Predicting three-year risk of incident colorectal precancerous conditions using non-invasive predictors and machine-learning algorithms: A multi-center retrospective study

Abstract

BACKGROUND

Current screening guidelines recommend a three-year follow-up for patients with non-advanced adenoma (non-AA), making it challenging to capture rapidly evolving precancerous conditions and provide timely preventive interventions.

AIM

To develop and validate a model using non-invasive predictors to predict the three-year risk of developing advanced adenoma (AA) among non-AA patients.

METHODS

This study retrospectively included 1808 Chinese patients from two hospitals aged 30-80 years who had newly diagnosed non-AA lesions. The patients were divided into a training cohort (n = 761), an internal validation cohort (n = 762), and an external validation cohort (n = 285). The least absolute shrinkage and selection operator (LASSO) procedure was used to select predictors from 65 non-invasive parameters. Five machine learning models and a stepwise logistic regression were applied to predict incident AA diagnosed within three years after the baseline screening.

RESULTS

The LASSO procedure selected seven predictors: (1) Serum thymidine kinase 1 protein; (2) Cytokeratin 19 fragment; (3) Carbohydrate antigen 19-9; (4) Age; (5) Diastolic blood pressure; (6) Fecal immunochemical test; and (7) Sex. The EXtreme Gradient Boosting (XGBOOST) machine-learning model achieved an area under the receiver operating characteristic curve of 0.988 (95%CI: 0.980-0.996) in the training cohort, 0.825 (0.766-0.883) in the internal validation cohort, and 0.767 (0.642-0.892) in the external validation cohort, higher than other models (P < 0.05). SHapley Additive exPlanations analysis was applied to quantify the contribution of each included predictor to the XGBOOST model’s output. An open-access web application was deployed for the XGBOOST model.

CONCLUSION

The XGBOOST model included seven non-invasive predictors for precise prediction of three-year AA risk and was validated internally and externally, providing a feasible tool to guide more intensive follow-up for high-risk non-AA patients.

Key Words: Early detection of cancer; Precancerous conditions; Colorectal neoplasms; Machine learning; Retrospective study

Core Tip: This retrospective study developed and validated a machine learning model using seven non-invasive predictors, serum thymidine kinase 1 protein, cytokeratin 19 fragment, carbohydrate antigen 19-9, age, diastolic blood pressure, fecal immunochemical test, and sex, to predict the three-year risk of incident advanced adenoma (AA) in patients with non-AA. The model achieved high discriminative performance [area under the receiver operating characteristic: (1) 0.988 in training; and (2) 0.767 in external validation] and superior clinical utility over traditional methods. An open-access web application was deployed to enable precise detection of patients at high-risk of developing AA, offering a novel tool to guide more intensive follow-up for this high-risk subgroup.

INTRODUCTION

Colorectal cancer (CRC) is the third most common incident malignancy and the second leading cause of cancer death worldwide, imposing considerable health and economic burdens[1]. Advanced adenomas (AAs) are typical precancerous conditions of CRC[2]. Therefore, AA is an important prevention target to reduce both CRC incidence and mortality. Current National Comprehensive Cancer Network guidelines recommend colonoscopy screening after 5-10 years for individuals who have newly diagnosed non-AA[3]. However, this strategy ignores the substantial heterogeneity in short-term risk. Population-based studies show that 1.5%-10.3% of non-AA individuals develop AA within only 3-5 years, whereas many others remain non-advanced for decades[4-6]. Shortening the rescreening time interval for everyone would not be feasible due to the low update rate (14%-19%) of invasive colonoscopy[7,8], whereas maintaining the long interval risks missing rapidly evolving lesions. A non-invasive strategy that identifies the high-risk subgroup of developing AA within a short-term could optimize resource allocation, yet no validated tool exists for non-AA populations.

Several risk prediction models have focused on five-year CRC incidence with modest discriminative power [area under the receiver operating characteristic (AUC): 0.63-0.70][9]. Few models target a short-term window specifically for incident AA. Recent models for AA have successfully used non-invasive predictors and machine learning algorithms (AUC: 0.85-0.98), but they focused on prevalent case diagnosis instead of the prediction of incident cases[10,11], making it hard to guide follow-up strategies. Moreover, prior models require a large number of predictors (16-30) that are less feasible in routine care. There is growing interest in exploring a refined model to assess short-term incident AA risk without additional patient burden.

The routinely used non-invasive screening method, the fecal blood test, showed a modest sensitivity for AA diagnosis (6%-56%)[12], and no evidence for incident AA prediction. Previous models for AA diagnosis commonly identified age, sex, fecal blood test, and blood pressure as main predictors[10,11,13]. Recent studies have revealed that alterations in serum cancer-related biomarkers such as cytokeratin 19 fragment (CYFRA21-1)[14], carbohydrate antigen 19-9 (CA19-9)[15], and cell proliferation marker serum thymidine kinase 1 protein (STK1p)[16], may precede CRC. Whether these non-invasive measurements can forecast short-term AA risk in a real-world screening cohort remains unknown. Machine learning algorithms are well-suited to detect interactions among these potential AA predictors, but their value for this specific task has not been explored.

Using data from two geographically different Chinese cohorts, we therefore conducted a retrospective analysis to develop and validate a machine learning prediction model that estimates the 3-year probability of incident AA among screening populations with newly diagnosed non-AA lesions. By integrating non-invasive predictors, we sought to create an open-access tool that can be immediately applied in practice to personalize follow-up strategies.

MATERIALS AND METHODS

Study design and participants

This was a retrospective analysis of clinical data from two geographically different centers. Center 1 was Shenzhen People’s Hospital, located in South China, and Center 2 was West China Longquan Hospital Sichuan University, located in West China. Both centers provided routine colonoscopy screenings for those without a history of cancer. We included participants aged 30-80 years who had newly diagnosed non-AA between January 2020 and December 2022. Participants were excluded if they had any baseline AA or CRC (Figure 1A). Individuals who did not complete a three-year follow-up colonoscopy were also excluded. Eligible participants from Center 1 were chronologically ordered by their enrollment date and then divided into a training cohort (n = 762) and an internal validation cohort (n = 761) in a 1:1 ratio. Eligible participants from Center 2 were assigned to an external validation cohort (n = 285).

Figure 1 Study flow charts. A: Flow chart of study participants. Center 1 was Shenzhen People’s Hospital and Center 2 was West China Longquan Hospital Sichuan University; B: Procedure for developing prediction models. CA19-9: Carbohydrate antigen 19-9; CYFRA21-1: Cytokeratin 19 fragment; DBP: Diastolic blood pressure; FIT: Fecal immunochemical test; LASSO: Least absolute shrinkage and selection operator; STK1p: Serum thymidine kinase 1 protein.

Data collection

Baseline data were extracted from electronic health records and laboratory databases. We collected 114 non-invasive parameters, including demographics, risk factors, anthropometric measures, routine laboratory test results (complete blood count, liver function, kidney function, lipids, etc.), fecal immunochemical test (FIT) results and tumor-associated biomarkers. Tumor-associated biomarkers included CA19-9, CYFRA21-1, STK1p, carcinoembryonic antigen, and other routinely measured markers. All biomarkers were measured using standardized protocols in the participating centers.

Outcome definition

The primary outcome was incident AA diagnosed at the follow-up colonoscopy performed within three years ± six months after the baseline screening. AA was defined as adenomas ≥ 10 mm in diameter, or with villous components, or exhibiting high-grade dysplasia, according to the National Comprehensive Cancer Network guidelines of CRC screening. Diameter data were extracted from the colonoscopy reports, and histopathological data were extracted from pathology reports. All colonoscopies were performed by certified endoscopists, and histopathological reports were provided by certified pathologists.

Feature selection

We excluded variables (n = 49) with > 40% missing data, which could introduce bias; this exclusion threshold was chosen based on established recommendations in statistical guidance to balance data retention with analytical reliability. For the remaining 65 variables, we employed K-nearest neighbor imputation because it preserves the underlying data structure by estimating missing values from similar observations, making it particularly suitable for our mixed-type clinical dataset. To identify the most informative feature combination from the 65 candidate variables, we applied least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation. LASSO was used to penalize less contributive parameters and select a parsimonious set of features for model development. Features with non-zero coefficients at the optimal lambda were retained.

Model development

Five machine learning models were developed using the selected predictors: (1) EXtreme Gradient Boosting (XGBOOST); (2) LASSO; (3) Decision tree; (4) Support vector machine; and (5) Neural network (Figure 1B). Additionally, a stepwise logistic regression model was developed for comparison. All models were trained in the training cohort using 10-fold cross-validation to prevent overfitting. Model performance was evaluated using the AUC curve, sensitivity, specificity, and decision curve analysis (DCA).

Model validation

The internal validation cohort was used to assess model generalizability within the same population. The external validation cohort was used to test the robustness of the models across different populations. AUC, sensitivity, specificity, and DCA were also provided for validation cohorts. The best model was defined as the one with the highest AUC in the external validation cohort. SHapley Additive exPlanations (SHAP) analysis was applied to quantify the contribution of each included predictor to the best model’s output. An open-access web application was deployed for the best model.

Statistical analysis

Continuous variables were presented as medians with interquartile range, and categorical variables as n (%). Comparisons were made using the Kruskal-Wallis test (comparing three or more groups) or the Mann-Whitney U-test (comparing two groups) for quantitative variables. Comparisons were made using the χ² test for categorical variables. P values < 0.05 were considered statistically significant. All statistical analyses were performed using R software (version 4.5.1) and Stata/MP (version 18). The statistical methods of this study were reviewed by Gui-Yuan Han, a medical statistician from the National Clinical Research Center for Cardiovascular Diseases, Fuwai Hospital Chinese Academy of Medical Sciences, Shenzhen, China.

RESULTS

Participant characteristics

A total of 1808 eligible participants were enrolled in this study, including 762 in the training cohort, 761 in the internal validation cohort, and 285 in the external validation cohort (Table 1). The median age was 48 years (interquartile range: 41-54), and 24.6% were female. Over the three-year follow-up period, 131 participants (7.3%) were diagnosed with incident AA. The incidence rates were similar across cohorts: (1) 7.0% in the training; (2) 6.8% in the internal validation; and (3) 9.1% in the external validation cohort (P = 0.410).

Table 1 Baseline characteristics and incident events of study participants, median (interquartile range)/n (%).

Characteristics	Training	Internal validation	External validation	P value	Overall
n	762	761	285		1808
Incident AA	53 (7.0)	52 (6.8)	26 (9.1)	0.410	131 (7.3)
Site of incident AA				0.012
Rectum	6 (0.8)	19 (2.5)	10 (3.5)		35 (1.9)
Left colon	28 (3.7)	14 (1.8)	6 (2.1)		48 (2.7)
Right colon	19 (2.5)	19 (2.5)	10 (3.5)		48 (2.7)
Age (years)	48 (41, 54)	47 (40, 53)	51 (43, 59)	< 0.001	48 (41, 54)
Female sex	175 (23.0)	164 (21.6)	106 (37.2)	< 0.001	445 (24.6)
DBP (mmHg)	74 (66, 82)	73 (66, 82)	76 (69, 84)	0.020	74 (66, 82)
STK1p (pmol/L)	0.16 (0.10, 0.24)	0.16 (0.10, 0.23)	0.33 (0.17, 0.72)	< 0.001	0.17 (0.11, 0.26)
FIT positive	213 (27.95)	216 (28.38)	77 (27.0)	0.908	506 (28.0)
CYFRA21-1 (μg/L)	2.00 (1.42, 2.70)	1.92 (1.44, 2.80)	1.26 (0.95, 1.42)	< 0.001	1.83 (1.26, 2.60)
CA19-9 (kU/L)	6.90 (3.62, 11.08)	6.72 (3.75, 11.27)	6.69 (3.50, 11.53)	0.959	6.77 (3.62, 11.24)

Advanced adenoma (including adenoma ≥ 10 mm, with villous components, or exhibiting high-grade dysplasia) diagnosis based on the National Comprehensive Cancer Network guideline. AA: Advanced adenoma; CA19-9: Carbohydrate antigen 19-9; CYFRA21-1: Cytokeratin 19 fragment; DBP: Diastolic blood pressure; FIT: Fecal immunochemical test; STK1p: Serum thymidine kinase 1 protein.

Baseline predictors are summarized in Table 1. Participants in the external validation cohort were significantly older, had a higher rate of female participants, showed higher values of diastolic blood pressure (DBP), STK1p, and CYFRA21-1, compared to the training and internal validation cohorts (P < 0.05). FIT-positive rates and CA19-9 values were similar among the cohorts (P ≥ 0.05).

Feature selection

LASSO regression with 10-fold cross-validation identified seven predictors among the 65 candidate features for machine learning model development (Table 2, Figure 2A). Correlations between each predictor were mild or insignificant (Figure 2B). Median values of CYFRA21-1, CA19-9, age, STK1p, DBP, and percentages of FIT-positive and female sex all showed an increasing trend in participants with incident AA compared to those without across the cohorts (Figure 2C). The univariate logistic model in the training cohort showed that CYFRA21-1, CA19-9, age, STK1p, and DBP had significant positive associations with incident AA, while FIT-positive and female sex had positive but insignificant associations (Table 3). The stepwise logistic model included CYFRA21-1, STK1p, and DBP as predictors (Table 3).

Figure 2 Model feature selection. A: Least absolute shrinkage and selection operator procedure for selecting predictors in the training cohort; B: Spearman correlations between included predictors in the training cohort; C: Distribution of selected predictors in participants with and without incident advanced adenoma in the training and validation cohorts (data are medians with interquartile ranges or n (%)). ^aP < 0.05. AA: Advanced adenoma; CA19-9: Carbohydrate antigen 19-9; CYFRA21-1: Cytokeratin 19 fragment; DBP: Diastolic blood pressure; FIT: Fecal immunochemical test; STK1p: Serum thymidine kinase 1 protein.

Table 2 Predictors selected by the least absolute shrinkage and selection operator procedure.

Lambda	Number of non-zero coefficient	Cross-validation mean deviance	Variables (A)dded, (R)emoved, or left (U)nchanged lambda
0.0513	1	0.5032	A	STK1p
0.0388	2	0.4959	A	CA19-9
0.0294	3	0.4938	A	CYFRA21-1
0.0268	4	0.4936	A	DBP
0.0153	5	0.4805	A	Age
0.0088	6	0.4764	A	FIT positive
0.00731	7	0.4762	A	Sex

¹Lambda selected by the minimum cross-validation mean deviance.

CA19-9: Carbohydrate antigen 19-9; CYFRA21-1: Cytokeratin 19 fragment; DBP: Diastolic blood pressure; FIT: Fecal immunochemical test; STK1p: Serum thymidine kinase 1 protein.

Table 3 Development of a logistic model in the training cohort.

Predictor	Odds ratio	95%CI	Coefficient	P value
Univariate regression
STK1p (pmol/L)	1.69	1.10-2.61	0.5262	0.017
CYFRA21-1 (μg/L)	1.31	1.11-1.56	0.2717	0.002
DBP (mmHg)	1.04	1.02-1.07	0.0407	0.001
Age (years)	1.04	1.01-1.08	0.0420	0.007
CA19-9 (kU/L)	1.03	1.003-1.05	0.0265	0.026
Female sex	1.10	0.57-2.10	0.0929	0.779
FIT positive	1.24	0.68-2.25	0.2119	0.489
Stepwise multivariate regression
STK1p (pmol/L)	1.84	1.15-2.94	0.6084	0.011
CYFRA21-1 (μg/L)	1.32	1.10-1.59	0.2794	0.003
DBP (mmHg)	1.04	1.01-1.07	0.0375	0.004
Constant	0.002	0.0002-0.01	-6.3842	< 0.001

CA19-9: Carbohydrate antigen 19-9; CYFRA21-1: Cytokeratin 19 fragment; DBP: Diastolic blood pressure; FIT: Fecal immunochemical test; STK1p: Serum thymidine kinase 1 protein.

Model development and validation

Among the six models developed, XGBOOST demonstrated the highest discriminative performance across the cohorts (P < 0.05; Table 4, Figure 3A). In the training cohort, it achieved an AUC of 0.988 (95%CI: 0.980-0.996), with a sensitivity of 96.23% and specificity of 99.86% at the risk threshold of 50%. In the internal validation cohort, the AUC was 0.825 (95%CI: 0.766-0.883), with sensitivity and specificity of 63.46% and 99.86%, respectively. In the external validation cohort, the AUC was 0.767 (95%CI: 0.642-0.892), with a sensitivity of 61.54% and specificity of 90.35%.

Figure 3 Model performance. A: Receiver operating characteristic curves of different models in the training, internal validation, and external validation cohorts; B: Decision analyses of different models in the training, internal validation, and external validation cohorts. AUC: Area under the receiver operating characteristic; LASSO: Least absolute shrinkage and selection operator; NNET: Neural network; SVM: Support vector machine; XGBOOST: EXtreme Gradient Boosting.

Table 4 Performance of different models for predicting advanced adenoma.

Model	Area under the curve (95%CI)	Threshold	Sensitivity	Specificity
Training
XGBOOST (probability)	0.988 (0.980-0.996)	≥ 50%	96.23%	99.86%
LASSO (probability)	0.682 (0.596-0.768)	≥ 7.43%	50.94%	84.77%
Decision tree (1 vs 0)	0.667 (0.602-0.733)	≥ 1	35.85%	97.60%
NNET (1 vs 0)	0.575 (0.527-0.624)	≥ 1	9.43%	99.58%
SVM (1 vs 0)	0.519 (0.493-0.545)	≥ 1	3.77%	100.00%
Logistic model (probability)	0.691 (0.607-0.774)	≥ 8.87%	50.94%	85.47%
Internal validation
XGBOOST (probability)	0.825 (0.766-0.883)	≥ 50%	63.46%	99.86%
LASSO (probability)	0.616 (0.538-0.693)	≥ 7.43%	26.92%	83.36%
Decision tree (1 vs 0)	0.546 (0.502-0.590)	≥ 1	11.54%	97.60%
NNET (1 vs 0)	0.492 (0.488-0.497)	≥ 1	0.00%	98.45%
SVM (1 vs 0)	0.499 (0.498-0.501)	≥ 1	0.00%	99.86%
Logistic model (probability)	0.622 (0.545-0.698)	≥ 8.87%	26.92%	84.06%
External validation
XGBOOST (probability)	0.767 (0.642-0.892)	≥ 50%	61.54%	90.35%
LASSO (probability)	0.500 (0.500-0.500)	≥ 7.43%	-	-
Decision tree (1 vs 0)	0.561 (0.477-0.646)	≥ 1	23.08%	89.19%
NNET (1 vs 0)	0.552 (0.489-0.615)	≥ 1	11.54%	98.84%
SVM (1 vs 0)	0.500 (0.500-0.500)	≥ 1	-	-
Logistic model (probability)	0.706 (0.602-0.809)	≥ 8.87%	69.23%	67.95%

LASSO: Least absolute shrinkage and selection operator; NNET: Neural network; SVM: Support vector machine; XGBOOST: EXtreme Gradient Boosting.

DCA demonstrated that the XGBOOST model provided a higher net benefit across a wide range of threshold probabilities compared to other models in all three cohorts (Figure 3B). This indicates superior clinical utility in identifying individuals at high risk for incident AA.

Model interpretation and application

The visualized XGBOOST model with 149 decision trees is presented in Supplementary Figure 1. Figure 4A presents SHAP dependence plots for the seven selected predictors, illustrating how the value of each predictor influences the model output. In each subplot, the X-axis denotes the actual feature value, the Y-axis represents the corresponding SHAP value (indicating the direction and magnitude of the feature’s contribution to the predicted risk), and the color gradient encodes the value of a second interacting feature (typically age or STK1p) to reveal potential interaction effects. For example, higher CYFRA21-1 and CA19-9 values are generally associated with positive SHAP values (increased predicted risk), whereas the relationship between age and SHAP values demonstrates a clear positive trend, and higher STK1p levels consistently drive predictions toward higher risk.

Figure 4 Model interpretation and application. A: Scatter plots showing the relationship of actual values of included predictors with SHapley Additive exPlanations (SHAP) values and a second predictor with the strongest correlation; B: Feature importance ranked by mean absolute SHAP values; C: SHAP beeswarm plots illustrating impact distributions of included predictors; D: Online EXtreme Gradient Boosting prediction tool accessible at https://precrc.shinyapps.io/precrc/. A predicted probability ≥ 50% and < 50% is coded as 1 and 0, respectively; E: Percentage of incident advanced adenoma in EXtreme Gradient Boosting predicted high-risk and low-risk groups (0 = low risk, 1 = high risk). CA19-9: Carbohydrate antigen 19-9; CYFRA21-1: Cytokeratin 19 fragment; DBP: Diastolic blood pressure; FIT: Fecal immunochemical test; SHAP: SHapley Additive exPlanations; STK1p: Serum thymidine kinase 1 protein; XGBOOST: EXtreme Gradient Boosting.

Figure 4B displays the SHAP feature importance ranking based on the mean absolute SHAP value for each predictor. Features are ordered by their overall contribution to the model’s predictions across the entire dataset. STK1p exhibits the highest mean absolute SHAP value, followed by CYFRA21-1, CA19-9, DBP, age, sex, and FIT, indicating that STK1p and the tumor-associated biomarkers collectively dominate model decision-making.

Figure 4C shows a SHAP beeswarm plot summarizing the distribution of SHAP values for each feature among all patients. Each point represents an individual patient, with the horizontal position indicating the SHAP value and the color coding reflecting whether the patient’s feature value was low or high. This visualization shows not only which features are most influential globally but also how specific patient characteristics drive individual predictions toward higher or lower risk estimates.

Figure 4D provides a screenshot of the open-access web-based risk calculator. The interface allows clinicians to input the seven non-invasive predictors (age, sex, CYFRA21-1, CA19-9, STK1p, DBP, and FIT) and obtain an individualized three-year risk estimate for incident AA, facilitating real-time clinical decision-making at the point of care.

Figure 4E illustrates the distribution of XGBoost-predicted risk categories (0 = low risk, 1 = high risk) stratified by actual outcome (non-AA vs AA) across the training, internal validation, and external validation cohorts. The stacked bar charts show the percentage of patients in each category, demonstrating that the model assigns a substantially higher proportion of AA patients to the high-risk category compared with non-AA patients among the three cohorts (P < 0.001).

DISCUSSION

We developed and externally validated a prediction tool based on an XGBOOST model. It uses only seven non-invasive parameters to identify non-AA patients at high risk of incident AA within three years after the first-time colonoscopy. Both discrimination and DCA indicated a consistent net benefit over traditional logistic regression and four alternative machine-learning algorithms. An open-access application was deployed for incident AA risk prediction to guide follow-up practice in non-AA populations.

Although AA is a well-recognized precancerous lesion of CRC, previous studies exploring its etiology are scarce. The associations of the seven selected predictors, CYFRA21-1, CA19-9, age, STK1p, DBP, FIT, and sex, align well with published literature, indirectly supporting their biological plausibility for AA. Cytokeratin-19 is one of the structural proteins found in the intracytoplasmic cytoskeleton of intestinal epithelium; when cells lose polarity and detach from the basement membrane (a hallmark of high-grade dysplasia), degradation of the cytokeratin-19 network is greatly increased[17]. The resulting fragments (measured as CYFRA 21-1) escape into the lamina propria and then the portal circulation. A systematic review and a prospective study suggested that CYFRA21-1 presents promise in clinical colonoscopy settings, which showed an AUC of 0.73 for detecting AA in a general cancer screening setting[18,19]. CA19-9, although conventionally utilized for pancreatic cancer, has been reported to be elevated in AA patients (adjusted odds ratio: 1.33, P = 0.045)[15]. An increase in tissue concentrations of CA19-9 has also been shown to progress from normal colonic mucosa through colorectal adenomas to carcinomas[20]. Age is a well-known risk factor for CRC; the risk of AA also rises markedly with advancing age in the current study[21]. The sex difference in CRC incidence is controversial in previous studies, and the reasons for this are not fully understood. Our identification of female sex as a predictor for AA is therefore justified and concurs with some epidemiological observations[21]. STK1p is a cell-proliferation enzyme frequently overexpressed in various malignancies, including CRC[22]. Our finding is consistent with a meta-analysis indicating that STK1p can serve as a biomarker for the progression from healthy to adenomas (pooled mean difference: -0.39, P < 0.001)[23]. DBP showed a statistically significant association with AA in the present study. Seretis et al[23] observed a positive relationship between blood pressure and the risk of several malignancies, including CRC, in their meta-analysis, lending indirect support to our finding. FIT is a well-established, non-invasive screening method for CRC (sensitivity: 79%, specificity: 94%)[24]. However, the sensitivity of FIT for detecting AA was low to moderate (6%-56%)[12]. Our study also found a positive but insignificant association between FIT-positive results and incident AA, but the LASSO procedure still selected FIT as a model predictor. This might be because machine learning algorithms could detect mild interactions among multiple predictors.

As published models have focused on prevalent AA diagnosis, there is no direct comparison between our model and other incident AA prediction studies. The XGBOOST model achieved an AUC of 0.988 in the training set, outperforming the logistic-regression and other machine-learning models. External validation yielded a more conservative but still clinically attractive value (AUC: 0.767), surpassing most published incident CRC prediction models (AUC: 0.63-0.70)[9]. High specificity is particularly important in screening settings, where false-positive colonoscopy referrals incur substantial cost and patient burden. Our model retained a high specificity of 90.35% and a moderate sensitivity of 61.54% in the external validation, while previous models for incident CRC prediction showed a low sensitivity (11.0%-26.4%) when the specificity was over 90%[9]. DCA further demonstrated that the highest net clinical benefit was for the XGBOOST model.

Unlike the existing risk prediction models that predominantly focus on CRC incidence[9], our model specifically targets incident AA within a three-year window. This focus is clinically important because incident AA represents the direct precursor lesion that determines subsequent surveillance intervals, yet it has been underrepresented in previous prediction tools. Our model integrates a concise panel of non-invasive parameters, ensuring patient acceptance and screening cost control. The inclusion of STK1p is particularly innovative, as it reflects cellular proliferative activity and may capture the biological progression from normal mucosa to AA sensitively. Furthermore, while most existing models were developed in populations with limited geographical diversity, our multi-center design across southern and western China provides valuable evidence for an understudied population. The clinical advantage of our model lies in its potential to enable truly individualized surveillance. By stratifying patients into risk categories at the point of adenoma diagnosis, clinicians can shift beyond uniform interval recommendations toward precision-based follow-up schedules. This may reduce unnecessary procedures in low-risk individuals, ensuring intensive monitoring for those at higher risk of rapid adenoma progression, thereby optimizing resource allocation and patient compliance in population-level screening programs. Moreover, the interpretable decision trees generated by XGBoost enhance physicians' understanding of the prediction process, thereby increasing clinical trust. In the future, the algorithm could be embedded in mobile applications or electronic health-record systems to provide real-time risk assessment and tailored screening recommendations.

There are several limitations in this study. First, regarding selection bias inherent to the retrospective design, our reliance on existing medical records may have introduced ascertainment bias, as patients who underwent regular follow-up colonoscopy may have differed from those who did not in terms of health-seeking behavior, socioeconomic status, or underlying risk perception. This could have resulted in a study population that is not fully representative of the broader screening-eligible population, potentially affecting the external validity of our risk estimates. Second, the exclusion of patients without complete follow-up data represents a limitation. If patients lost to follow-up were predominantly low-risk individuals, our model may overestimate the true incidence of AA and overstate predictive performance in real-world settings where adherence to surveillance is imperfect. Third, although our multi-center design included hospitals in southern and western China, all participants were of Chinese ethnicity, and the two centers may not capture the full spectrum of genetic, environmental, and lifestyle variations present across other regions of China or other countries. This homogeneity limits the direct generalizability of our model to other ethnic groups or populations with different CRC epidemiology, dietary patterns, and screening infrastructures. Forth, novel biomarkers, such as serum or fecal DNA methylation markers[25], were not available. Missed key predictors could have reduced the model’s performance. The high missingness threshold (> 40%) was selected to preserve analytical power, yet it may have excluded potentially informative predictors. Furthermore, KNN imputation assumes that missing values are missing at random; if this assumption was violated, the imputed values could have introduced subtle biases that inflated model performance in the training cohort. This may partially explain the observed performance drop in external validation, where the missing data patterns and patient characteristics differed from the training cohort. Future studies should consider complete data strategies or sensitivity analyses to further quantify the impact of missing data on model robustness and generalizability. Fifth, the notably high AUC (0.988) in the training cohort raises concerns about potential overfitting, although LASSO regularization and 10-fold cross-validation were used during model development. The substantial performance drop observed in external validation (AUC: 0.767) likely reflects a combination of overfitting to the training data distribution and genuine demographic and clinical discrepancy between cohorts. The observed external validation performance was lower than training performance, which remains clinically acceptable and superior to existing risk prediction models (AUC: 0.63-0.70) in the CRC screening literature[9]. Finally, the small number of incident AA cases restricted the subgroup analyses by lesion site and led to wide confidence intervals for some comparison analyses.

The AUC of 0.500 observed for some models such as the support vector machine and LASSO models in the external cohort indicates complete loss of discriminative ability, effectively reducing predictions to random chance. For the LASSO model, we hypothesize that the feature selection process, effective for the training distribution, may have produced a sparse set of predictors whose coefficients were highly sensitive to covariate shifts between cohorts. Specifically, the demographic differences in age, sex, and STK1p levels may have altered the linear relationships that LASSO assumed, leading to systematic miscalibration. In the support vector machine model, its reliance on kernel-based transformation of the feature space likely resulted in a decision boundary that was over-optimized to the training data topology and became non-generalizable when the external cohort’s feature distribution diverged, particularly in the context of the higher proportion of females and older patients with different baseline biomarker profiles. This model-specific fragility underscores the importance of comparing multiple algorithms during development, as the relative robustness of XGBOOST, likely attributable to its ensemble regularization, would not have been apparent without this comparative framework. Simpler linear models may be particularly vulnerable to external population shifts in multi-center studies, suggesting that future model development in similar settings should prioritize algorithms with demonstrated generalizability.

The open-access web application of the XGBOOST model was designed with a streamlined user interface that requires only the input of seven clinical and laboratory variables, enabling rapid risk stratification without specialized computational treatment. However, we acknowledge that current usability has not been formally evaluated through standardized instruments. Regarding real-world validation, the present study represents a retrospective proof-of-concept, and prospective implementation studies are essential to assess the tool’s performance in routine clinical environments where data quality, patient heterogeneity, and workflow dynamics may differ substantially among different research settings. We have planned a multi-center prospective study that will embed the model within routine care to evaluate its impact on clinician decision-making, patient adherence to recommended surveillance intervals, and ultimately, the detection rate of incident AA. For clinical workflow integration, we envision the model being deployed at two critical decision points: (1) Immediately following the first-time colonoscopy to generate personalized surveillance recommendations; and (2) During follow-up visits to prioritize high-risk patients for earlier colonoscopy scheduling.

Future studies are required to improve the model performance and application. The external validation cohort exhibited a higher median age, higher proportion of females, elevated levels of DBP and STK1p, and a lower level of CYFRA21-1. These differences are clinically meaningful because age and sex are established risk factors for CRC, and STK1p is a novel biomarker associated with CRC and AA. Such distributional shifts may have contributed to the observed performance attenuation in external validation and could lead to under-estimation or over-estimation of model performance if applied to populations with divergent baseline characteristics. There is a need for future calibration studies to enhance model transportability across diverse clinical settings. Future experiments are required to further explore the biological plausibility for associations between included biomarkers and the onset of AA. Further external validations of the model in larger populations and across Asian and Western databases are also required to minimize geographic bias.

CONCLUSION

We provide an externally validated XGBOOST model, including seven non-invasive parameters, to predict three-year AA risk among patients with newly diagnosed non-AA lesions. This may provide a novel and effective tool to guide more intensive follow-up strategies for high-risk non-AA patients. Prospective studies are planned to verify the model performance and explore the effectiveness of a stratified follow-up strategy combining this model.

ACKNOWLEDGEMENTS

We would like to thank all physicians and nurses at Shenzhen People’s Hospital and West China Longquan Hospital Sichuan University for providing high-quality data of routine colonoscopy screening.