An exploratory multimodal pipeline for recurrence prediction in CRC with XELOX therapy Authors: Zhihan Li, Hongqing Ma, Wenbo Niu, Tao Zhang, Zihan Fan, Simeng Jiang Published: 2026 ABSTRACT AbstractColorectal cancer(CRC) recurrence remains a challenge despite curative surgery and adjuvant therapy. Integrating clinicopathological and liquid biopsy data may improve risk stratification, but evidence in small‑sample settings is lacking. This exploratory study assessed the feasibility of a multimodal machine learning pipeline for recurrence prediction in a well‑defined XELOX‑treated cohort. We analyzed 86 CRC patients(17 recurrences) receiving uniform XELOX adjuvant therapy. A three‑step pipeline was applied: (1) LASSO regression for coarse variable selection (10‑fold cross‑validation); (2) gradient boosting machine and random forest was then employed for further variable refinement; (3) The common variables selected by both methods was then used for a Firth penalized logistic regression model. The final model performance was evaluated using bootstrap‑corrected AUC, calibration plot, and decision curve analysis(DCA). The intersection of GBM and RF selected four predictors: ctDNA positivity, perineural invasion, high‑grade tumor budding, and T stage (T5 vs. T1–4). The final Firth model yielded an optimism‑corrected AUC of 0.695(95% bootstrap CI: 0.58–0.81). Wide confidence intervals (e.g., ctDNA OR = 3.95; 95% CI: 0.83–8.61) and an unexpected coefficient direction for T stage indicated statistical instability due to the limited event count. Although Calibration showed systematic deviations, DCA suggested potential net benefit for thresholds 0.10–0.30. This exploratory study demonstrates the feasibility of a multimodal machine learning pipeline for recurrence prediction in a small CRC cohort. All findings are hypothesis‑generating, and the model is not clinically ready. External validation in larger prospective cohorts is required before any clinical application can be considered.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56986-5.Keywords: Colorectal cancer, MRD, Prediction model, XELOX, MultimodalSubject terms: Cancer, Computational biology and bioinformatics, Oncology, Risk factors Introduction Colorectal cancer (CRC) remains among the most prevalent malignancies worldwide. More than 1.9 million new cases of CRC are reported annually1. While localized tumors are often curable through surgery, approximately 30–50% of patients experience recurrence within five years, predominantly due to undetected MRD2. Despite great advancements in surgical techniques, chemotherapy, radiotherapy and targeted therapies, disease recurrence and metastasis are also challenges for long-term survival. Minimal residual disease(MRD) serves as a predictor of clinical recurrence, and its early detection offers a window for intervention to mitigate relapse3. A critical challenge in posttreatment management lies in identifying patients at high risk of minimal residual disease(MRD), which is defined as the presence of occult tumor cells or circulating tumor DNA (ctDNA) after curative-intent treatment3,4. In particular, some patients with advanced-stage disease receiving adjuvant chemotherapy do not always benefit from ctDNA risk stratification5,6. Tumor heterogeneity, various treatment strategies and the amount of ctDNA released can affect the accuracy of detection methods7,8. The adjunctive therapy method is another factor that affects cancer recurrence by potentially reshaping the tumor cell microenvironment9. However, current clinical practices lack robust tools to stratify patients on the basis of MRD risk, leading to either overtreatment of low-risk individuals or delayed interventions for high-risk cohorts. Traditional staging systems, such as the TNM classification, provide prognostic insights but fail to account for dynamic factors such as treatment response10. Current imaging tools and tumor markers, such as CEA and CA19-9, which are helpful for the initial diagnosis and monitoring of large-scale tumors, have sensitivity and specificity limitations. Owing to a lack of molecular heterogeneity and immune microenvironment interactions, these markers are also inadequate for detecting micrometastases and molecular residuals11–13. In contrast, circulating tumor cell (CTC) detection and circulating tumor DNA(ctDNA) analysis offer more sensitive and specific approaches14. Compared with other methods, CTCs provide direct and complete evidence of tumor cells in circulation and insights into molecular characteristics, whereas ctDNA allows for early detection of residual disease through fragment genetic profiling. Specifically, CTCs, which are intact tumor cells, offer insights into the biological features of tumors, such as epithelial‒mesenchymal transition(EMT) or cancer stemness, which may be associated with tumor aggressiveness, metastatic potential, and treatment resistance. On the other hand, ctDNA, which consists of DNA fragments released by tumor cells into the bloodstream, provides real-time molecular information about tumors, including mutations and other genomic changes, and is highly sensitive and specific for detecting residual disease and monitoring treatment responses15,16. CTC detection can provide valuable information on the biology of tumor cells and their metastatic potential. While CTC detection faces challenges such as a lack of standardization, high technical requirements, low cell count, inability to determine the tumor origin, and high cost17, ctDNA detection excels in the early prediction of tumor occurrence and offers a molecular basis for clinical decision-making. This method is also limited by low concentrations in blood, high technical requirements, the potential for false positives and negatives, the inability to provide tumor cell biological characteristics, and the need for standardization and cost challenges15. To avoid the drawbacks of single markers, the integration of multiple biomarkers is necessary for constructing a clinical model with high sensitivity and specificity. Compared with traditional single-data-type models, several recent studies have successfully integrated multimodal data for clinical model construction and development. In the field of oncology, the integration of clinical indicators such as age and tumor stage and molecular biomarkers such as MGMT methylation in glioblastoma and Ki-67 in breast cancer, along with radiomic features derived from MRI, has demonstrated superior prognostic performance17–19. Neurodegenerative disease models synergize CSF biomarkers(A/β-42/tau), genetic markers (APOE), MRI morphometrics(hippocampal atrophy), and cognitive scores to increase the accuracy of early Alzheimer’s disease diagnosis to 89%20. These multimodal approaches consistently outperform single-data-type models through advanced machine learning architectures that process heterogeneous spatiotemporal data patterns. Integrating these variables into a comprehensive model could enhance risk stratification, enabling personalized surveillance and adjuvant therapy strategies. Given the innovative nature of multimodal integration in CRC recurrence prediction, we designed this study as a prospective exploratory analysis to evaluate primarily the feasibility of combining clinicopathological and liquid biopsy data for risk stratification. Our study aims to provide preliminary evidence and methodological insights for future large-scale validation studies. Materials and methods Study cohort and clinical context This exploratory study aimed to develop a multimodal risk prediction model for recurrence in CRC patients following curative surgery and during/after the completion of XELOX adjuvant therapy. The envisioned clinical application, if validated, would be to identify high-risk patients for intensified surveillance(e.g., enhanced imaging frequency, closer ctDNA monitoring) during postadjuvant follow-up. This is a retrospective cohort study with a prospectively defined analytical plan for model development and validation. An exploratory clinical dataset from 86 patients who were ultimately diagnosed with colorectal cancer was constructed. To avoid the influence of different treatment strategies on variable selection, only patients who received XELOX-based adjuvant therapy were included in the study. These patients were from the Affiliated Hospital of Hebei University, Baoding, Hebei Province, China, from August 2020 to August 2023. Patients were included on the basis of the following criteria: (1) histologically confirmed primary colorectal adenocarcinoma; (2) complete clinicopathological staging data, including tumor size (T stage), lymph node involvement (N stage), and metastasis status (M stage); (3) documented records of pre- and posttreatment serum biomarkers carcinoembryonic antigen (CEA) and carbohydrate antigen 19 − 9 (CA19-9); (4) availability of follow-up data on recurrence status; and (5) completion of standard surgical resection with XELOX-based adjuvant therapy. (6) The exclusion criteria were as follows: (1) incomplete clinical or pathological records (e.g., missing T/N/M staging, biomarker values, or recurrence status); (2) the presence of synchronous malignancies or nonadenocarcinoma histology; (3) incomplete adjuvant therapy cycles or loss to follow-up; and (4) perioperative mortality within 30 days. Data, including demographic details, tumor characteristics (e.g., lymphovascular invasion, perineural invasion, and tumor budding grade), treatment modalities (surgery and adjuvant therapy), longitudinal blood test results and circulating tumor cell(CTC) counts, were extracted from the patients’ electronic medical records. Biomarker thresholds were defined as follows: elevated CEA (> 5 ng/mL) and CA19-9 (> 37 U/mL) levels were classified as abnormal. Lymph node involvement (N stage) was categorized as N0 (no metastasis), N1 (1–3 nodes), N2 (4–6 nodes), or N3 (> 7 nodes) on the basis of the TNM guidelines. Recurrence was confirmed by imaging or histopathology. All staging information was based on pathological staging (pTNM) following surgical resection. To ensure data quality, all liquid biopsy assays were standardized using the same platforms throughout the study: the GEDI microfluidic platform for CTC enumeration and the OncoScreen target panel for ctDNA detection21. Following the manufacturer’s protocols to minimize batch effects and technical variability was employed. Ethical approval and informed consent were obtained in accordance with institutional review board protocols. A comprehensive table of baseline clinical and sociodemographic characteristics(Table 1) and a CONSORT-style flowchart illustrating patient selection(Supplementary File1) are provided to ensure transparent reporting of cohort characteristics and the selection process. Table 1Baseline clinical and pathological characteristics of the study cohort (N = 86).CharacteristicCategoryNumber of Patients (%) or Median (Range) Age (years) 62 (21–84) Sex Male62 (72.09%)Female24 (26.91%) Tumor Location Colon51 (59.30%)Rectum31 (36.04%)Colorectum (Unspecified)4 (4.65%) Pathological Features Grade of Differentiation Well/Moderate57 (67.28%)Poor8 (9.30%)Missing/Not specified21 (24.41%) Lymphovascular Invasion Present15 (16.27%)Absent66 (76.74%) Perineural Invasion Present19 (22.09%)Absent62 (72.09%) Tumor Budding Grade Low-Grade57 (67.28%)Intermediate-Grade17 (19.76%)High-Grade12 (13.95%) Pathological Stage (AJCC) clinical Stage I4 (4.65%)II46 (53.48%)III29 (33.72%)IV7 (8.13%) T Stage T1-T214 (16.27%)T333 (38.37%)T439 (45.34%) N Stage N045 (52.32%)N126 (30.23%)N213 (15.11%)N32 (2.32%) Preoperative Tumor Markers CEAElevated (> 5 ng/mL)18/58 (31.0%)CA19-9Elevated (> 37 U/mL)15/55 (27.3%) Circulating Tumor Cells (CTC) 026 (30.23%)≥ 125 (29.07%)Not Tested/Missing35 (40.70%) Follow-up & Recurrence Confirmed Recurrence/Metastasis17 (19.77%)No Recurrence69 (80.23%) Baseline clinical and pathological characteristics of the study cohort (N = 86). Dataset preparation A consistent data splitting strategy was employed across all the models to ensure comparable performance evaluation. The dataset was partitioned into 70% training and 30% testing sets using stratified sampling to preserve the event rate distribution. Given the limited sample size and exploratory nature of this study, we employed repeated stratified k-fold cross-validation rather than a single hold-out validation set, as this provides more robust performance estimation in small samples. Data preprocessing and cleaning were performed via R(version 4.3.1) with the tidyverse and janitor packages. The raw dataset containing 86 colorectal cancer patients was imported and processed to address inconsistencies, including the removal of nonprintable characters(e.g., NUL characters) and the retention of valid ASCII text. Data boundaries were dynamically identified by detecting header rows and excluding metadata annotations. Missing values, encoded as “NA” or invalid annotations, were systematically addressed. Any variable with > 30% missing data was excluded entirely from the analysis. Among the included variables, owing to the missing data rate being less than 30% and the small number of missing values, we adopted a complete case analysis approach. For variables retained(< 30% missing), a complete-case analysis approach was used. The final dataset of 11 variables had minimal missing data that were handled through complete-case analysis. For the final multivariable analysis, several continuous or ordinal predictors were dichotomized on the basis of clinical relevance and observed recurrence rates. Neurological invasion(a5) and first postoperative ctDNA status(a14) were categorized as “present” (any non-zero value) versus “absent” (zero). Tumor budding (a6) was dichotomized as high- grade(Grade 4) versus low/intermediate- grade(Grade 1- 3) because Grade 4 exhibited a substantially higher recurrence rate (54.5%) than lower grades (< 12%). T stage(a2) was dichotomized as T5 versus T1-4; the recurrence rate in T5 patients(12/45, 26.7%) was more than twice that in T1-4 patients (5/41, 12.2%), justifying this split. After dichotomization, the final cohort comprised 86 patients whose complete data included all four variables. Variable selection via LASSO regression LASSO regression was used solely for dimension reduction to mitigate overfitting, not as a final prediction model. With 17 events available, we aimed to select a parsimonious set of predictors. Using the glmnet package in R, we performed 10-fold cross-validation to determine the optimal penalty parameter(lambda.min). Variables with non-zero coefficients at lambda.min were retained as candidate predictors. This step is justified in our small-sample setting because LASSO applies regularization that shrinks irrelevant coefficients to zero, reducing the risk of spurious associations. Gradient boosting machine (GBM) modeling GBM was implemented using the gbm, caret, and ROSE R packages. Hyperparameters were set as follows: number of trees = 100, learning rate(shrinkage) = 0.1, interaction depth = 3, and minimum node size = 5. Repeated 5-fold cross-validation (5 repeats) was used for training. No additional variable selection was performed beyond the LASSO-selected set. Random forest modeling Random forest was implemented using the randomForest and caret packages. A grid search over mtry (2, 3, 4) was conducted with 5-fold cross-validation(3 repeats). The final model used 200 trees. The same LASSO-selected predictor set was used as input. Firth’s regression model Given the low number of recurrence events(n = 17) relative to the number of candidate predictors (events-per-variable ratio < 5), standard maximum likelihood estimation may produce biased coefficients and overly narrow confidence intervals. To mitigate these issues, we employed Firth’s penalized likelihood logistic regression using the logistf R package. The predictors entered into the Firth model were the intersection of variables selected by both GBM and random forest. Specifically, four variables were consistently retained by both methods: postoperative ctDNA status, perineural invasion, tumor budding(Grade 4 vs. Grade 1–3), and T stage(T5 vs. T1–4). These four variables were entered simultaneously into the Firth model; no stepwise selection or further filtering was applied.Odds ratios(ORs) and 95% confidence intervals(CIs) were calculated from the penalized estimates. DCA regression DCA was performed to evaluate the clinical utility of the Firth logistic regression model. Net benefit was calculated across threshold probabilities(0–1, step 0.01) using the dcurves R package, comparing the model against “treat all” and “treat none” strategies. Statistical methods Given the exploratory nature and limited sample size, all analyses should be interpreted as hypothesis-generating. All statistical tests were two-sided, and p < 0.05 was considered nominally significant without adjustment for multiple testing, consistent with the exploratory aim. Analyses were performed in R version 4.3.1 using the pROC, logistf, randomForest, gbm, caret, and dcurves packages. Random seeds were set(set.seed(123)) for reproducibility. Results Variable selection via LASSO regression methods As a prospective exploratory analysis, our primary focus was on evaluating the feasibility of multimodal integration rather than establishing definitive clinical thresholds. Among the 86 patients analyzed, 17 experienced recurrence during follow-up, yielding an event rate of 19.8%. With 6 final predictors selected by LASSO, the event per variable (EPV) ratio was approximately 2.83. All subsequent results focus exclusively on these six LASSO-selected variables. To ensure the robustness of the model, we employed 10-fold cross-validation to determine the optimal regularization parameter (lambda). This process involved evaluating the model’s performance across different lambda values to identify the one that minimized the cross-validation error. The optimal lambda value (Lambda.min) was determined to be 0.0267387, which corresponded to the minimum cross-validation error (Fig. 1A). The coefficient paths of the LASSO regression model were visualized to illustrate the regularization process. As shown in the coefficient path plot, the coefficients of different variables changed with the logarithm of lambda (Log(λ)). With increasing Log(λ), the coefficients of some variables decreased to zero, indicating that these variables were excluded from the model. The paths of the coefficients help us understand the behavior of the variables during the regularization process (Fig. 1B). Using this optimal lambda value, we extracted the nonzero coefficient variables from the LASSO regression model. The variables selected at this optimal lambda value were considered the most significant predictors in the dataset. The selected variables included clinical stage, T stage, N stage, neurological invasion status, tumor budding stage, CTC number and the first ctDNA detection results. In terms of variable importance, the magnitude of the LASSO regression coefficients was used as an indicator. The coefficients reflect the relative importance of each variable in predicting the outcome, with larger absolute values suggesting a stronger influence. The sorted coefficient magnitudes are presented in Fig. 1C. Variable selection via LASSO regression methods. (A) Cross-validated misclassification error vs. the regularization parameter (Lambda). (B) Lasso coefficient trajectories. (C) Bar plot of the importance of the selected variables. GBM modeling results The selected variables were subsequently used for clinical model construction. Given its robust performance in predicting tumor recurrence, the gradient boosting machine (GBM) model was used first. XGBoost analysis revealed critical predictors of the outcome, with variable importance scores ranging from 0.00 to 0.20 (Fig. 2A), where the tumor budding grade (a6), neurological invasion status (a5), and first ctDNA detection results (a14) emerged as the three most influential variables. Other clinical indices, such as treatment time (a13), advanced clinical stage (a1), degree of differentiation (a4), N stage (a4) and T stage (a3), were also ranked as important variables. The threshold selection curve indicated a balanced trade-off between sensitivity (0.75) and specificity (0.72) at a threshold of 0.25. The curve also indicates that the model can effectively discriminate between positive and negative class recognition capabilities at this threshold (Fig. 2B). The F1 score reached its peak value of 0.8 at a threshold of 0.3, highlighting the model’s superior ability to harmonize precision and recall for positive class identification (Fig. 2C). This divergence underscores the model’s adaptability to diverse operational requirements: a lower threshold (0.25) prioritizes the error control, whereas a higher threshold (0.3) maximizes the detection efficacy of critical cases. The clinical model was constructed via a GBM diagram. (A) Bar plot of the importance in the model. (B) Threshold selection curve (threshold = 0.25). (C) The relationship curve between the F1 score and threshold. Random forest modeling results The random forest analysis, which uses MeanDecreaseAccuracy (≥ 2) and MeanDecreaseGini (≥ 1) as thresholds, indicated that the selected variables were ranked in Fig. 3A. On the basis of the feature importance rankings, clinical stage and N classification presented the highest mean decrease in accuracy values, followed by T classification, CTC status, and first ctDNA detection; neurological invasion status and tumor budding stage were also ranked prominently in terms of the mean decrease in the Gini score. The error rate visualization demonstrated a steady decline from an initial rate of 0.8 to stabilization at approximately 0.2 after 500 trees, with minimal fluctuations beyond 200 trees, confirming model convergence and robustness (Fig. 3B). These results indicate that the dominance of pathological staging, tumor microenvironment characteristics, and liquid biopsy-derived markers can be efficiently used to predict tumor recurrence, which is supported by the model’s low final error rate and absence of overfitting. The clinical model was constructed via a random forest diagram (A) Feature importance scores for the input variables ranked by the MeanDecreaseGini and MeanDecreaseAccuracy methods. Higher values indicate stronger contributions to model predictions. (Threshold: MeanDecreaseAccuracy ≥ 2; MeanDecreaseGini ≥ 1) (B) The error rate of a random forest varies with the number of trees. Firth’s regression modeling The final Firth regression model included four predictors: postoperative ctDNA status, perineural invasion, tumor budding (Grade 4 vs. Grade 1–3), and T stage (T5 vs. T1–4). These variables represent the intersection of those selected by GBM and random forest after LASSO coarse screening. Results are presented in Table 2; Fig. 4. The ctDNA positivity was associated with an increased odds of recurrence (OR = 3.95; 95% bootstrap CI: 0.83–8.61). Perineural invasion (OR = 2.52; 95% CI: 0.48–5.72) and high-grade tumor budding (OR = 2.19; 95% CI: 0.43–10.65) also showed positive associations. However, all confidence intervals were very wide, reflecting the limited statistical precision due to the small sample size and low event count. The p values(ranging from 0.026 to 0.038) should be interpreted as exploratory and hypothesis-generating, not as confirmatory evidence. T stage(T5 vs. T1–4) exhibited an unexpected inverse association(OR = 0.53; 95% CI: 0.15–1.71). Table 2Multivariable Firth logistic regression models for recurrence.VariablesORCI_lowerCI_upperp_valuectDNA (positive vs. negative)3.950.838.610.0379Perineural invasion (present vs. absent)2.520.485.720.0296Tumor budding (Grade4 vs. Grade1-3)2.190.4310.650.0323T stage (T5 vs. T1-4)0.530.151.710.0264 Note: Estimates are from Firth’s penalized likelihood estimation. Wide confidence intervals reflect the limited number of events (n = 17) and possible separation. Multivariable Firth logistic regression models for recurrence. Note: Estimates are from Firth’s penalized likelihood estimation. Wide confidence intervals reflect the limited number of events (n = 17) and possible separation. Coefficients of the final Firth logistic regression model. Points represent coefficient estimates on the log-odds scale; horizontal lines indicate 95% confidence intervals. The variables are ordered by the magnitude of the coefficient. ctDNA positivity and perineural invasion were positively associated with recurrence, whereas T stage (T5 vs. T1-4) was negatively correlated. Wide confidence intervals reflect the limited sample size (n = 86, 17 events) and the resulting statistical uncertainty. Evaluation of the clinical model The ROC curve of the regression model supported its robust predictive capacity (AUC = 0.695) (Fig. 5A). The Hosmer-Lemeshow goodness-of-fit test indicated no significant departure from perfect calibration (χ²=3.33, df = 4, p = 0.503). However, visual inspection of the calibration plot (Fig. 5B) revealed some systematic deviations: the model tended to overestimate risk in the low-to-intermediate range (a predicted probability of 0.10–0.45) and underestimate risk at the high end (a predicted value of 0.50). Given the limited number of events (n = 17), the power of the Hosmer-Lemeshow test is low, and graphical assessment may be more informative in this exploratory setting. Decision curve analysis indicated a positive net benefit across threshold probabilities of 0.10–0.40, suggesting potential clinical utility(Fig. 5C). These performance estimates should be interpreted as exploratory given the modest sample size, and external validation in larger cohorts is warranted. Performance evaluation of the final prediction model. (A) Receiver operating characteristic (ROC) curve. The model achieved an area under the curve (AUC) of 0.695 (bootstrap 95% CI: 0.63–0.76). (B) Calibration plot. The dashed diagonal line represents perfect calibration; the solid line (with 95% confidence interval) shows observed versus predicted probabilities using loess smoothing. The Hosmer-Lemeshow test was non-significant (χ² = 3.33, df = 4, p = 0.503), indicating no significant miscalibration at the global level, although local deviations were visible. (C) Decision curve analysis. The model’s net benefit (solid blue line) is compared with the “treat all” (dashed line) and “treat none” (horizontal line) strategies across threshold probabilities. Positive net benefit was observed in the 0.10–0.30 range, suggesting potential clinical utility. Discussion This prospective exploratory study provides preliminary evidence supporting the feasibility of integrating multi modal data for CRC recurrence risk prediction. Our findings should be interpreted as generating hypotheses for future validation rather than providing definitive clinical guidance. The enrolled variables included clinical, pathological stage, molecular biomarkers, and treatment-related variables. Given the complexity, incompleteness, and redundant information of the dataset, LASSO regression was initially employed for variable selection. The selected variables were subsequently used for clinical model construction via three machine learning diagrams—XGBoost(XGBM), random forest(RF), and logistic regression(LR). Our results identify the critical variables associated with integrating pathological staging, several clinical features, and liquid biopsy markers in MRD risk stratification. Our study also reveals methodological challenges inherent in evaluating the risk of cancer recurrence. The variable importance analyses across the three models consistently identified pathological staging parameters (clinical stage and T/N stage), tumor pathology features (neurological invasion status and tumor budding), and liquid biopsy markers (CTC status and first ctDNA detection) as dominant predictors of recurrence. These findings are consistent with emerging evidence emphasizing the prognostic value of integrating the broad TNM stage with dynamic molecular and microenvironmental features22,23. In addition, neurological invasion and tumor budding status, which are strongly associated with both the Gini impurity (RF) index and coefficient magnitude (XGBM), are strongly associated with aggressive tumor behavior, metastatic potential and poor prognosis24,25. Similarly, ctDNA and CTC detection demonstrated robust predictive power across the models, which is consistent with its role as a sensitive indicator of MRD and early recurrence14,15. The final model included four prespecified variables—postoperative ctDNA status, perineural invasion, tumor budding (Grade 4 vs. Grade 1–3), and T stage (T5 vs. T1–4)—selected on the basis of prior literature and clinical rationale. Although none of the adjusted odds ratios reached conventional statistical significance (all p > 0.05), the direction of effects for ctDNA (OR = 3.95), perineural invasion (OR = 2.52), and tumor budding (OR = 2.19) was consistent with increased recurrence risk. These findings were also consistent with a substantial body of evidence14,24,25,. The wide confidence intervals reflect the limited number of events (n = 17) and the presence of separation, which are common challenges in early-phase exploratory studies. T stage exhibited an unexpected inverse association (OR = 0.53), which may stem from the specific dichotomization (T5 vs. T1-4) or from residual confounding; these findings underscore the instability of the estimates in small samples and should not be overinterpreted. Our two-stage selection approach (LASSO for coarse dimension reduction, followed by parallel GBM and random forest screening) was designed to identify a parsimonious set of predictors in a small-sample setting. The final Firth regression model was built using only the intersection of variables selected by GBM and random forest, rather than all LASSO-selected variables. This conservative strategy reduces the risk of including noise predictors and improves interpretability. Firth’s penalized likelihood was then applied to this small set of four predictors to obtain bias-reduced coefficient estimates. Despite these precautions, the wide confidence intervals and the unexpected coefficient direction for T stage indicate that residual instability remains, underscoring the exploratory nature of the analysis. Recent methodological work by van Smeden et al. has challenged the rigid application of the events-per-variable(EPV) rule, demonstrating that predictive performance depends on a combination of factors, including total sample size, event fraction, and the strength of predictor–outcome associations26. Our approach aligns with this nuanced view: by reducing the dimensionality via ML screening and applying Firth’s correction, we obtained estimates that are more stable than standard logistic regression would provide while retaining interpretability. Similar strategies have been successfully employed in other small-sample settings, such as hip fracture mortality prediction and early GDM risk assessment27,28. The final model demonstrated moderate discriminative ability (AUC = 0.695; 95% CI: 0.63–0.76), which is lower than the 0.82 originally reported but more realistic given the sample size. This attenuation highlights the risk of optimistic performance estimates when internal validation is insufficient—a key lesson from our analysis. The calibration plot revealed systematic overestimation of risk in the low-to-intermediate range and underestimation at the high end, yet the Hosmer-Lemeshow test was not significant (p = 0.503). This apparent discrepancy illustrates the low power of global goodness-of-fit tests in small samples and reinforces the importance of visual inspection. Decision curve analysis suggested potential clinical utility for threshold probabilities between 0.10 and 0.30, but these findings must be validated in external cohorts before any clinical application. Our findings also support several prior studies advocating multimodal integration for CRC prognostication. For example, the emphasis on the first positive ctDNA result potentially aligns with a study that demonstrated that postoperative early ctDNA positivity predicted increased recurrence risk29. Furthermore, we evaluated the contribution of ctDNA detection time to cancer recurrence. The results indicated that only the earliest positive result can efficiently predict CRC recurrence. Although XELOX therapy was continuously employed, a strong correlation between positive results and recurrence was also established. These findings potentially indicate that more efficient therapeutic strategies should be considered for treating ctDNA-positive patients. Similarly, the prominence of tumor budding has been established as an independent prognostic indicator in stage II CRC30. However, our models diverge from previous works in two key aspects, namely, liquid biopsy and traditional biomarkers. Unlike studies that rank CEA or CA19-9 as important predicating markers, our models relegate these serological markers to secondary roles, although they have been included in statistics. This may be due to their limited sensitivity and indirect relationship with MRD31. On the other hand, another important liquid biopsy marker, the CTC number, was not always ranked as an important marker by the three diagrams. This may be related to the low detection rate and diverse heterogeneous cell populations in clinical practice16,32. The primary strength of this study is its methodological transparency. By explicitly reporting the challenges encountered—separation, wide CIs, and calibration deviations—we provide a realistic account of exploratory modeling in small samples. The use of Firth regression and bootstrap resampling enhances the robustness of our estimates. However, several limitations must be acknowledged. First, the small sample size (86 patients, only 17 events) is the most critical limitation. It results in a low events-per-variable ratio (approximately 2.8 for the final model), which leads to wide confidence intervals, unstable coefficient estimates, and an increased risk of overfitting. The unexpected direction of the T stage coefficient is a direct manifestation of this instability. Second, despite the use of Firth’s penalized likelihood, the model remains statistically fragile; external validation in a much larger cohort would be required to obtain stable estimates. Third, the single-center retrospective design introduces potential selection bias and limits generalizability. Fourth, the dichotomization of continuous or ordinal predictors (e.g., T stage) may have contributed to information loss and unexpected effect directions. Fifth, we were unable to include CTC count due to missing data, precluding assessment of this potentially important marker. Sixth, the use of recurrence as a binary outcome ignores time-to-event information; survival analysis would be more informative but requires larger samples. Finally, all performance metrics (AUC, calibration, DCA) are derived from internal validation only; external validation in independent cohorts is essential before any clinical application. Readers should interpret the findings as exploratory and hypothesis-generating, not as a clinically deployable prediction tool. In summary, while our findings are hypothesis-generating, the methodological framework presented here offers a roadmap for conducting rigorous exploratory modeling in small datasets. We hope that this work will stimulate further methodological refinement and ultimately contribute to the development of clinically useful risk stratification tools for colorectal cancer. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions All the authors made substantial contributions to the study conception and design and the acquisition, analysis, and interpretation of the data; Zhihan L and Hongqing Ma drafted the article and revised it critically for important intellectual content; and Wenbo Niu conducted the statistical analysis. Tao Zhang conducted the data visualization. Wenbo Niu revised the manuscript. Wenbo Niu and Tao Zhang supplied funding for the study. Zihan Fan and Simeng Jiang interpreted mainly the data. All the authors agreed to submit the final manuscript to the current journal; Funding This study was supported by the Medical Science Research Project of Hebei (Project No. 20240186) Data availability Deidentified data and the analysis code used in this study are available from the corresponding author upon reasonable request, subject to the execution of appropriate data sharing agreements and institutional approval. Competing interests The authors declare no competing interests. Ethics approval and consent to participate The present study was approved by the Ethics Committee of the Fourth Hospital of Hebei Medical University. The research was performed in accordance with the World Medical Association Declaration of Helsinki. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Zhihan Li and Hongqing Ma contributed equally to this work. Contributor Information Wenbo Niu, Email: 47900878@hebmu.edu.cn. Tao Zhang, Email: zhtdou@163.com. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Deidentified data and the analysis code used in this study are available from the corresponding author upon reasonable request, subject to the execution of appropriate data sharing agreements and institutional approval.