Clinical decision support for precolonoscopy cancer triage: A rule-out-oriented machine learning model for colorectal cancer risk

Abstract

BACKGROUND

Colonoscopy is the gold standard for colorectal cancer diagnosis but is invasive and subject to capacity constraints. Noninvasive precolonoscopy triage tools are needed to prioritize high-risk patients and safely manage those with a low likelihood of malignancy.

AIM

To develop an extreme gradient boosting (XGBoost)-based prediction model using routine laboratory parameters to estimate precolonoscopy malignancy risk.

METHODS

This retrospective cohort study included 1604 consecutive patients who underwent colonoscopy at Ankara Training and Research Hospital (January 2022 to December 2025). The predictors included age, sex, complete blood count components, liver enzymes, lipid profiles, fasting glucose levels, and derived inflammatory/metabolic indices (neutrophil to lymphocyte ratio, systemic immune-inflammation index, neutrophil-to-high-density lipoprotein ratio, platelet to lymphocyte ratio, triglyceride-glucose, atherogenic index of plasma, and Fibrosis-4). An XGBoost classifier with class-weight adjustment was evaluated using repeated stratified K-fold cross-validation (5 folds, 50 repeats). SHapley Additive exPlanations (SHAP) analysis guided feature reduction to ten-feature (SHAP10) and five-feature (SHAP5) models.

RESULTS

Malignancy was present in 23 patients (1.43%). The full model achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.734 and an area under the precision-recall curve of 0.078; at the Youden-optimized threshold, the positive predictive value was 0.050, and the negative predictive value (NPV) was 0.993. The discrimination ability of the SHAP10 model was comparable (ROC-AUC: 0.736; NPV: 0.997), whereas that of the SHAP5 model was lower (ROC-AUC: 0.711). The key predictors included glucose level, platelet count, neutrophil to lymphocyte ratio, age, neutrophil count, alanine aminotransferase level, aspartate aminotransferase level, systemic immune-inflammation index, neutrophil-to-high-density lipoprotein ratio, and high-density lipoprotein level. After post hoc Platt recalibration, all three models achieved near-ideal calibration (intercept ≈ 0, slope ≈ 1.00).

CONCLUSION

An XGBoost model using routine precolonoscopy laboratory data achieved a very high NPV, supporting its potential as a rule-out-oriented triage tool. A parsimonious ten-feature model preserved discrimination while enhancing clinical applicability. Prospective, multicenter validation is warranted.

Key Words: Colorectal cancer; Pre-colonoscopy triage; Machine learning; Gastroenterology; Precision medicine; Oncology; Decision support; Inflammatory indices

Core Tip: Colonoscopy capacity is limited and most procedures find no malignancy. Using only routinely available pre-colonoscopy laboratory parameters from 1604 consecutive patients (1.43% malignancy prevalence), we developed an extreme gradient boosting-based machine learning model with explainable SHapley Additive exPlanations analysis. Performance was validated through repeated stratified cross-validation, calibration analysis, and decision-curve analysis. The 10-feature SHapley Additive exPlanations-reduced model achieved a high sensitivity of 91% and a very low negative likelihood ratio of 0.17, with a calibrated negative predictive value of 99.7%. This rule-out-oriented tool may safely defer 40 to 67 colonoscopies per 100 patients in resource-limited settings.

INTRODUCTION

Colorectal cancer accounts for approximately 1.93 million new cases, representing 9.6% of all cancers, and is the third most commonly diagnosed cancer worldwide[1]. With approximately 904000 deaths annually, it is the second leading cause of cancer-related mortality. The growing geriatric population and the structure of industrialized societies are associated with colorectal cancer risk factors and exacerbate the drivers that increase disease incidence[2]. Survival outcomes achieved after surgery and adjuvant therapy in the localized stage are significantly better than those in the metastatic stage. Screening programs are critically important for detecting the disease at an early stage.

The United States Preventive Services Task Force recommends annual gFOBT or FIT, flexible sigmoidoscopy every 5 years, and total colonoscopy every 10 years[3]. While screening recommendations are well established for individuals aged 50-75 years, the benefit of screening is less clear for those aged 45-50 years and for adults older than 75 years. In addition, early-onset colorectal cancer occurring outside these age groups and without overt clinical symptoms constitutes a major cause of morbidity and mortality worldwide, and asymptomatic patients who are not included in screening age ranges are often diagnosed at a metastatic stage[4].

Among screening tests, gFOBT and FIT are easy-to-apply, noninvasive methods; however, their false-positive rates are relatively high, which often leads to referrals for colonoscopy[5]. Although colonoscopy has high sensitivity and specificity, it is an invasive procedure and is operator dependent[6]. Colonoscopic screening is typically performed by gastroenterologists; therefore, in settings where endoscopy capacity and clinician resources are insufficient, triage and decision-support approaches (e.g., FIT-based pathways) are needed to prioritize urgent/high-risk cases and safely manage low-risk patients who are unlikely to have severe disease[7-9].

Patients who present with alarm symptoms such as rectal bleeding, iron deficiency anemia, weight loss, and changes in bowel habits are typically prioritized for urgent colonoscopic evaluation; however, these findings can also be attributed to benign causes[10]. Symptom-based approaches may lead to diagnostic delay because a proportion of patients with colorectal cancer may be asymptomatic or present with only subtle symptoms, and such patients - who may also be missed by screening programs - can ultimately be diagnosed at an advanced stage.

Various precolonoscopy risk stratification approaches have been developed to estimate the likelihood of malignancy and to ensure timely referral in settings with limited endoscopy resources. Examples include quantitative FIT and multivariable models that combine demographic characteristics (e.g., age/sex), symptoms, examination findings, and routine laboratory parameters (e.g., COLONPREDICT and the simpler FAST score)[11,12]. Machine learning-based models have also been investigated for this purpose[13-15]. Nevertheless, no guideline-endorsed test to determine precolonoscopy malignancy risk is currently available.

Machine learning methods can leverage complex, high-dimensional clinical data to generate individualized risk estimates that are difficult to capture with rule-based approaches. In medicine, these models are being increasingly used to support decision-making by enabling early risk stratification, triage, and more efficient use of limited diagnostic resources. In this context, we aimed to develop an XGBoost-based precolonoscopy malignancy prediction model using routinely obtainable variables. This approach may help prioritize high-risk patients for expedited colonoscopy while supporting the safe management of patients with a low likelihood of malignancy.

MATERIALS AND METHODS

Study design, setting, and period

This retrospective cohort study included consecutive patients who presented to the gastroenterology outpatient clinic at Ankara Research and Training Hospital between January 1, 2022, and December 31, 2025.

Ethics approval

Data were obtained from the colonoscopy cohort established at Ankara Training and Research Hospital under the approval of the Institutional Ethics Committee (approval No. E-25/629, dated September 24, 2025). The study was conducted in accordance with the Declaration of Helsinki. Given the retrospective design and the use of deidentified data, the requirement for informed consent was waived by the Ethics Committee.

Data source and variables

Demographic variables (age and sex) and colonoscopy findings were extracted from the institutional retrospective dataset. Precolonoscopy laboratory parameters and derived metabolic/inflammatory indices were used as candidate predictors. Data were entered by a single investigator using a standardized data collection form and subsequently anonymized; all direct personal identifiers were removed and excluded from analysis. Statistical analyses were performed by an independent investigator who was blinded to patient identifiers.

Derived indices from biochemical parameters

In addition to raw precolonoscopy laboratory values, several composite metabolic and inflammatory indices were calculated from routine biochemical and complete blood count parameters. Specifically, the triglyceride-glucose index was derived from fasting glucose and triglyceride levels; the atherogenic index of plasma was derived from triglyceride and high-density lipoprotein (HDL) cholesterol levels; the Fibrosis-4 score was calculated using age, aspartate aminotransferase (AST) level, alanine aminotransferase (ALT) level, and platelet count; and inflammatory ratios/indices, including the neutrophil to lymphocyte ratio (NLR), platelet to lymphocyte ratio (PLR), systemic immune-inflammation index (SII), and neutrophil-to-high-density lipoprotein ratio (NHR), were computed from neutrophil, lymphocyte, platelet, and HDL measurements. These derived indices were included as candidate predictors alongside the original laboratory variables.

Model development and validation

We developed a precolonoscopy prediction model using an extreme gradient boosting (XGBoost) classifier. Missing predictor values were handled using median imputation fit within each training split. Given the substantial class imbalance, class weights were applied during training using the negative-to-positive ratio. All analyses were implemented in Python 3 using the scikit-learn and XGBoost libraries. Model performance was assessed using repeated stratified K-fold cross-validation (5 folds, 50 repeats). For each split, predicted probabilities were generated for the held-out fold, yielding out-of-fold (OOF) predictions for all individuals. Discrimination was quantified using the area under the receiver operating characteristic curve (ROC-AUC) and area under the precision-recall curve (PR-AUC) (average precision), with the latter emphasized because of the low outcome prevalence.

Threshold selection and diagnostic metrics

A single operating threshold for each model was selected using the Youden index (J = sensitivity + specificity - 1) calculated on OOF predictions. At the Youden threshold, the positive predictive value (PPV) and negative predictive value (NPV) were reported.

Explainability and feature reduction

Model explainability was evaluated using SHapley Additive exPlanations (SHAP). To explore parsimonious models, reduced predictor sets were derived on the basis of SHAP importance (top-10 and top-5 features), and their performances were compared with those of the full model.

Calibration and clinical utility analyses

Model calibration was evaluated on the basis of the OOF predictions in four complementary ways. First, the Cox (1958) calibration intercept (“calibration-in-the-large”), fitted as a binomial GLM on the outcome with the OOF logits as offset, was reported with 95% Wald confidence intervals (CIs); an intercept of zero indicates the absence of systematic bias. Second, the calibration slope (binomial GLM of the outcome on the OOF logits) was reported with 95% Wald CIs; a slope of one indicates correctly scaled predictions, whereas values below one indicate overdispersed (too-extreme) probability estimates. Third, a quintile (5-group) calibration plot was produced with per-bin n and event counts annotated, and Wilson 95%CIs were calculated for the observed event rate in each bin; five bins were chosen to retain a nonzero number of events per bin given that there were only 23 observed events. Fourth, as a sensitivity analysis, Platt scaling (a single logistic regression on the OOF logits) was applied, and the calibration intercept, slope, and Brier score were recomputed on the Platt-recalibrated probabilities.

Clinical utility was assessed with decision curve analysis (Vickers and Elkin 2006), which reported both the standard net benefit across threshold probabilities pt ∈ [0.001, 0.10] and a rule-out-oriented yield metric defined as TN per 100 - [(1 - pt)/pt] × FN per 100. Likelihood ratios (LR+, LR-) and the diagnostic odds ratio were calculated at the Youden-optimized threshold; their 95%CIs were derived from 2000 stratified bootstrap resamples of the OOF predictions. Posttest probabilities across five plausible pretest prevalences (1.43%, 3%, 5%, 10%, and 20%) were derived from the Youden-threshold LR+ and LR- using Bayes’ rule. All additional analyses were performed in Python 3.10 using scikit-learn 1.4, statsmodels 0.14, and XGBoost 2.0.

RESULTS

Patient characteristics and colonoscopy findings

The cohort comprised 1604 patients. The median patient age was 61 years (interquartile range: 51-70); a total of 774 (48.2%) were male, and 830 (51.7%) were female (Table 1). The colonoscopy/pathology findings recorded in the dataset included malignancy in 23 patients (1.4%), diverticulosis in 52 (3.2%), benign ulcers in 33 (2.1%), and normal colonoscopy in 995 (62.0%) (Table 1). The precolonoscopy laboratory parameters stratified by malignancy status, derived inflammatory/metabolic indices, and summary of the variable-wise missing data are presented in detail in Supplementary Tables 1 and 2. All patients diagnosed with colorectal cancer had adenocarcinoma pathology.

Table 1 Demographics and colonoscopy findings, n (%).

Variable	Overall (n = 1604)	Malignancy+ (n = 23)	Malignancy- (n = 1581)
Demographics
Age, years, median (IQR)	61 (51-70)	67 (60-74)	61 (51-70)
Male	774 (48.3)	13 (56.5)	761 (48.1)
Female	830 (51.7)	10 (43.5)	820 (51.9)
Colonoscopy findings1
Polyp present	523 (32.6)	-	-
Malignancy present	23 (1.4)	-	-
Diverticulosis	52 (3.2)	-	-
Benign ulcer	33 (2.1)	-	-
Normal colonoscopy	995 (62.0)	-	-
Among polyp-positive patients (n = 523)
Adenoma present	365 (69.8)	-	-

¹Colonoscopy findings are not mutually exclusive; 30 patients had ≥ 2 concurrent findings. Eight patients had non-standard findings (operated colon, ulcerative colitis in remission, or unrecorded) and were retained with malignancy = 0. Adenoma percentage is calculated among the 523 polyp-positive patients, not the full cohort. IQR: Interquartile range.

Model performance (repeated stratified CV; OOF-average)

Across repeated stratified cross-validation, the full model achieved an ROC-AUC of 0.734 and a PR-AUC of 0.078 (Figures 1 and 2). The discrimination ability of the SHAP10 model was similar (ROC-AUC: 0.736; PR-AUC: 0.077), whereas that of the SHAP5 model was lower (ROC-AUC: 0.711; PR-AUC: 0.055) (Table 2). At the Youden-selected threshold, the PPV remained modest, and the NPV remained high across the models, which is consistent with the very low malignancy prevalence (1.43%) (Table 2). Full model: Youden cutoff 0.004597 → PPV 0.050, NPV 0.993; SHAP10: Youden cutoff 0.000485 → PPV 0.026, NPV 0.997; SHAP5: Youden cutoff 0.002415 → PPV 0.028, NPV 0.994.

Figure 1 Receiver operating characteristic curves for malignancy prediction. Receiver operating characteristic curves based on out-of-fold predicted probabilities from repeated stratified 5-fold cross-validation (50 repeats) comparing the full-feature model, SHAP10, and SHAP5 models. Area under the curve values are reported in the text/Table 2. ROC: Receiver operating characteristic curve; CV: Cross-validation; OOF: Out-of-fold; AUC: Area under the curve; SHAP: SHapley Additive exPlanations; FPR: False-positive rate.

Figure 2 Precision-recall curves based on out-of-fold predicted probabilities from repeated stratified cross-validation. The dashed horizontal line indicates the baseline precision equal to the outcome prevalence (1.43%). Area under the precision-recall curve values are reported in the text/Table 2. PR: Precision-recall; SHAP: SHapley Additive exPlanations.

Table 2 Discrimination and classification performance at the Youden-optimised threshold (repeated stratified 5-fold × 50-repeat cross-validation; 95% confidence intervals from 2000 bootstrap resamples).

Model	Youden cut-off	ROC-AUC (95%CI)	PR-AUC (95%CI)	Brier score (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)	PPV (95%CI)	NPV (95%CI)
All features	0.0046	0.734 (0.637-0.835)	0.078 (0.022-0.191)	0.0152 (0.0099-0.0208)	0.565 (0.455-1.000)	0.846 (0.388-0.905)	0.051 (0.017-0.086)	0.993 (0.989-1.000)
SHAP10	0.0005	0.736 (0.642-0.826)	0.076 (0.020-0.183)	0.0153 (0.0098-0.0208)	0.913 (0.538-1.000)	0.503 (0.485-0.872)	0.026 (0.018-0.067)	0.997 (0.991-1.000)
SHAP5	0.0025	0.711 (0.597-0.812)	0.055 (0.023-0.152)	0.0154 (0.0101-0.0211)	0.739 (0.400-0.963)	0.629 (0.357-0.944)	0.028 (0.017-0.087)	0.994 (0.989-0.999)

All metrics are computed on out-of-fold predictions pooled across 250 cross-validation folds (5 × 50). The Youden cut-off (maximum sensitivity + specificity - 1) is reported as a statistical reference only; clinically relevant operating points (pt = 0.02-0.05) are discussed in the decision curve analysis. CI: Confidence interval; ROC-AUC: Area under the receiver operating characteristic curve; PR-AUC: Area under the precision-recall curve; PPV: Positive predictive value; NPV: Negative predictive value; SHAP: SHapley Additive exPlanations.

Across repeated resampling, the most consistently selected predictors in the SHAP top-10 set included glucose, platelet count, NLR, age, neutrophil count, ALT, AST, SII, NHR, and HDL, supporting the clinical relevance of combined metabolic and inflammatory signals for precolonoscopy malignancy risk stratification. The data in Figure 3 (SHAP summary plot) indicate that the model’s malignancy predictions are driven primarily by a combination of metabolic and inflammatory signals. Higher values of glucose, platelet count, SII, age, AST, and NLR generally shift the model output toward a higher predicted probability of malignancy (positive SHAP direction), whereas higher values of HDL, ALT, NHR, and neutrophil count in this model generally shift predictions toward a lower malignancy probability (negative SHAP direction).

Figure 3 Beeswarm plot showing the distribution of SHapley Additive exPlanations values for the top-10 predictors (English labels). Each dot represents an individual patient; X-axis indicates SHapley Additive exPlanations value (impact on model output), and color denotes the feature value (low to high). SHAP: SHapley Additive exPlanations; NLR: Neutrophil to lymphocyte ratio; HDL: High-density lipoprotein; SII: Systemic immune-inflammation index; NHR: Neutrophil-to-high-density lipoprotein ratio; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase.

Calibration

The results of the calibration analysis revealed that the uncalibrated class-weighted XGBoost models rank-ordered patients correctly but produced probability estimates that were systematically overdispersed (Table 3). The calibration-in-the-large was acceptable for SHAP10 (intercept +0.363, 95%CI: -0.102 to +0.827) and SHAP5 (-0.042, 95%CI: -0.493 to +0.409) and modestly positive for the full model (+0.654, 95%CI: +0.174 to +1.133). The calibration slope, however, was consistently and substantially less than one for all three models (range: 0.31-0.33), indicating that the raw XGBoost probabilities are too extreme relative to the underlying event rates - a well-recognized consequence of training tree ensembles under class weighting for rare outcomes.

Table 3 Calibration performance before and after post-hoc Platt scaling (Cox 1958 framework; fitted on out-of-fold predictions).

Model	Uncalibrated (raw XGBoost output)			After Platt recalibration
	Intercept (95%CI)	Slope (95%CI)	Brier score	Intercept	Slope	Brier score
All features	+0.654 (+0.174 to +1.133)	0.329 (0.171-0.487)	0.0152	0.000	1.003	0.0140
SHAP10	+0.363 (-0.102 to +0.827)	0.315 (0.157-0.472)	0.0153	0.000	1.006	0.0140
SHAP5	-0.042 (-0.493 to +0.409)	0.314 (0.147-0.480)	0.0154	-0.001	1.006	0.0140

The calibration intercept (“calibration-in-the-large”) was obtained as the intercept of a binomial generalised linear model fitted on the observed outcome with the out-of-fold logit predictions as offset; an intercept of 0 indicates no systematic bias. The calibration slope was obtained as the coefficient of a binomial GLM regressing the outcome on the out-of-fold logit; a slope of 1 indicates well-calibrated predictions, whereas values below 1 indicate over-dispersed (too-extreme) predicted probabilities. Platt scaling was applied as a single logistic regression on the out-of-fold logits and restored near-ideal calibration (intercept ≈ 0, slope ≈ 1) for all three models without loss of discrimination, reducing the Brier score uniformly to 0.0140. CI: Confidence interval; XGBoost: Extreme gradient boosting; SHAP: SHapley Additive exPlanations.

Post hoc Platt scaling corrected both parameters to near-ideal values in all three models (intercept ≈ 0; slope: 1.003-1.006) and uniformly improved the Brier score from 0.0152-0.0154 to 0.0140 (Table 3), confirming that the rank ordering of the OOF probabilities carries the full discriminative signal and that a simple fixed recalibration layer is sufficient for clinical deployment. Quintile calibration plots before and after Platt scaling (Supplementary Figures 1 and 2) revealed a clear monotonic gradient of the observed event rate across the predicted probability quintiles, confirming that the model stratifies genuine high- and low-risk patients rather than producing a constant probability.

Clinical utility: Decision curve analysis and likelihood ratios

Decision curve analysis (Figure 4) demonstrated that all three models provided a positive net benefit over both the “refer all” and “refer none” strategies across the clinically relevant threshold range (pt = 0.02-0.05). In the rule-out-oriented yield framing (Figure 4, panel B), the SHAP10 model safely avoided approximately 40.8 colonoscopies per 100 patients at pt = 0.02, which increased to 59.4 at pt = 0.03 and 67.5 at pt = 0.05 - net of the harm weight applied to missed cancers.

Figure 4 Decision curve analysis. Decision curve analysis based on out-of-fold predicted probabilities from repeated stratified cross-validation. A: Standard net benefit curves across threshold probabilities, comparing the full-feature, SHAP10, and SHAP5 models with “refer all” and “refer none” strategies; B: Rule-out-oriented clinical yield expressed as net avoided colonoscopies per 100 patients, calculated as true negatives gained minus harm-weighted false negatives. OOF: Out-of-fold; SHAP: SHapley Additive exPlanations; TN: True negative; FN: False negative.

Likelihood ratios at the Youden threshold (Table 4) were prevalence independent and therefore directly addressed the concern that the high NPVs reported in Table 2 might simply reflect the low malignancy prevalence. The SHAP10 model achieved the most favorable rule-out profile, with an LR- = 0.173 (95%CI: 0.000-0.445) and a diagnostic odds ratio of 10.6; the full model achieved the highest LR+ of 3.66 (95%CI: 2.34-5.19). Across pretest prevalence scenarios ranging from 1.43% to 20%, the posttest probability of malignancy after a negative SHAP10 result remained below 4.15%, corresponding to a 79%-83% reduction relative to the pretest probability (Supplementary Table 3). Taken together, the DCA yield, the LR-, and the Fagan posttest probability table confirm that the model adds clinically meaningful information beyond the prevalence floor.

Table 4 Likelihood ratios and diagnostic odds ratios at the Youden-optimised threshold (95% confidence intervals from 2000 bootstrap resamples).

Model	Threshold	Sensitivity (95%CI)	Specificity (95%CI)	LR+ (95%CI)	LR- (95%CI)	DOR
All features	0.0046	0.565 (0.364-0.783)	0.846 (0.828-0.863)	3.66 (2.34-5.19)	0.514 (0.257-0.752)	7.1
SHAP10	0.0005	0.913 (0.778-1.000)	0.503 (0.476-0.528)	1.84 (1.55-2.07)	0.173 (0.000-0.445)	10.6
SHAP5	0.0025	0.739 (0.538-0.909)	0.629 (0.605-0.652)	1.99 (1.45-2.46)	0.415 (0.149-0.735)	4.8

Likelihood ratios are prevalence-independent and therefore provide evidence that the model’s performance is not simply a mathematical consequence of the low malignancy prevalence (1.43%). The SHAP10 negative likelihood ratio of 0.173 indicates that a negative model result reduces the post-test odds of malignancy by approximately 5.8-fold, consistent with a rule-out-oriented triage role. CI: Confidence interval; LR+: Positive likelihood ratio = sensitivity/(1 - specificity); LR-: Negative likelihood ratio = (1 - sensitivity)/specificity; DOR: Diagnostic odds ratio. SHAP: SHapley Additive exPlanations.

DISCUSSION

Colonoscopy is the gold standard for the diagnosis of colorectal cancer; however, it is invasive, operator dependent, and resource intensive and has notable risks of complications such as perforation and bleeding[16]. Global screening recommendations and the inability to exclude malignancy in patients who present with ambiguous symptoms have led to an increased demand for colonoscopy[8]. The inability to meet this growing demand results in prolonged waiting times, which may lead to diagnostic delays and detection at more advanced stages. For a cancer such as colorectal cancer, where favorable outcomes are achieved when it is diagnosed at an early stage, such delays may ultimately contribute to increased mortality[7,9].

Baron et al[8] reported that approximately 28% of colonoscopy referrals in an open-access endoscopy system were inappropriate and emphasized that guideline-based triage could optimize resource utilization. Furthermore, models developed using hematological and biochemical parameters have been shown to reduce unnecessary colonoscopies by approximately 30%, thereby providing substantial benefits in terms of health care costs and patient comfort[17,18]. In this context, there is a clear clinical need for reliable, noninvasive, and cost-effective precolonoscopy decision-support tools capable of accurately stratifying patients according to their risk of malignancy.

Our study aims to evaluate the performance of a machine learning-based model that utilizes clinical, hematological, and biochemical parameters identified during routine assessment to classify the presence of malignancy risk prior to colonoscopy. In terms of developing the precolonoscopy prediction score, XGBoost was preferred because of its ability to model nonlinear relationships and complex interactions among variables with high accuracy, its robustness to missing data, and its strong generalization performance.

The model achieved an ROC-AUC of 0.734 and, more importantly, for a rule-out-oriented tool, a NPV of 0.993 at the Youden-optimized threshold. The SHAP-based reduced model retaining only SHAP10 preserved comparable discrimination (ROC-AUC: 0.736) while achieving an NPV of 0.997, suggesting that a parsimonious set of metabolic and inflammatory markers can capture most of the predictive signal contained in the full feature space.

These findings indicate that a parsimonious set of metabolic and inflammatory markers can capture most of the predictive signal present in the full feature space. Our results suggest that, rather than serving as a diagnostic tool for detecting malignancy, the model may be positioned as a potential clinical decision-support tool aimed at identifying patients with a very low probability of colorectal cancer, thereby enabling the safe deferral or deprioritization of colonoscopy in resource-limited settings.

Several precolonoscopy risk stratification models have been developed. The COLONPREDICT model, which integrates the fecal hemoglobin concentration, clinical symptoms, physical examination findings, and serum markers (including carcinoembryonic antigen and blood hemoglobin), achieved an AUC of 0.92 for colorectal cancer detection in symptomatic patients[11]. The simpler FAST score, which is based solely on the fecal hemoglobin concentration, age, and sex, had an AUC of 0.88-0.91 in the derivation and validation cohorts[12]. A recent systematic review of FIT-based and non-FIT-based prediction models in symptomatic patients confirmed that models combining FIT with demographic and clinical variables consistently outperform symptom-only referral criteria, such as the NICE guidelines[19]. However, models relying on fecal hemoglobin levels are limited to settings where FIT is routinely performed prior to referral to colonoscopy, and FIT-negative cancers remain a recognized diagnostic gap[20]. Our model differs from these approaches in that it uses exclusively routine precolonoscopy blood-based parameters - complete blood count components, liver enzymes, lipid profiles, glucose, and derived inflammatory/metabolic indices - without requiring fecal testing, making it applicable to any clinical setting where a standard laboratory workup is available.

Machine learning and artificial intelligence approaches using routine hematological data for colorectal cancer detection have gained increasing attention[21]. The ColonFlag algorithm, which uses age, sex, and 20 complete blood count parameters, was validated across multiple populations, with c-statistics ranging from 0.736 to 0.82[22]. A recent systematic review confirmed that compared with established screening tools, ColonFlag can detect colorectal cancer before clinical diagnosis, but it has variable performance across populations and low overall sensitivity[23]. Hornbrook et al[15] demonstrated that a machine learning model using sex, age, and complete blood count data achieved an AUC of 0.80 for colorectal cancer detection and identified right-sided cancers more accurately than left-sided lesions. Li et al[24] evaluated five machine learning algorithms using conventional laboratory test data, including liver enzymes, lipid profiles, complete blood counts, and tumor biomarkers, and reported that a logistic regression model achieved the best performance, with an AUC of 0.98, a PPV of 0.92, and an NPV of 0.97, although the case-control design with balanced groups limits generalizability to real-world prevalence settings. More recently, Li et al[25] in 2025 developed XGBoost models using clinical laboratory data that achieved AUCs of 0.966 for differentiating healthy controls from colorectal cancer patients and 0.881 for distinguishing polyp patients from cancer patients, outperforming CEA and FOBT alone. Our study extends this body of work by evaluating model performance in a consecutive cohort with real-world class imbalance (1.43% malignancy prevalence), emphasizing the PR-AUC and NPV as the primary performance metrics suited to rule-out applications in low-prevalence settings.

A distinctive feature of our model is the inclusion of derived inflammatory and metabolic indices alongside raw laboratory values. SHAP analysis revealed the NLR, SII, neutrophil count, and HDL level as consistently important predictors (Table 5), reflecting the established link between systemic inflammation, metabolic dysregulation, and colorectal carcinogenesis. Inflammation-based biomarkers such as the NLR, PLR, and SII have been extensively studied as prognostic indicators in established colorectal cancer. A meta-analysis of 27 studies demonstrated that an elevated pretreatment SII was significantly associated with poor overall survival, progression-free survival, and disease-free survival in colorectal cancer patients[26]. Passardi et al[27] similarly reported that the NLR, PLR, and SII were significant prognostic and predictive markers in metastatic colorectal cancer patients. While most studies have focused on postdiagnostic prognosis, emerging evidence supports the role of these markers in prediagnostic risk prediction. Güzelyüz Bodur et al[28] in 2026 reported that inflammation-based indices, including the NLR and SII, were higher in patients with malignant colonoscopic findings; however, their discriminatory performance was limited, and their independent associations with malignancy were only modest after adjustment, suggesting that such markers are more informative when they are interpreted alongside other clinical or laboratory variables. Our findings are consistent with this recommendation: The discriminative ability of the model is derived from the combined signal of multiple metabolic and inflammatory features rather than from any single biomarker.

Table 5 Direction of feature effects in the SHapley Additive exPlanations summary.

Feature (Figure 3)	Global direction1	Spearman ρ	High-tail direction2	Pos(SHAP) > 0 in top 20%
Glucose (mg/dL)	Increases	0.718	Decreases	0.234
Platelet (103/μL)	Increases	0.640	Decreases	0.137
SII	Increases	0.490	Decreases	0.019
Age (years)	Increases	0.424	Decreases	0.006
AST (U/L)	Increases	0.377	Decreases	0.216
NLR	Increases	0.368	Increases	0.931
HDL (mg/dL)	Decreases	-0.133	Decreases	0.142
ALT (U/L)	Decreases	-0.187	Decreases	0.000
NHR	Decreases	-0.263	Decreases	0.143
Neutrophil (103/μL)	Decreases	-0.525	Decreases	0.031

¹Global direction was defined by the sign of Spearman correlation between the feature value and its SHapley Additive exPlanations value (positive = higher values tend to increase the predicted malignancy probability; negative = higher values tend to decrease it).

²High-tail direction summarizes SHapley Additive exPlanations sign behavior among observations in the top 20% of the feature distribution, highlighting potential non-linear or threshold effects.

SHAP: SHapley Additive exPlanations; SII: Systemic immune-inflammation index; ALT: Alanine aminotransferase; NLR: Neutrophil to lymphocyte ratio; HDL: High-density lipoprotein; AST: Aspartate aminotransferase; NHR: Neutrophil-to-high-density lipoprotein ratio.

The very high NPV of our model (0.993-0.997) deserves particular attention. In a population where malignancy prevalence is low, as in unselected colonoscopy referrals, a high NPV means that patients classified as low risk by the model have a residual malignancy probability of only 0.3%-0.7%. This rule-out-oriented framework is conceptually aligned with the triage paradigms proposed for FIT-based pathways, where the primary objective is to safely exclude significant pathology rather than to confirm cancer[7,19]. The modest PPV (0.026-0.050) reflects the mathematical effect of low prevalence on the positive predictive value and does not undermine the model’s intended function; rather, it indicates that a positive screen should prompt expedited investigation rather than serve as a stand-alone diagnostic. This asymmetric performance profile - favoring sensitivity and NPV over specificity and PPV - is precisely what is required for a triage tool in which the clinical cost of a false-negative (missed malignancy) far exceeds the cost of a false-positive (an unnecessary but otherwise safe colonoscopy).

Notably, the low PPV at the Youden cutoff (0.050) implies that if the model was used as a rule-in tool, the large majority of positively flagged patients would ultimately prove to be nonmalignant, creating a substantial false-alarm burden in settings where positive flags trigger expedited work-up. This is an unavoidable mathematical consequence of the 1.43% malignancy prevalence and would apply to any test operating in such a population. Two strategies mitigate this concern within our framework: First, we explicitly position the model as a rule-out rather than a rule-in tool, with clinically recommended thresholds chosen to maximize safe rule-out yield (pt = 0.02-0.05) rather than statistical discrimination at the Youden point; second, for settings that require case-finding, the all-features model - with its higher LR+ of 3.66 (95%CI: 2.34-5.19) - is better suited for flagging patients for expedited evaluation, albeit at a lower sensitivity.

From a clinical implementation perspective, the model offers several pragmatic advantages. First, it relies exclusively on variables that are routinely collected as part of standard precolonoscopy evaluation, imposing no additional cost or logistic burden on the health care system. Second, the reduced SHAP10 model demonstrates that ten features suffice to maintain or even marginally improve performance relative to the full model, facilitating integration into electronic health record systems with minimal data requirements. Third, unlike FIT-dependent models, this approach can be applied regardless of whether fecal testing has been performed, expanding its potential utility in settings where FIT is unavailable or where patients have already been referred for colonoscopy on clinical grounds. Potential deployment scenarios include assisting gastroenterologists in prioritizing waiting lists, supporting primary care referral decisions, and serving as a complementary tool alongside FIT-based pathways.

Limitations

This study has several limitations that must be acknowledged. First, the retrospective, single-center design limits external generalizability, and the model requires validation in independent, multicenter cohorts with differing demographics and referral patterns. Second, the low malignancy prevalence (23/1604; 1.43%), while representative of a real-world consecutive colonoscopy cohort, constrains the statistical precision of performance estimates and inflates the width of confidence intervals around sensitivity and PPV. Third, certain clinically important predictors - including symptoms, fecal hemoglobin concentration, family history, and endoscopic quality indicators - were not available in the dataset and could improve model performance if incorporated. Fourth, the inflammatory indices used as model predictors (NLR, SII, PLR, and NHR) are nonspecific markers of systemic inflammation and can be elevated in conditions unrelated to colorectal malignancy, including inflammatory bowel disease (active or in remission), recent infections, rheumatologic conditions, and uncontrolled diabetes. Eight patients in our cohort had documented ulcerative colitis in remission; these were retained as nonmalignant cases. In real-world deployment - particularly in populations with a higher burden of chronic inflammatory comorbidity than our cohort - such patients may contribute disproportionately to false-positive flags. Future implementations should consider either excluding patients with active inflammatory comorbidities or incorporating baseline inflammatory status (e.g., C-reactive protein, erythrocyte sedimentation rate, known inflammatory bowel disease) as an additional feature or an exclusion criterion. Fifth, we did not perform temporal external validation; all evaluations relied on repeated stratified cross-validation, which, although rigorous, does not fully simulate prospective deployment. Sixth, the PR-AUC values (0.055-0.078) remain modest, reflecting the inherent difficulty of predicting rare events from nonspecific biomarkers. Finally, this model was designed as a decision-support aid and is explicitly not intended to replace colonoscopy or to be used as a sole criterion for denying investigation.

CONCLUSION

We have demonstrated that an XGBoost model using routinely available precolonoscopy laboratory parameters and derived inflammatory/metabolic indices can achieve a very high NPV for colorectal malignancy in a real-world, low-prevalence cohort. The rule-out-oriented design of the model is aligned with the growing need for noninvasive triage tools to optimize endoscopy resource allocation and prioritize high-risk patients. A parsimonious ten-feature model preserves discrimination while enhancing clinical applicability. Prospective, multicenter validation studies are warranted to evaluate the clinical impact and safety of integrating such a tool into colonoscopy referral pathways.