0.8 range. Panel C, CRA Training calibration: same axes; RF rises steeply near 0.4 to 0.6 predicted probability, while XGB stays closer to the diagonal. Panel D, CRA Validation calibration: same axes; curves show notable crossings, with XGB and RF showing closer alignment to the diagonal than LR, SVM and KNN. Panel E, CRC Training decision curve: x-axis is threshold probability 0.0 to 1.0, y-axis is net benefit. RF holds the highest net benefit from approximately 0.1 to 0.6, followed by XGB. Panel F, CRC Validation decision curve: same axes; RF and XGB maintain the highest net benefit through mid-range thresholds, with several curves crossing below zero near 0.9. Panel G, CRA Training decision curve: same axes; RF peaks near 0.4 net benefit at low thresholds, followed by XGB. Panel H, CRA Validation decision curve: same axes; RF and XGB offer the greatest net benefit across thresholds, with LR and SVM dropping below zero near 0.8. — Calibration plots and DCA of five machine learning models. Calibration plots for (A) CRC in the training cohort, (B) CRC in the validation cohort, (C) CRA in the training cohort, and (D) CRA in the validation cohort. DCA for (E) CRC in the training cohort, (F) CRC in the validation cohort, (G) CRA in the training cohort, and (H) CRA in the validation cohort. The dashed diagonal line in the calibration plots indicates perfect calibration. In the DCA plots, the dashed gray line indicates the treat-all strategy, and the solid black line indicates the treat-none strategy.

SHAP-Based Model Interpretation

To quantify the contribution of each feature to model predictions, SHAP analysis was applied to the final XGB models for both global and local interpretation. In the CRC model, the SHAP summary bar plot showed that, apart from CEA and CA19-9, PLR was the most important inflammatory biomarker, followed by LMR, platelets, BMI, PNR, WBC, lymphocytes, monocytes, SII, and NLR (Figure 9A). In the CRA model, the most important clinical features, in descending order, were CA19-9, CEA, age, monocytes, PLR, LMR, and AISI (Figure 9D). The SHAP summary dot plots further illustrated the direction and magnitude of these effects in the CRC and CRA models (Figure 9B and E). Higher PLR values were associated with higher SHAP values, thereby increasing the model output toward CRC, whereas higher LMR and lymphocyte values were associated with a lower predicted probability of CRC (Figure 9B). Notably, multiple inflammatory markers, including monocytes, PLR, LMR, and AISI, were retained in the final model, indicating that inflammation-related alterations already carried predictive value at the adenoma stage. SHAP dependence plots further showed that the effects of these inflammatory markers on model output were nonlinear rather than simple monotonic increases or decreases (Figures 10 and 11).

Six plots showing SHAP feature importance, SHAP value distributions and prediction waterfalls for two models. Image A: Bar chart of mean SHAP values. CEA leads at 0.20, followed by CA19-9 at 0.10, PLR at 0.07 and NLR at 0.01. Image B: SHAP dot plot with values from -0.25 to 0.50. CEA ranges -0.25 to 0.50, CA19-9 -0.10 to 0.35, PLR -0.05 to 0.30, others near 0. Image C: SHAP waterfall plot, prediction 0.4 to 1.0. Key values: CEA 37.99, Platelets 352, PLR 19.26. Base 0.500, final 0.991, largest step +0.380. Image D: Bar chart of mean SHAP values. CA19-9 at 0.10, CEA 0.05, Age 0.03, AISI 0.01. Image E: SHAP dot plot, values -0.2 to 0.3. CA19-9 -0.2 to 0.3, CEA -0.15 to 0.25, Age -0.10 to 0.10. Image F: SHAP waterfall plot, prediction 0.4 to 0.8. Key values: CA19-9 2.33, CEA 2.7, Age 47. Base 0.442, final 0.788, largest step +0.266. — SHAP-based interpretation of the final XGB models. For CRC, (A) SHAP summary bar plot, (B) SHAP summary dot plot, and (C) SHAP waterfall plot. For CRA, (D) SHAP summary bar plot, (E) SHAP summary dot plot, and (F) SHAP waterfall plot. Bar plots rank features by mean absolute SHAP value; dot plots display the distribution and direction of SHAP values; waterfall plots show feature-level contributions to individual predictions.

Twelve scatter plots showing SHAP values for features like CEA, CA19-9, PLR and others in the CRC model. The image features twelve scatter plots showing SHAP dependence for CRC model features: CEA, CA19-9, PLR, LMR, Platelets, BMI, PNR, WBC, Lymphocytes, Monocytes, SII and NLR. Each plot′s x-axis is the feature value and the y-axis is the SHAP value. Points are color-coded from purple to yellow, indicating secondary feature values. Plot A (CEA) and Plot B (CA19-9) show high SHAP values at lower feature values. Plot C (PLR) has a positive correlation, while Plot D (LMR) shows a negative correlation. Plots E (Platelets) and F (BMI) have nonlinear relationships. Plots G (PNR) and H (WBC) show decreasing SHAP values with increasing feature values. Plot I (Lymphocytes) has a dense cluster at lower values. Plots J (Monocytes), K (SII) and L (NLR) display varied nonlinear patterns. Outliers and clusters highlight significant feature impacts on model predictions. — SHAP dependence plots for the CRC model. SHAP dependence plots for (A) CEA, (B) CA19-9, (C) PLR, (D) LMR, (E) Platelets, (F) BMI, (G) PNR, (H) WBC, (I) Lymphocytes, (J) Monocytes, (K) SII, and (L) NLR in the final CRC model. Each point represents one observation. The x-axis indicates the feature value, and the y-axis indicates the corresponding SHAP value.

Scatter plots showing SHAP values for CA19-9, CEA, Age, Monocytes, PLR, LMR and AISI in the CRA model. Scatter plots illustrate SHAP values for various features against their respective values, highlighting feature contributions to the CRA model. Plot A shows CA19-9 values (0-150) with SHAP values (-0.2 to 0.3), where lower CA19-9 values have higher SHAP contributions. Plot B depicts CEA values (0-20) with SHAP values (-0.2 to 0.1), showing significant contributions at lower CEA values. Plot C presents Age (20-80) with SHAP values (-0.1 to 0.1), indicating older age correlates with lower SHAP values. Plot D shows Monocytes (0.0-1.5) with SHAP values (-0.05 to 0.05), where higher levels correspond to lower SHAP values. Plot E illustrates PLR (0-500) with SHAP values (-0.05 to 0.10), showing an upward trend with higher PLR values. Plot F displays LMR (0-200) with SHAP values (0.00 to 0.10), concentrated at lower values. Plot G shows AISI (0-2000) with SHAP values (-0.02 to 0.04), clustered at lower values. These plots demonstrate varying feature importance in the model. — SHAP dependence plots for the CRA model. SHAP dependence plots for (A) CA19-9, (B) CEA, (C) Age, (D) Monocytes, (E) PLR, (F) LMR, and (G) AISI in the final CRA model. Each point represents one observation. The x-axis indicates the feature value, and the y-axis indicates the corresponding SHAP value.

In addition, local interpretation further clarified the model’s decision-making process at the individual-patient level. The SHAP waterfall plot for CRC illustrated how the prediction for case 253 was derived (Figure 9C). Elevated CEA (37.99 ng/mL), elevated CA19-9 (58.36 U/mL), high platelet count (352 × 10⁹/L), and high PLR (189.25) collectively increased the prediction score from the baseline value (E[f(x)] = 0.500) to a final value of 0.991. Other features, including BMI, WBC, LMR, PNR, NLR, SII, lymphocytes, and monocytes, also contributed to varying degrees. For CRA, the SHAP waterfall plot showed that the baseline prediction of 0.442 increased to 0.788 (Figure 9F). In this case, CA19-9 (2.33, +0.206) made the strongest positive contribution, followed by CEA (2.70, +0.051), age (47, +0.039), and monocyte count (0.37 × 10⁹/L, +0.028), whereas LMR made a relatively small negative contribution. In summary, SHAP analysis indicated that PLR was a shared key inflammatory feature in both CRC and CRA prediction, whereas PNR was more specifically associated with CRC and AISI with CRA, suggesting that the composition of inflammatory signatures differs across stages of colorectal tumorigenesis.

Discussion

This study systematically evaluated the associations of eight blood cell-derived inflammatory indices, including NLR, NMR, PLR, PNR, LMR, SII, SIRI, and AISI, with colorectal neoplasia, and further developed interpretable ML models for identifying CRA and CRC. Among the five candidate algorithms, XGB showed the most robust performance in the validation cohort, with AUCs of 0.928 (95% CI, 0.900–0.957) for CRC and 0.829 (95% CI, 0.777–0.880) for CRA, as well as good calibration and clinical net benefit. Overall, these findings further support an association between systemic inflammatory imbalance and colorectal neoplasia development, while highlighting the potential of ML models based on routinely available blood cell-derived inflammatory indices for noninvasive risk stratification.

Previous studies have demonstrated that the initiation and progression of CRC are closely linked to inflammation and immune dysregulation.²⁹^,³⁰ Platelets release vascular endothelial growth factor (VEGF), transforming growth factor-β (TGF-β), and multiple chemokines, thereby promoting tumor angiogenesis, epithelial-mesenchymal transition, and metastasis.³¹^,³² Lymphocytes, particularly cytotoxic T cells and natural killer cells, are key mediators of antitumor immune surveillance.³³ After infiltrating tumor tissue, monocytes can differentiate into tumor-associated macrophages, which promote immune suppression, stromal remodeling, and tumor progression through inflammatory mediators such as interleukin-6, tumor necrosis factor-α, and TGF-β.³⁴^,³⁵ Neutrophils also display substantial functional plasticity within the tumor microenvironment. They can promote angiogenesis and tumor invasion by secreting vascular endothelial growth factor, reactive oxygen species, and other mediators, and can also inhibit T-cell proliferation through immune checkpoint-related pathways, thereby fostering an immunosuppressive microenvironment that favors tumor progression.³⁶^,³⁷ Accordingly, tumor-associated inflammation is involved not only in remodeling the local tumor microenvironment, but also in driving systemic changes in peripheral blood cell components and their derived indices. Previous studies have shown that increased NLR and PLR, along with decreased LMR, are generally associated with more advanced disease, poorer survival, and diminished treatment response in CRC.³⁸ Rezazadeh et al investigated the relationship between peripheral blood inflammatory indices and the clinicopathological characteristics of CRC, and found that higher SIRI levels were associated with an increased risk of metastasis.³⁹ Compared with previous studies, our study incorporated eight blood cell-derived inflammatory indices and extended the analysis to both precancerous lesions and invasive cancer. PLR increased progressively from HC to CRA to CRC, suggesting that platelet-related inflammatory activation may emerge as early as the adenoma stage. In contrast, NLR, LMR, SII, SIRI, and AISI differed primarily in CRC compared with the other two groups, suggesting that neutrophil and monocyte-mediated inflammatory activity becomes more pronounced as lesions progress to invasive cancer, whereas lymphocyte-mediated antitumor immune surveillance is relatively weakened. Composite inflammatory indices are more sensitive than single blood cell parameters in capturing the early inflammatory disequilibrium underlying colorectal tumorigenesis.

RCS analysis further showed that the relationship between inflammatory indices and colorectal neoplasia was not simply linear. Previous studies have often categorized continuous variables using specific thresholds or assumed a linear risk relationship by default. However, the cut-off values used across studies have been inconsistent. For example, one study reported that the optimal cut-off values of NLR, PLR, and SII for distinguishing patients with CRC from controls were 2.28, 122.40, and 616.50, respectively.⁴⁰ In a colonoscopy-based population, Güzelyüz Bodur et al used exploratory thresholds of 1.997 for NLR, 121.960 for PLR, and 543.625 for SII to identify malignant lesions, with corresponding AUCs ranging from 0.633 to 0.691.⁴¹ These data indicate that although analyses based on a single cut-off point may facilitate result interpretation and clinical translation, they also lead to loss of information and make it difficult to accurately characterize the true pattern of risk variation across different levels of these indices. In the present study, clear nonlinear associations were observed between PNR, LMR, SIRI, and AISI and CRC, as well as between LMR and SIRI and CRA. These findings suggest that the effects of some inflammatory indices on disease risk may not be constant across their distributions and may instead be more pronounced within specific value ranges. Such nonlinearity may explain why previous studies based on single cut-off points or conventional linear models have often yielded heterogeneous results and limited diagnostic performance.

ML can extract complex patterns from high-dimensional, nonlinear clinical data and therefore offers advantages over conventional statistical methods in risk prediction.⁴²^,⁴³ For example, Schallmoser et al constructed a risk stratification model for young-onset CRC based on demographic characteristics, symptoms, and laboratory data, in which the RF algorithm achieved an AUC of 0.859 in internal validation.⁴⁴ Another study based on routine laboratory parameters reported that the RF model achieved an AUC of 0.863 for CRC diagnosis, outperforming both CEA alone and the combination of CEA and CA19-9.⁴⁵ Nevertheless, previous studies have provided limited evidence regarding the systematic evaluation of inflammation-related features and their value for risk stratification in precancerous lesions. In the present study, we compared five algorithms, including LR, SVM, KNN, RF, and XGB, and found that XGB demonstrated good discriminative performance in the prediction of both CRC and CRA. Notably, these models were based primarily on routine hematologic parameters, blood cell-derived inflammatory indices, and tumor markers, all of which are readily available and relatively low-cost in routine clinical practice. These features may make the models useful as adjuncts in screening settings where colonoscopy resources are limited or patient adherence is suboptimal. In addition, LASSO regression and the Boruta algorithm were jointly used for feature selection to prioritize stable and informative variables. Given the strong intercorrelations among inflammatory indices, this dual-selection strategy may reduce feature redundancy and mitigate multicollinearity, thereby improving model robustness and reproducibility.

To improve model interpretability, SHAP was applied to the XGB model at both the global and individual levels. Model interpretability is essential for the clinical application of ML models because predictions are more likely to be accepted when they can be linked to specific features and the direction of their effects.⁴⁶^,⁴⁷ SHAP analysis showed that the key feature composition differed between the CRC and CRA models. PLR was an important shared inflammatory feature in both models, suggesting that platelet-to-lymphocyte imbalance may represent a common inflammatory signal across different stages of colorectal neoplasia. Beyond PLR, PNR, platelet count, and lymphocyte-related variables contributed more prominently to the CRC model, whereas AISI and monocyte-related features were relatively more important in the CRA model. These findings suggest that the pattern of inflammatory response evolves during the transition from precancerous lesions to invasive cancer. In addition to ranking feature importance, SHAP can identify the main drivers of individual risk predictions, thereby making model outputs clinically traceable. Local SHAP analysis further showed that combinations of CEA, CA19-9, PLR, and platelet count could jointly drive the model toward a positive prediction. These findings suggest that inflammatory indices may provide complementary information beyond CEA and CA19-9 by reflecting the host immune-inflammatory status, which may help improve the discrimination of colorectal neoplasia.

Several limitations should be acknowledged. First, this was a single-center retrospective study, and all participants were recruited from the same hospital, which may have introduced selection bias. Second, although internal validation was performed, these models have not yet been tested in an independent external cohort. Therefore, model performance may be influenced by patient characteristics, laboratory measurement procedures, and disease spectrum, and their generalizability remains uncertain. Prospective multicenter validation would be valuable for further assessing the predictive performance and potential clinical applicability of these models. Third, the healthy control group comprised individuals undergoing routine health examinations, but these individuals did not undergo colonoscopy. The disease spectrum of the control population may differ from that of patients evaluated in clinical practice for colorectal symptoms or suspected colorectal neoplasia. This difference may limit the representativeness of our findings in complex clinical diagnostic settings. Accordingly, prospective studies should include colonoscopy-confirmed controls, particularly individuals evaluated for suspected colorectal neoplasia and patients with other non-neoplastic colorectal conditions, to further assess the clinical utility of these models in more challenging diagnostic settings. Fourth, because molecular subtype information was unavailable, we could not assess whether the predictive contribution of inflammatory indices varied according to microsatellite instability status, CpG island methylator phenotype, or consensus molecular subtype. Integrating peripheral inflammatory indices with molecular and pathological features in subsequent studies may help clarify their diagnostic value in specific biological subgroups.

Conclusions

In conclusion, blood cell-derived inflammatory indices showed potential clinical value in identifying colorectal neoplasia. Machine-learning models based on these inflammatory indices may support risk stratification and clinical decision-making in colorectal neoplasia. Further prospective multicenter studies are needed to validate their robustness and clinical utility.

Acknowledgments

We extend our sincere appreciation to all the researchers for their contributions.

Funding Statement

This study was funded by the National Natural Science Foundation of China (No. 82260579), Joint Project on Regional High-Incidence Diseases Research of Guangxi Natural Science Foundation under Grant (No. 2024GXNSFDA010004), and the Fangchenggang Science and Technology Programme (Grant No. fangke AB24002030).

Abbreviations

AISI, aggregate index of systemic inflammation; BMI, body mass index; CA19-9, carbohydrate antigen 19-9; CEA, carcinoembryonic antigen; CRA, colorectal adenoma; CRC, colorectal cancer; DCA, decision curve analysis; HC, healthy controls; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; LMR, lymphocyte-to-monocyte ratio; LR, logistic regression; ML, machine learning; NLR, neutrophil-to-lymphocyte ratio; NMR, neutrophil-to-monocyte ratio; PLR, platelet-to-lymphocyte ratio; PNR, platelet-to-neutrophil ratio; RCS, restricted cubic spline; RF, random forest; ROC, receiver operating characteristic; SHAP, Shapley additive explanations; SII, systemic immune-inflammation index; SIRI, systemic inflammation response index; SVM, support vector machine; TNM, tumor-node-metastasis; WBC, white blood cell count; XGB, extreme gradient boosting.

Data Sharing Statement

Data can be obtained from the corresponding author upon reasonable request.

Ethical Statement

The study was reviewed and approved by the medical ethics committee of the Second Affiliated Hospital of Guangxi Medical University (2023-KY-0684). As this study involved retrospective data analysis, and the personal information of all patients and controls had been anonymized and de-identified before analysis, the need for informed consent was waived by the ethics committee of the Second Affiliated Hospital of Guangxi Medical University.

Author Contributions

Xinya Zeng: Conceptualization; Formal analysis; Software; Methodology; Writing-original draft. Zongshou Li: Formal analysis;Visualization; Methodology; Writing-original draft. Lulu Cai: Formal analysis; Validation; Methodology; Writing-original draft. Shengmei Liang: Software; Methodology; Writing-review and editing. Siyan Wang: Investigation; Writing-review and editing. Kai Li: Investigation; Writing-review and editing. Ni Chen: Data curation; Project administration; Writing-review and editing. Xiaomei Yang: Validation; Supervision; Writing-review & editing. Shiquan Liu: Conceptualization; Supervision; Funding acquisition; Writing-review & editing. All authors took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare no conflict of interest.

References

1.Force LM, Kocarnik JM, May ML. et al. The global, regional, and national burden of cancer, 1990–2023, with forecasts to 2050: a systematic analysis for the global burden of disease study 2023. Lancet. 2025;406(10512):1565–22. doi: 10.1016/S0140-6736(25)01635-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Shin AE, Giancotti FG, Rustgi AK. Metastatic colorectal cancer: mechanisms and emerging therapeutics. Trends Pharmacol Sci. 2023;44(4):222–236. doi: 10.1016/j.tips.2023.01.003 [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Chen Y, Zhang J, Ding Y, Zhu F, Chen Y. Colorectal cancer pathogenesis, oncogenic signaling networks and targeted therapeutic advances. Mol Biomed. 2026;7(1):32. doi: 10.1186/s43556-026-00433-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Krul MF, Elferink MAG, Kok NFM, et al. Initial impact of national CRC screening on incidence and advanced colorectal cancer. Clin Gastroenterol Hepatol. 2023;21(3):797–807.e3. doi: 10.1016/j.cgh.2022.08.046 [DOI] [PubMed] [Google Scholar]
5.Shi J, Løberg M, Kalager M, et al. Effect of colonoscopy screening on risks of colorectal cancer and related death: instrumental variable estimation of per-protocol effects. Eur J Epidemiol. 2025;40(4):419–425. doi: 10.1007/s10654-025-01209-w [DOI] [PubMed] [Google Scholar]
6.Ola I, Cardoso R, Hoffmeister M, Brenner H. Utilization of colorectal cancer screening tests: a systematic review and time trend analysis of nationally representative data. Eclinicalmedicine. 2024;75:102783. doi: 10.1016/j.eclinm.2024.102783 [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Westerberg M, Ludvigsson JF, Metcalfe C, et al. Colonoscopy and fecal immunochemical testing versus usual care in diagnostic colorectal cancer screening: the SCREESCO randomized controlled trial. Nat Med. 2026;2026:1. doi: 10.1038/s41591-026-04225-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Levy BT, Xu Y, Daly JM, et al. Comparative performance of common fecal immunochemical tests: a cross-sectional study. Ann Intern Med. 2024;177(10):1350–1360. doi: 10.7326/M24-0080 [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Lech G. Colorectal cancer tumour markers and biomarkers: recent therapeutic advances. WJG. 2016;22(5):1745. doi: 10.3748/wjg.v22.i5.1745 [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Schmitt M, Greten FR. The inflammatory pathogenesis of colorectal cancer. Nat Rev Immunol. 2021;21(10):653–667. doi: 10.1038/s41577-021-00534-x [DOI] [PubMed] [Google Scholar]
11.Bardelčíková A, Šoltys J, Mojžiš J. Oxidative stress, inflammation and colorectal cancer: an overview. Antioxidants. 2023;12(4):901. doi: 10.3390/antiox12040901 [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Burgos-Molina AM, Téllez Santana T, Redondo M, Bravo Romero MJ. The crucial role of inflammation and the immune system in colorectal cancer carcinogenesis: a comprehensive perspective. Int J Mol Sci. 2024;25(11):6188. doi: 10.3390/ijms25116188 [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Wang J, Xiao H, Cui S, Wu C, Xiang D. The bridging role of neutrophils in the progression of inflammation-induced colorectal cancer. Biochim Biophys Acta Rev Cancer. 2025;1880(6):189460. doi: 10.1016/j.bbcan.2025.189460 [DOI] [PubMed] [Google Scholar]
14.Sylman JL, Mitrugno A, Atallah M, et al. The predictive value of inflammation-related peripheral blood measurements in cancer staging and prognosis. Front Oncol. 2018;8:78. doi: 10.3389/fonc.2018.00078 [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Yu L, Yuan J, Lou S, et al. Factors affecting survival in patients with colorectal cancer: an umbrella review. J Transl Med. 2025;23(1):1185. doi: 10.1186/s12967-025-06876-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Yamamoto T, Kawada K, Obama K. Inflammation-Related Biomarkers for the Prediction of Prognosis in Colorectal Cancer Patients. Ijms. 2021;22(15):8002. doi: 10.3390/ijms22158002 [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Menyhart O, Fekete JT, Győrffy B. Inflammation and colorectal cancer: a meta-analysis of the prognostic significance of the systemic immune–inflammation index (SII) and the systemic inflammation response index (SIRI). Int J Mol Sci. 2024;25(15):8441. doi: 10.3390/ijms25158441 [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Yan L, Yi B, Fang Y. From biomarkers to mechanisms: deepening the clinical translation of inflammatory indices in colorectal cancer pathogenesis. Int J Surg. 2026;112(1):2076–2077. doi: 10.1097/JS9.0000000000003405 [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Zhou M, Gu Q, Zhou M, et al. Extensive study on the associations of 12 composite inflammatory indices with colorectal cancer risk and mortality: a cross-sectional analysis of NHANES 2001–2020. Int J Surg. 2025;111(11):7559–7575. doi: 10.1097/JS9.0000000000002996 [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Wang J, Zeng Z, Li Z, et al. The clinical application of artificial intelligence in cancer precision treatment. J Transl Med. 2025;23(1):120. doi: 10.1186/s12967-025-06139-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Zhang C, Xu J, Tang R, et al. Novel research and future prospects of artificial intelligence in cancer diagnosis and treatment. J Hematol Oncol. 2023;16(1):114. doi: 10.1186/s13045-023-01514-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Liu K, Luo Z, Zhang W, et al. Artificial intelligence models: transforming early diagnosis and precise treatment of gastrointestinal cancers. Mol Cancer. 2026;2026:1. doi: 10.1186/s12943-026-02589-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Allgaier J, Mulansky L, Draelos RL, Pryss R. How does the model make predictions? A systematic literature review on the explainability power of machine learning in healthcare. Artif Intell Med. 2023;143:102616. doi: 10.1016/j.artmed.2023.102616 [DOI] [PubMed] [Google Scholar]
24.Toumaj S, Heidari A, Jafari Navimipour N. Leveraging explainable artificial intelligence for transparent and trustworthy cancer detection systems. Artif Intell Med. 2025;169:103243. doi: 10.1016/j.artmed.2025.103243 [DOI] [PubMed] [Google Scholar]
25.Zabor EC, Reddy CA, Tendulkar RD, Patil S. Logistic Regression in Clinical Studies. Int J Radiat Oncol Biol Phys. 2022;112(2):271–277. doi: 10.1016/j.ijrobp.2021.08.007 [DOI] [PubMed] [Google Scholar]
26.Rodríguez-Pérez R, Vogt M, Bajorath J. Support vector machine classification and regression prioritize different structural features for binary compound activity and potency value prediction. ACS Omega. 2017;2(10):6371–6379. doi: 10.1021/acsomega.7b01079 [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Maleki F, Ovens K, Najafian K, Forghani B, Reinhold C, Forghani R. Overview of machine learning part 1. Neuroimaging Clin N Am. 2020;30(4):e17–e32. doi: 10.1016/j.nic.2020.08.007 [DOI] [PubMed] [Google Scholar]
28.Eloranta S, Boman M. Predictive models for clinical decision making: deep dives in practical machine learning. J Intern Med. 2022;292(2):278–295. doi: 10.1111/joim.13483 [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Lasry A, Zinger A, Ben-Neriah Y. Inflammatory networks underlying colorectal cancer. Nat Immunol. 2016;17(3):230–240. doi: 10.1038/ni.3384 [DOI] [PubMed] [Google Scholar]
30.Zhou RW, Harpaz N, Itzkowitz SH, Parsons RE. Molecular mechanisms in colitis-associated colorectal cancer. Oncogenesis. 2023;12(1):48. doi: 10.1038/s41389-023-00492-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Xue J, Deng J, Qin H, et al. The interaction of platelet-related factors with tumor cells promotes tumor metastasis. J Transl Med. 2024;22(1):371. doi: 10.1186/s12967-024-05126-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Contursi A, Tacconelli S, Di Berardino S, De Michele A, Patrignani P. Platelets as crucial players in the dynamic interplay of inflammation, immunity, and cancer: unveiling new strategies for cancer prevention. Front Pharmacol. 2024;15:1520488. doi: 10.3389/fphar.2024.1520488 [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Ferkel SAM, Holman EA, Sojwal RS, Rubin SJS, Rogalla S. Tumor-Infiltrating Immune Cells in Colorectal Cancer. Neoplasia. 2025;59:101091. doi: 10.1016/j.neo.2024.101091 [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Hou S, Zhao Y, Chen J, Lin Y, Qi X. Tumor-associated macrophages in colorectal cancer metastasis: molecular insights and translational perspectives. J Transl Med. 2024;22(1):62. doi: 10.1186/s12967-024-04856-x [DOI] [PMC free article] [PubMed] [Google Scholar]
35.He Q, Xiang L, Luo Y, et al. Tumor-associated macrophages in colon cancer immunotherapy: mechanisms, natural product interventions, and microenvironment remodeling. Front Immunol. 2025;16:1642091. doi: 10.3389/fimmu.2025.1642091 [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Masui H, Kawada K, Obama K. Neutrophil and Colorectal Cancer. Ijms. 2024;26(1):6. doi: 10.3390/ijms26010006 [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Zheng W, Wu J, Peng Y, Sun J, Cheng P, Huang Q. Tumor-Associated Neutrophils in Colorectal Cancer Development, Progression and Immunotherapy. Cancers. 2022;14(19):4755. doi: 10.3390/cancers14194755 [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Arianmanesh F, Bagheri S, Karimi MA, Izadi S, Ahmadi MH. The Evaluation of Diagnostic, Prognostic, and Predictive Role of Hematologic Inflammatory Indices NLR, PLR, and LMR in Common Solid Tumors. Cancer Rep. 2025;8(12):e70407. doi: 10.1002/cnr2.70407 [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Rezazadeh M, Kamyabi A, Pisheh RG, et al. Diagnostic value of peripheral blood inflammatory indices for clinicopathological profile of colorectal cancer: a retrospective observational study. BMC Gastroenterol. 2025;25(1):127. doi: 10.1186/s12876-025-03681-w [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Hernandez-Ainsa M, Velamazan R, Lanas A, Carrera-Lasfuentes P, Piazuelo E. Blood-Cell-Based Inflammatory Markers as a Useful Tool for Early Diagnosis in Colorectal Cancer. Front Med. 2022;9:843074. doi: 10.3389/fmed.2022.843074 [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Güzelyüz Bİ, Ünlü U, Koca B, YeniOva AÖ, Köseoğlu RD. The predictive value of complete blood count and the systemic immune-inflammation index (SII) for diagnosing malignancy in patients undergoing colonoscopy. BMC Prim Care. 2026;27(1):80. doi: 10.1186/s12875-026-03189-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Sun K, Wang Y, Qu R, et al. Comprehensive application of artificial intelligence in colorectal cancer: a review. iScience. 2025;28(7):112980. doi: 10.1016/j.isci.2025.112980 [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Merabet A, Saighi A, Saad H, et al. AI for colon cancer: a focus on classification, detection, and predictive modeling. Int J Med Inform. 2026;206:106115. doi: 10.1016/j.ijmedinf.2025.106115 [DOI] [PubMed] [Google Scholar]
44.Zhen J, Li J, Liao F, et al. Development and validation of machine learning models for young-onset colorectal cancer risk stratification. Npj Precis Onc. 2024;8(1):239. doi: 10.1038/s41698-024-00719-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Zhang X, Tong X, Mou J, et al. ColoLDB: a machine learning-based predictive model for colorectal cancer using routine laboratory parameters. J Gastrointest Oncol. 2026;17(1):12. doi: 10.21037/jgo-2025-611 [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Caterson J, Lewin A, Williamson E. The application of explainable artificial intelligence (XAI) in electronic health record research: a scoping review. Digital health. 2024;10:20552076241272657. doi: 10.1177/20552076241272657 [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Hettikankanamage N, Shafiabady N, Chatteur F, Rmx W, Ud Din F, Zhou J. eXplainable Artificial Intelligence (XAI): a Systematic Review for Unveiling the Black Box Models and Their Relevance to Biomedical Imaging and Sensing. Sensors. 2025;25(21):6649. doi: 10.3390/s25216649 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

Data can be obtained from the corresponding author upon reasonable request.

[cit0001] 1.Force LM, Kocarnik JM, May ML. et al. The global, regional, and national burden of cancer, 1990–2023, with forecasts to 2050: a systematic analysis for the global burden of disease study 2023. Lancet. 2025;406(10512):1565–22. doi: 10.1016/S0140-6736(25)01635-6 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0002] 2.Shin AE, Giancotti FG, Rustgi AK. Metastatic colorectal cancer: mechanisms and emerging therapeutics. Trends Pharmacol Sci. 2023;44(4):222–236. doi: 10.1016/j.tips.2023.01.003 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0003] 3.Chen Y, Zhang J, Ding Y, Zhu F, Chen Y. Colorectal cancer pathogenesis, oncogenic signaling networks and targeted therapeutic advances. Mol Biomed. 2026;7(1):32. doi: 10.1186/s43556-026-00433-4 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0004] 4.Krul MF, Elferink MAG, Kok NFM, et al. Initial impact of national CRC screening on incidence and advanced colorectal cancer. Clin Gastroenterol Hepatol. 2023;21(3):797–807.e3. doi: 10.1016/j.cgh.2022.08.046 [DOI] [PubMed] [Google Scholar]

[cit0005] 5.Shi J, Løberg M, Kalager M, et al. Effect of colonoscopy screening on risks of colorectal cancer and related death: instrumental variable estimation of per-protocol effects. Eur J Epidemiol. 2025;40(4):419–425. doi: 10.1007/s10654-025-01209-w [DOI] [PubMed] [Google Scholar]

[cit0006] 6.Ola I, Cardoso R, Hoffmeister M, Brenner H. Utilization of colorectal cancer screening tests: a systematic review and time trend analysis of nationally representative data. Eclinicalmedicine. 2024;75:102783. doi: 10.1016/j.eclinm.2024.102783 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0007] 7.Westerberg M, Ludvigsson JF, Metcalfe C, et al. Colonoscopy and fecal immunochemical testing versus usual care in diagnostic colorectal cancer screening: the SCREESCO randomized controlled trial. Nat Med. 2026;2026:1. doi: 10.1038/s41591-026-04225-9 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0008] 8.Levy BT, Xu Y, Daly JM, et al. Comparative performance of common fecal immunochemical tests: a cross-sectional study. Ann Intern Med. 2024;177(10):1350–1360. doi: 10.7326/M24-0080 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0009] 9.Lech G. Colorectal cancer tumour markers and biomarkers: recent therapeutic advances. WJG. 2016;22(5):1745. doi: 10.3748/wjg.v22.i5.1745 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0010] 10.Schmitt M, Greten FR. The inflammatory pathogenesis of colorectal cancer. Nat Rev Immunol. 2021;21(10):653–667. doi: 10.1038/s41577-021-00534-x [DOI] [PubMed] [Google Scholar]

[cit0011] 11.Bardelčíková A, Šoltys J, Mojžiš J. Oxidative stress, inflammation and colorectal cancer: an overview. Antioxidants. 2023;12(4):901. doi: 10.3390/antiox12040901 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0012] 12.Burgos-Molina AM, Téllez Santana T, Redondo M, Bravo Romero MJ. The crucial role of inflammation and the immune system in colorectal cancer carcinogenesis: a comprehensive perspective. Int J Mol Sci. 2024;25(11):6188. doi: 10.3390/ijms25116188 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0013] 13.Wang J, Xiao H, Cui S, Wu C, Xiang D. The bridging role of neutrophils in the progression of inflammation-induced colorectal cancer. Biochim Biophys Acta Rev Cancer. 2025;1880(6):189460. doi: 10.1016/j.bbcan.2025.189460 [DOI] [PubMed] [Google Scholar]

[cit0014] 14.Sylman JL, Mitrugno A, Atallah M, et al. The predictive value of inflammation-related peripheral blood measurements in cancer staging and prognosis. Front Oncol. 2018;8:78. doi: 10.3389/fonc.2018.00078 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0015] 15.Yu L, Yuan J, Lou S, et al. Factors affecting survival in patients with colorectal cancer: an umbrella review. J Transl Med. 2025;23(1):1185. doi: 10.1186/s12967-025-06876-7 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0016] 16.Yamamoto T, Kawada K, Obama K. Inflammation-Related Biomarkers for the Prediction of Prognosis in Colorectal Cancer Patients. Ijms. 2021;22(15):8002. doi: 10.3390/ijms22158002 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0017] 17.Menyhart O, Fekete JT, Győrffy B. Inflammation and colorectal cancer: a meta-analysis of the prognostic significance of the systemic immune–inflammation index (SII) and the systemic inflammation response index (SIRI). Int J Mol Sci. 2024;25(15):8441. doi: 10.3390/ijms25158441 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0018] 18.Yan L, Yi B, Fang Y. From biomarkers to mechanisms: deepening the clinical translation of inflammatory indices in colorectal cancer pathogenesis. Int J Surg. 2026;112(1):2076–2077. doi: 10.1097/JS9.0000000000003405 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0019] 19.Zhou M, Gu Q, Zhou M, et al. Extensive study on the associations of 12 composite inflammatory indices with colorectal cancer risk and mortality: a cross-sectional analysis of NHANES 2001–2020. Int J Surg. 2025;111(11):7559–7575. doi: 10.1097/JS9.0000000000002996 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0020] 20.Wang J, Zeng Z, Li Z, et al. The clinical application of artificial intelligence in cancer precision treatment. J Transl Med. 2025;23(1):120. doi: 10.1186/s12967-025-06139-5 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0021] 21.Zhang C, Xu J, Tang R, et al. Novel research and future prospects of artificial intelligence in cancer diagnosis and treatment. J Hematol Oncol. 2023;16(1):114. doi: 10.1186/s13045-023-01514-5 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0022] 22.Liu K, Luo Z, Zhang W, et al. Artificial intelligence models: transforming early diagnosis and precise treatment of gastrointestinal cancers. Mol Cancer. 2026;2026:1. doi: 10.1186/s12943-026-02589-7 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0023] 23.Allgaier J, Mulansky L, Draelos RL, Pryss R. How does the model make predictions? A systematic literature review on the explainability power of machine learning in healthcare. Artif Intell Med. 2023;143:102616. doi: 10.1016/j.artmed.2023.102616 [DOI] [PubMed] [Google Scholar]

[cit0024] 24.Toumaj S, Heidari A, Jafari Navimipour N. Leveraging explainable artificial intelligence for transparent and trustworthy cancer detection systems. Artif Intell Med. 2025;169:103243. doi: 10.1016/j.artmed.2025.103243 [DOI] [PubMed] [Google Scholar]

[cit0025] 25.Zabor EC, Reddy CA, Tendulkar RD, Patil S. Logistic Regression in Clinical Studies. Int J Radiat Oncol Biol Phys. 2022;112(2):271–277. doi: 10.1016/j.ijrobp.2021.08.007 [DOI] [PubMed] [Google Scholar]

[cit0026] 26.Rodríguez-Pérez R, Vogt M, Bajorath J. Support vector machine classification and regression prioritize different structural features for binary compound activity and potency value prediction. ACS Omega. 2017;2(10):6371–6379. doi: 10.1021/acsomega.7b01079 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0027] 27.Maleki F, Ovens K, Najafian K, Forghani B, Reinhold C, Forghani R. Overview of machine learning part 1. Neuroimaging Clin N Am. 2020;30(4):e17–e32. doi: 10.1016/j.nic.2020.08.007 [DOI] [PubMed] [Google Scholar]

[cit0028] 28.Eloranta S, Boman M. Predictive models for clinical decision making: deep dives in practical machine learning. J Intern Med. 2022;292(2):278–295. doi: 10.1111/joim.13483 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0029] 29.Lasry A, Zinger A, Ben-Neriah Y. Inflammatory networks underlying colorectal cancer. Nat Immunol. 2016;17(3):230–240. doi: 10.1038/ni.3384 [DOI] [PubMed] [Google Scholar]

[cit0030] 30.Zhou RW, Harpaz N, Itzkowitz SH, Parsons RE. Molecular mechanisms in colitis-associated colorectal cancer. Oncogenesis. 2023;12(1):48. doi: 10.1038/s41389-023-00492-0 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0031] 31.Xue J, Deng J, Qin H, et al. The interaction of platelet-related factors with tumor cells promotes tumor metastasis. J Transl Med. 2024;22(1):371. doi: 10.1186/s12967-024-05126-6 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0032] 32.Contursi A, Tacconelli S, Di Berardino S, De Michele A, Patrignani P. Platelets as crucial players in the dynamic interplay of inflammation, immunity, and cancer: unveiling new strategies for cancer prevention. Front Pharmacol. 2024;15:1520488. doi: 10.3389/fphar.2024.1520488 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0033] 33.Ferkel SAM, Holman EA, Sojwal RS, Rubin SJS, Rogalla S. Tumor-Infiltrating Immune Cells in Colorectal Cancer. Neoplasia. 2025;59:101091. doi: 10.1016/j.neo.2024.101091 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0034] 34.Hou S, Zhao Y, Chen J, Lin Y, Qi X. Tumor-associated macrophages in colorectal cancer metastasis: molecular insights and translational perspectives. J Transl Med. 2024;22(1):62. doi: 10.1186/s12967-024-04856-x [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0035] 35.He Q, Xiang L, Luo Y, et al. Tumor-associated macrophages in colon cancer immunotherapy: mechanisms, natural product interventions, and microenvironment remodeling. Front Immunol. 2025;16:1642091. doi: 10.3389/fimmu.2025.1642091 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0036] 36.Masui H, Kawada K, Obama K. Neutrophil and Colorectal Cancer. Ijms. 2024;26(1):6. doi: 10.3390/ijms26010006 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0037] 37.Zheng W, Wu J, Peng Y, Sun J, Cheng P, Huang Q. Tumor-Associated Neutrophils in Colorectal Cancer Development, Progression and Immunotherapy. Cancers. 2022;14(19):4755. doi: 10.3390/cancers14194755 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0038] 38.Arianmanesh F, Bagheri S, Karimi MA, Izadi S, Ahmadi MH. The Evaluation of Diagnostic, Prognostic, and Predictive Role of Hematologic Inflammatory Indices NLR, PLR, and LMR in Common Solid Tumors. Cancer Rep. 2025;8(12):e70407. doi: 10.1002/cnr2.70407 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0039] 39.Rezazadeh M, Kamyabi A, Pisheh RG, et al. Diagnostic value of peripheral blood inflammatory indices for clinicopathological profile of colorectal cancer: a retrospective observational study. BMC Gastroenterol. 2025;25(1):127. doi: 10.1186/s12876-025-03681-w [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0040] 40.Hernandez-Ainsa M, Velamazan R, Lanas A, Carrera-Lasfuentes P, Piazuelo E. Blood-Cell-Based Inflammatory Markers as a Useful Tool for Early Diagnosis in Colorectal Cancer. Front Med. 2022;9:843074. doi: 10.3389/fmed.2022.843074 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0041] 41.Güzelyüz Bİ, Ünlü U, Koca B, YeniOva AÖ, Köseoğlu RD. The predictive value of complete blood count and the systemic immune-inflammation index (SII) for diagnosing malignancy in patients undergoing colonoscopy. BMC Prim Care. 2026;27(1):80. doi: 10.1186/s12875-026-03189-6 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0042] 42.Sun K, Wang Y, Qu R, et al. Comprehensive application of artificial intelligence in colorectal cancer: a review. iScience. 2025;28(7):112980. doi: 10.1016/j.isci.2025.112980 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0043] 43.Merabet A, Saighi A, Saad H, et al. AI for colon cancer: a focus on classification, detection, and predictive modeling. Int J Med Inform. 2026;206:106115. doi: 10.1016/j.ijmedinf.2025.106115 [DOI] [PubMed] [Google Scholar]

[cit0044] 44.Zhen J, Li J, Liao F, et al. Development and validation of machine learning models for young-onset colorectal cancer risk stratification. Npj Precis Onc. 2024;8(1):239. doi: 10.1038/s41698-024-00719-2 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0045] 45.Zhang X, Tong X, Mou J, et al. ColoLDB: a machine learning-based predictive model for colorectal cancer using routine laboratory parameters. J Gastrointest Oncol. 2026;17(1):12. doi: 10.21037/jgo-2025-611 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0046] 46.Caterson J, Lewin A, Williamson E. The application of explainable artificial intelligence (XAI) in electronic health record research: a scoping review. Digital health. 2024;10:20552076241272657. doi: 10.1177/20552076241272657 [DOI] [PMC free article] [PubMed] [Google Scholar]

[cit0047] 47.Hettikankanamage N, Shafiabady N, Chatteur F, Rmx W, Ud Din F, Zhou J. eXplainable Artificial Intelligence (XAI): a Systematic Review for Unveiling the Black Box Models and Their Relevance to Biomedical Imaging and Sensing. Sensors. 2025;25(21):6649. doi: 10.3390/s25216649 [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Interpretable Machine Learning Models Based on Blood Cell-Derived Inflammatory Indices for Identifying Colorectal Neoplasia: A Retrospective Study

Xinya Zeng

Zongshou Li

Lulu Cai

Shengmei Liang

Siyan Wang

Kai Li

Ni Chen

Xiaomei Yang

Shiquan Liu

Abstract

Purpose

Patients and Methods

Results

Conclusion

Introduction

Methods

Study Population

Figure 1.

Data Collection and Definitions

Data Preprocessing

Variable Selection

Model Development and Performance

Model Interpretability

Statistical Analysis

Results

Baseline Characteristics

Table 1.

Table 2.

Comparison of Inflammatory Indices Among HC, CRA, and CRC Groups

Figure 2.

Figure 3.

RCS Analysis of the Associations of Inflammatory Indices with CRC and CRA

Figure 4.

Figure 5.

Feature Selection

Figure 6.

Model Development and Validation

Figure 7.

Table 3.

Table 4.

Figure 8.

SHAP-Based Model Interpretation

Figure 9.

Figure 10.

Figure 11.

Discussion

Conclusions

Acknowledgments

Funding Statement

Abbreviations

Data Sharing Statement

Ethical Statement

Author Contributions

Disclosure

References

Associated Data

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases