Colorectal cancer sidedness: prognostic implications and the predictive role of artificial intelligence Authors: Marcelo Portes Rocha Martins, Rafaela Lopes de Figueiredo Andrade, Pedro Henrique Villar Delfino, Laurence Rodrigues do Amaral, Letícia da Conceição Braga, Roberta Rayra Martins-Chaves Published: 2026 ABSTRACT BackgroundColorectal cancer (CRC) is a biologically heterogeneous disease in which tumor sidedness has emerged as a relevant prognostic factor. Conventional TNM staging does not incorporate several clinically and biologically meaningful variables that may influence outcomes. In this context, artificial intelligence (AI) based approaches offer an opportunity to integrate complex clinicopathological data and improve prognostic stratification. This study aimed to evaluate clinicopathological variables associated with tumor sidedness and to identify clinical predictors of high-risk disease using an AI-based decision-tree model.MethodsThis retrospective cohort study included 71 adults who underwent surgical resection for colorectal adenocarcinoma at a tertiary oncology center between 2020 and 2024 and had complete clinicopathological data available for analysis. Overall and progression-free survival were estimated using the Kaplan–Meier method, and associations between categorical variables were assessed using Fisher’s exact test. Decision-tree models were constructed using the J48 (C4.5) algorithm, and model performance was evaluated by leave-one-out cross-validation (LOOCV).ResultsLeft-sided tumors were predominant and more frequently associated with alcohol ingestion (p = 0.04), the use of neoadjuvant chemoradiotherapy (p < 0.01), and higher mortality (p = 0.04), despite more intensive treatment strategies. Right-sided tumors were prevalent in women and were associated with angiolymphatic invasion. In prognostic modeling, positive surgical margins emerged as the strongest predictor of mortality (Full 85.18%; LOOCV 74.07%). Among patients with negative margins, tumor laterality represented the most influential prognostic factor, with right-sided tumors associated with improved survival. Interestingly, younger patients showed shorter progression-free survival (Full 89.09%; LOOCV 76.36%).ConclusionsTumor sidedness constitutes a meaningful prognostic dimension in CRC when integrated with established pathological factors. AI-based decision-tree models can capture clinically coherent prognostic signatures and complement traditional staging systems, supporting their role as hypothesis-generating tools for individualized risk assessment and guiding future prospective validation. Introduction Colorectal cancer (CRC) ranks third worldwide in incidence and is the second major cause of mortality, following lung cancers [1]. Risk factors for CRC can be broadly classified into modifiable and non-modifiable factors. Modifiable risk factors include alcohol [2] and tobacco consumption, psychological stress, obesity [3], and increased intake of red meat, smoked foods [2]. Non-modifiable risk factors comprise advanced age, male sex, the presence of inflammatory bowel disease, genetic conditions such as familial adenomatous polyposis, a family history of CRC in first-degree relatives, and prior exposure to abdominal or pelvic radiotherapy [3]. The multifactorial etiological nature of CRC results in marked molecular heterogeneity, primarily driven by three major carcinogenic pathways: chromosomal instability (CIN), characterized by sequential mutations in APC, KRAS, TP53, and SMAD4; microsatellite instability (MSI), arising from mismatch repair deficiency due to germline mutations or MLH1 promoter hypermethylation and the CpG island methylator phenotype (CIMP), frequently associated with BRAF mutations and the serrated pathway [4]. In addition to these pathways, alterations in PIK3CA, ultra-mutated tumors harboring POLE/POLD1 mutations, and additional epigenetic changes further contribute to the biological diversity of CRC according to the World Health Organization Classification of Tumors Editorial Board [5]. This interplay of genetic and epigenetic mechanisms underpins clinically relevant differences in tumor behavior, which are also differentially expressed across colon tumor locations [6]. Right-sided colon tumors (RCC) exhibit a higher frequency of PIK3CA mutations, PTEN loss, and KRAS and BRAF alterations, as well as a high prevalence of MSI and mismatch repair deficiency. These features are associated with an increased mutational burden, enhanced neo-antigen generation, and a prominent tumor-infiltrating lymphocytic response [7]. Conversely, left-sided colon tumors (LCC) are linked to APC mutations, an early and defining event in the Wnt/β-catenin signaling pathway, as well as later alterations involving TP53 and, in a subset of cases, HER2 amplification [8, 9]. Collectively, these molecular differences support the existence of distinct biological pathways driving proximal and distal colorectal carcinogenesis, with important implications for prognosis and therapeutic decision-making [7]. Distinct clinical characteristics according to tumor sidedness have also been previously reported [10]. Approximately 83% of early-onset CRC occurring in individuals without a family history were diagnosed in the LCC. In the same study, RCC was more frequently observed in older individuals and in females [11]. Furthermore, RCC predominantly originated from sessile and/or serrated polyps and tended to present as flatter lesions with a mucinous component, increased T-cell infiltration, and a higher incidence of peritoneal metastases. In contrast, LCCs more commonly affected younger, male patients and were typically sporadic tumors arising from tubular or villous polyps, presenting with a polypoid or exophytic growth pattern into the bowel lumen, with lower immunogenicity and a greater propensity for distant metastases [11, 12]. Alcohol and tobacco consumption have also been more strongly associated with distal colon and rectal tumors [13, 14]. Taken together, these consistent clinicopathological and behavioral differences according to tumor sidedness highlight a complex and multifactorial landscape that extends beyond anatomical location, underscoring the need for integrative analytical approaches capable of capturing non-linear interactions among diverse prognostic variables, an area in which artificial intelligence–based methods have shown particular promise [15–17]. Machine learning (ML) can detect complex patterns in large clinical datasets. These algorithms are trained on previously labeled datasets, enabling them to learn relationships that can subsequently be used to predict outcomes in new patients [18]. Predictive methods, such as decision-tree models, have demonstrated high accuracy in forecasting CRC recurrence and in identifying carcinoembryonic antigen (CEA) levels, tumor location, depth of invasion, lymphatic and venous invasion, and lymph node involvement as relevant predictors [19]. A decision-tree model was developed using clinical indicators and preoperative blood cell parameters, demonstrating superior performance in distinguishing benign from malignant ovarian tumors compared with individual imaging or biomarker assessments [20]. Decision-tree models stand as a foundational paradigm in, uniquely positioned at the intersection of computational rigor and clinical epistemology. Unlike ‘black box’ models such as neural networks, which often obscure the decision-making process, decision-tree models operate as ‘white box’ models, offering high explainability. By generating explicit IF-THEN classification rules, the model allows clinicians to intuitively validate the algorithmic logic against established medical knowledge, facilitating the trustworthy translation of these findings into clinical practice [20, 21]. In this context, the clinicopathological characteristics that distinguish RCC and LCC constitute a meaningful set of variables that can be incorporated into predictive algorithms to stratify risk of recurrence, disease progression, and mortality. Therefore, the objective of the present study is to systematically characterize the clinicopathological profile of CRC by tumor sidedness, thereby providing a foundation for developing more accurate and clinically applicable intelligent models, as well as for our own AI-based predictive model. Materials and methods Study design This is an observational retrospective cohort study conducted among individuals aged 18 years or older who underwent surgical treatment at Instituto Mário Penna and had a histopathological diagnosis of colorectal adenocarcinoma. Sample selection A total of 71 patients with complete clinicopathological data who agreed to participate in the study were selected from a cohort of 228 individuals who underwent surgeries classified as partial colectomies and/or abdominal rectosigmoidectomies with curative intent at Instituto Mário Penna (Belo Horizonte, Minas Gerais, Brazil) between 2020 and 2024. Complete clinicopathological data were available and analyzed for 71 of these patients. Clinicopathological information was obtained from the patients’ electronic medical records. Inclusion criteria comprised a histopathological diagnosis of colorectal adenocarcinoma, absence of distant metastatic disease, adequate clinical performance status, age ≥ 18 years, and surgical specimens confirming the retrieval of at least 12 lymph nodes during lymphadenectomy. Exclusion criteria included non-adenocarcinoma histological subtypes, prior radiotherapy, metastatic disease at diagnosis, synchronous or metachronous malignancies at the time of surgery, receipt of chemotherapy within five years before sample collection, recurrent CCR, and lack of informed consent. Ethics approval and consent to participate This study underwent full ethical review and was approved by the Research Ethics Committee of Instituto Mário Penna (Belo Horizonte, Brazil), under CAEE number: 25004619.3.3001.5121. All participants provided written informed consent. The study was conducted using anonymized data, ensuring confidentiality and the protection of participants’ identities. All procedures were conducted in accordance with the ethical standards of the institutional and national committees responsible for human experimentation and in compliance with the principles of the Declaration of Helsinki of 1975, as revised in 2008. Data analysis Clinicopathological variables were systematically analyzed, including histological subtype, tumor stage and grade, perineural and angiolymphatic invasion, the presence of small infiltrative tumor cell clusters at the tumor–stroma interface consistent with an active invasion phenotype (tumor budding), the number of positive lymph nodes relative to the total number of dissected nodes, CEA levels, and the use of chemotherapy and radiotherapy. Survival outcomes, including progression-free survival and overall survival, were assessed as time-to-event variables. Survival probabilities were estimated using the Kaplan–Meier method, and comparisons between survival curves were performed when appropriate. Associations between categorical variables were assessed using Fisher’s exact test. Statistical analyses were conducted using GraphPad Prism software (version 9.0), and a two-tailed p-value < 0.05 was considered statistically significant. Supervised and unsupervised ML techniques were employed to explore and predict clinicopathological data from patients with CRC. Decision-tree models and corresponding IF–THEN classification rules were developed using established methodologies from the literature. Decision-tree construction and class relationship analysis were performed using the J48 algorithm, an implementation of the C4.5 algorithm, in WEKA (Waikato Environment for Knowledge Analysis, version 3.8.3; University of Waikato, New Zealand) [22, 23]. The hierarchical structure of the decision-tree models were driven by computation. In building the decision tree models, pre-modeling collinearity tests were omitted due to the nature of the J48 algorithm. At each step, the algorithm uses the ‘Information Gain Ratio’ based on entropy, evaluating all available clinicopathological variables to automatically select the one that best splits the patient data into homogeneous outcome classes. Node splitting thresholds were computationally derived using the algorithm’s default entropy minimization protocols. Therefore, the root and subsequent nodes reflect a mathematical hierarchy of predictive strength, capturing the most discriminative variables at each specific level of the patient subgroup. To improve the representation of the minority class in the training set, we used ‘Cost-Sensitive Learning’ during the modeling phase. The penalty weights for the minority classes were systematically assigned based on the inverse frequency ratio of the target outcomes, effectively balancing the influence of both groups during the training phase. Model performance, robustness, and generalizability were assessed using leave-one-out cross-validation (LOOCV), with predictive models achieving LOOCV overall accuracy values greater than 70% considered acceptable for further interpretation. To ensure a comprehensive evaluation of model performance, we calculated Sensitivity, Specificity, Matthews Correlation Coefficient (MCC) to indicate the correlation between predicted and observed data, and the F1-Score to assess the model’s ability to correctly identify the class. Additionally, the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) was used to evaluate the overall discriminative power of the decision-tree models, independent of the classification threshold. Results Sample characterization and clinicopathological profile of CCR patients We analyzed the complete clinicopathological data of 71 patients diagnosed with colorectal adenocarcinoma (Table 1). The cohort comprised 37 females (52%) and 34 (48%) males, with ages ranging from 24 to 95 years (mean 61 ± 12). A significant majority of tumors were located on the left side of the colon, accounting for 61 cases (86%), while 10 tumors (14%) were located on the right side. The assessment of tumor laterality in relation to key clinicopathological variables revealed several descriptive patterns. The RCC occurred more frequently in women, whereas the LCC occurred more frequently in men (p = 0.08). Although no clear age predilection was identified, 39 patients with CCR were aged 62 years or younger. Interestingly, younger patients (≤ 62 years) showed a trend toward reduced overall survival (p = 0.07) and significantly shorter progression-free survival (p = 0.01) (Fig. 1A and B). Distributions of tumor size, angiolymphatic invasion, perineural invasion, and tumor budding were broadly comparable across tumor sides. Similarly, CEA levels and the presence of blood in the stool were observed at comparable frequencies in both groups. Collectively, these findings indicate that baseline clinicopathological features were balanced between tumor laterality groups. Fig. 1The outcomes at 60 months after treatment, by age group (≤ 62 years vs. >62 years). (a) Kaplan-Meier curves showing overall survival stratified by age. Time is expressed in months. Survival distributions were compared using the log-rank (Mantel-Cox) test (P = 0.0775). (b) Kaplan-Meier curves showing progression-free survival stratified by age. Time is expressed in months. Survival distributions were compared using the log-rank (Mantel-Cox) test (P = 0.0197). The median follow-up time for the entire cohort was 41 months (range, 3–64 months). Over the study period, 8 patients (1 RCC and 7 LCC) were censored due to loss to follow-up The outcomes at 60 months after treatment, by age group (≤ 62 years vs. >62 years). (a) Kaplan-Meier curves showing overall survival stratified by age. Time is expressed in months. Survival distributions were compared using the log-rank (Mantel-Cox) test (P = 0.0775). (b) Kaplan-Meier curves showing progression-free survival stratified by age. Time is expressed in months. Survival distributions were compared using the log-rank (Mantel-Cox) test (P = 0.0197). The median follow-up time for the entire cohort was 41 months (range, 3–64 months). Over the study period, 8 patients (1 RCC and 7 LCC) were censored due to loss to follow-up Lifestyle exposures, treatment, and outcomes of CCR patients associated with tumor-sidedness Tobacco exposure was more prevalent in LCC (46%) compared to RCC (10%) patients (p = 0.09). Alcohol consumption further differentiated LCC, with current or recent use being significantly more prevalent in this group (p = 0.04). Therapeutic strategies differed markedly according to tumor laterality. Patients with LCC were substantially more likely to receive neoadjuvant chemotherapy (p < 0.01), whereas RCC were commonly managed with primary surgery followed by adjuvant treatment. Despite this intensified therapy and a statistically more extensive lymph node dissection being performed on LCC (p < 0,01), clinical outcomes remained less favorable for this group. Patients harboring LCC experienced a markedly higher death rate than those with RCC (p = 0.04). Although tumor recurrence and distant metastases occurred more frequently in LCC, these events did not clearly segregate by laterality. Table 1Association between tumor sidedness and clinicopathological variables in 71 CCR patientsClinical-pathological FeaturesCategoryRCC (n)LCC (n)p-valueSexMale2320.08Female829Age≤ 62 years5340.74> 62 years527Tumor size0–5 cm6480.336–10 cm411> 11 cm02GradeWell-differentiated1140.50Moderately differentiated942Poorly differentiated05Angiolymphatic invasionYesNo4636250.31Perineural invasionYes8400.48No221Tumor buddingYes2170.71No844Number of lymph nodes retrieved12–2713460.00428–4269> 4316TNM stagingI/IIIII/IV3729320.49CEA (ng/mL)≤ 34350.563–529> 5417ChemotherapyNo chemotherapy313< 0.001Neoadjuvant chemotherapy028Adjuvant chemotherapy720Palliative chemotherapy00Smoking statusNever smoker9330.09Current smoker07Former smoker121Alcohol consumptionNon-drinker10350.04Current drinker017Former drinker09Blood in stoolAbsent380.17Present753Distant metastasis*Yes1200.26No941Tumor recurrence**Yes0140.19No1047DeathYes1280.04No933* Distant Metastasis is defined as the presence of tumor deposits in organs or lymph nodes outside the primary tumor site and its regional lymphatic drainage (e.g., liver, lungs, peritoneum, or non-regional nodes such as para-aortic), indicating systemic disease spread** Tumor recurrence is strictly defined as the reappearance of colorectal cancer at or near the site of the primary tumor following a complete (R0) resection, without evidence of systemic dissemination. This includes recurrence within the tumor bed, anastomosis, adjacent pericolic/perirectal tissues, or regional lymph nodes within the primary drainage basin Table 1Association between tumor sidedness and clinicopathological variables in 71 CCR patientsClinical-pathological FeaturesCategoryRCC (n)LCC (n)p-valueSexMale2320.08Female829Age≤ 62 years5340.74> 62 years527Tumor size0–5 cm6480.336–10 cm411> 11 cm02GradeWell-differentiated1140.50Moderately differentiated942Poorly differentiated05Angiolymphatic invasionYesNo4636250.31Perineural invasionYes8400.48No221Tumor buddingYes2170.71No844Number of lymph nodes retrieved12–2713460.00428–4269> 4316TNM stagingI/IIIII/IV3729320.49CEA (ng/mL)≤ 34350.563–529> 5417ChemotherapyNo chemotherapy313< 0.001Neoadjuvant chemotherapy028Adjuvant chemotherapy720Palliative chemotherapy00Smoking statusNever smoker9330.09Current smoker07Former smoker121Alcohol consumptionNon-drinker10350.04Current drinker017Former drinker09Blood in stoolAbsent380.17Present753Distant metastasis*Yes1200.26No941Tumor recurrence**Yes0140.19No1047DeathYes1280.04No933* Distant Metastasis is defined as the presence of tumor deposits in organs or lymph nodes outside the primary tumor site and its regional lymphatic drainage (e.g., liver, lungs, peritoneum, or non-regional nodes such as para-aortic), indicating systemic disease spread** Tumor recurrence is strictly defined as the reappearance of colorectal cancer at or near the site of the primary tumor following a complete (R0) resection, without evidence of systemic dissemination. This includes recurrence within the tumor bed, anastomosis, adjacent pericolic/perirectal tissues, or regional lymph nodes within the primary drainage basin Association between tumor sidedness and clinicopathological variables in 71 CCR patients Yes No 4 6 36 25 I/II III/IV 3 7 29 32 * Distant Metastasis is defined as the presence of tumor deposits in organs or lymph nodes outside the primary tumor site and its regional lymphatic drainage (e.g., liver, lungs, peritoneum, or non-regional nodes such as para-aortic), indicating systemic disease spread ** Tumor recurrence is strictly defined as the reappearance of colorectal cancer at or near the site of the primary tumor following a complete (R0) resection, without evidence of systemic dissemination. This includes recurrence within the tumor bed, anastomosis, adjacent pericolic/perirectal tissues, or regional lymph nodes within the primary drainage basin Predictive modeling reveals laterality and prognostic clinicopathologic signatures To further delineate clinicopathological features associated with tumor laterality and prognosis, ML techniques were initially developed to integrate clinically and pathologically relevant variables and to estimate individualized outcome probabilities. Models focused on tumor laterality achieved the highest performance across both training and validation analyses. Particularly, alcohol exposure emerged as a dominant predictor of LCC (Fig. 2A and B). The performance metrics for all generated decision-tree models are summarized in Table 2. As observed, the models demonstrated stability during internal validation, with LOOCV accuracies (70.37% and 89.09%, respectively) consistent with those from the full training dataset, although the model designed to predict tumor laterality in the general cohort (Fig. 2A) demonstrated the highest robustness. Prognostic modeling supports the central role of tumor sidedness in stratifying adverse clinical outcomes. In RCC, recurrence risk was lower, whereas in LCC, subsequent risk stratification was primarily driven by tumor size. Tumors larger than 5 cm were associated with an increased risk of recurrence, while tumors ≤ 5 cm were further classified according to perineural invasion status. Notably, and counterintuitively, the absence of perineural invasion was more strongly associated with recurrence than its presence. (Fig. 2C). This algorithm correctly classified 88.9% (24/27) of cases during training and retained a high predictive accuracy of 62.9% (17/27) under cross-validation (Table 2). In the predictive model for mortality stratified by CCR laterality, RCC was associated with more favorable outcomes. LCC was subsequently stratified by angiolymphatic invasion status. In the absence of angiolymphatic invasion, radiotherapy exposure emerged as the next discriminative factor, with improved survival observed among patients who did not receive radiotherapy. Among those who underwent radiotherapy, the pathological stage further stratified risk, with stages I and II associated with superior survival. In the presence of angiolymphatic invasion, the algorithm identified sex as the principal stratifier, with poorer outcomes in men, while women were further stratified by age, with younger patients exhibiting reduced survival probability. (Fig. 2D). As shown in Table 2, the models were stable during internal validation, with LOOCV accuracies (72.73%) consistent with those from the full training dataset (87.27%). To mitigate bias arising from class imbalance, a Cost-Sensitive Learning strategy was employed during training. Misclassification penalty weights were assigned to the minority class for each model (3.0 for Figs. 2A and C and 6.6 for Fig. 2B; and 1.4 for Fig. 2D), while the majority class retained a weight of 1.0. Consequently, the values displayed in the leaf nodes represent re-weighted instances normalized to the total sample size, rather than raw counts. The clinical translation of these values into integer patient counts is detailed in each terminal node of the figures. Fig. 2Decision-tree models for clinicopathological risk stratification in CCR. All models were generated using the J48 algorithm. A cost-sensitive learning strategy was employed to mitigate class imbalances; consequently, values in parentheses within nodes denote weighted sums (total weight/error weight) rather than raw instance counts. The clinical translation of these values into patient counts is detailed in each terminal node of the figures. (A) Tumor laterality prediction in adjuvant therapy patients (n = 27). A misclassification penalty weight of 3 was assigned to the minority class (RCC). The history of alcohol consumption and radiotherapy served as primary discriminators. (B) Tumor laterality prediction in the comprehensive cohort (n = 55). Utilizing similar weighting parameters, the model identifies radiotherapy as a strong predictor of left-sided/rectal involvement. (C) Recurrence prediction in patients undergoing adjuvant chemotherapy (n = 27). The minority event (Recurrence) was assigned an absolute weight of 3.0. Risk is stratified by tumor laterality, size, and perineural invasion. (D) Mortality prediction in patients undergoing chemotherapy (n = 55). A weight ratio of 1.4 balanced death vs. survival events. Prognosis relied on laterality, staging, sex, and age. Decision-tree models for clinicopathological risk stratification in CCR. All models were generated using the J48 algorithm. A cost-sensitive learning strategy was employed to mitigate class imbalances; consequently, values in parentheses within nodes denote weighted sums (total weight/error weight) rather than raw instance counts. The clinical translation of these values into patient counts is detailed in each terminal node of the figures. (A) Tumor laterality prediction in adjuvant therapy patients (n = 27). A misclassification penalty weight of 3 was assigned to the minority class (RCC). The history of alcohol consumption and radiotherapy served as primary discriminators. (B) Tumor laterality prediction in the comprehensive cohort (n = 55). Utilizing similar weighting parameters, the model identifies radiotherapy as a strong predictor of left-sided/rectal involvement. (C) Recurrence prediction in patients undergoing adjuvant chemotherapy (n = 27). The minority event (Recurrence) was assigned an absolute weight of 3.0. Risk is stratified by tumor laterality, size, and perineural invasion. (D) Mortality prediction in patients undergoing chemotherapy (n = 55). A weight ratio of 1.4 balanced death vs. survival events. Prognosis relied on laterality, staging, sex, and age. Predictive performance of decision-tree models The performance metrics for all induced decision-tree models are summarized in Table 2. The laterality tumor prediction models in the general cohort (Fig. 2B) demonstrated the highest overall diagnostic accuracy (89.09%), with a robust MCC (0.627) and a high AUC-ROC (0.83). Notably, this model maintained balanced class sensitivities (0.896 for left-sided tumors and 0.857 for right-sided tumors) and an MCC of 0.421, indicating a moderate-to-strong correlation between predictions and observations. Conversely, models trained on smaller subsets, specifically for recurrence prediction in patients undergoing adjuvant therapy (Fig. 2C), exhibited signs of overfitting, as indicated by a significant discrepancy between training (88.9%) and validation (62.96%) accuracies and an AUC-ROC. The performance metrics for this model indicate suboptimal predictive power, with a negative MCC (-0.227) and an AUC-ROC below the discrimination threshold (0.464). These results suggest significant discriminatory challenges, particularly in classifying tumor recurrence, where the model failed to distinguish between categories effectively. The mortality prediction model (Fig. 2D), however, showed promising generalizability, with a validation accuracy of 72.73% and an F1-score of 0.754 for death, along with a high AUC-ROC, suggesting that the selected clinicopathological variables have significant prognostic value for survival outcomes. Table 2Performance metrics and discriminatory power of decision-tree models for predicting clinicopathological characteristics in CCR patientsDecision-tree modelFull trainingLOOCVClassSpecificitySensibilityF1-scoreMCCAUC-ROCLaterality of tumors in adjuvant therapy patients81.48% (22/27)70.37% (19/27)LCC0.750.570.500.2990.732RCC0.570.750.790.2990.732Laterality of tumor in the general cohort90.9% (50/55)89.09 (49/55)LCC0.8960.8570.6670.6270.832RCC0.8570.8960.9350.6270.832Recurrence of the tumor88.9% (24/27)62.963%(17/27)Yes00.7730.773-0.2270.464No0.77300-0.2270.464Post-chemotherapy mortality87.27% (48/55)72.73% (40/55)Yes0.7390.7190.7540.4520.808No0.7190.7390.6940.4520.787 Table 2Performance metrics and discriminatory power of decision-tree models for predicting clinicopathological characteristics in CCR patientsDecision-tree modelFull trainingLOOCVClassSpecificitySensibilityF1-scoreMCCAUC-ROCLaterality of tumors in adjuvant therapy patients81.48% (22/27)70.37% (19/27)LCC0.750.570.500.2990.732RCC0.570.750.790.2990.732Laterality of tumor in the general cohort90.9% (50/55)89.09 (49/55)LCC0.8960.8570.6670.6270.832RCC0.8570.8960.9350.6270.832Recurrence of the tumor88.9% (24/27)62.963%(17/27)Yes00.7730.773-0.2270.464No0.77300-0.2270.464Post-chemotherapy mortality87.27% (48/55)72.73% (40/55)Yes0.7390.7190.7540.4520.808No0.7190.7390.6940.4520.787 Performance metrics and discriminatory power of decision-tree models for predicting clinicopathological characteristics in CCR patients 62.963% (17/27) Discussion Colorectal tumors are responsible for approximately 900,000 deaths per year, according to data from the World Health Organization [6]. However, if we delve deeper into this pathology, we will realize that colorectal tumors are not a single entity but behave differently depending on their location. Clinicopathological characteristics differ substantially between tumor locations [24]. RCC is more frequently observed in older patients, shows a female predominance, and is often associated with hereditary predisposition. It typically arises from serrated precursor lesions, exhibits increased T-cell infiltration, and demonstrates a higher propensity for peritoneal metastasis [25]. In contrast, LCC more commonly affects younger male patients, is predominantly sporadic, arises from tubular or villous adenomatous polyps, is less immunogenic, and metastasizes more frequently to the liver or lungs [11]. Our study highlighted and reinforced the heterogeneity of this disease through a clinical-pathological approach, combined with innovative ML using decision-tree models, not only to describe it but also to develop tools to predict CRC behavior and prognosis based on laterality. Traditional multivariable regression was not performed due to sample size constraints and low event rates, which would violate the Events Per Variable (EPV) assumption. Therefore, decision-tree models was preferred to capture clinical interactions. This strategy extends beyond descriptive analysis to develop predictive tools that anticipate tumor behavior and prognosis based on the primary tumor laterality. Our model stratified CCR into two distinct groups using predictive decision-tree models. LCC was predominantly associated with behavioral factors, particularly regular alcohol consumption, unlike RCC. These findings could be consistent with divergent pathogenic pathways. Another line of evidence supporting the biological relevance of tumor location derives from studies exploring the tumor microenvironment. Differences in angiogenesis and hypoxia according to tumor sidedness have been described in other malignancies. For instance, a recent investigation evaluating lung neuroendocrine tumors demonstrated that tumors arising on different sides of the lung parenchyma exhibited distinct microenvironmental profiles, with right-sided tumors showing higher microvessel density, whereas left-sided tumors were more frequently associated with hypoxia and reduced vascularization. These findings suggest that anatomical tumor location may reflect underlying differences in vascular architecture and oxygenation within the tumor microenvironment, potentially influencing tumor behavior and clinical outcomes. Although these observations were reported in pulmonary neuroendocrine tumors, they reinforce the broader concept that tumor sidedness may represent a surrogate marker of biological heterogeneity, providing additional context for the prognostic differences observed between right- and left-sided CCR [26]. RCC is frequently associated with BRAF mutations and high-frequency microsatellite instability resulting from defective DNA mismatch repair, leading to widespread genomic alterations [27]. In contrast, left-sided tumors typically arise through the conventional adenoma–carcinoma sequence, driven by chromosomal instability with gains, amplifications, losses, or deletions affecting a more limited set of genes, including APC, KRAS, TP53, SMAD4, and PIK3CA [27, 28]. These findings can alter not only the treatment but also the prognosis of CRC. The integration of molecular biomarkers into predictive models could refine prognostic stratification and enhance the biological interpretability of ML approaches applied to CCR. Unfortunately, comprehensive molecular profiling was not consistently available for all patients in our retrospective cohort, which precluded the inclusion of these variables in the current analysis. Nevertheless, future studies incorporating genomic and molecular data may improve the predictive accuracy of decision-tree models and provide deeper insight into the biological mechanisms underlying tumor sidedness and clinical outcomes in colorectal cancer. In our cohort, despite patients with LCC receiving more intensive treatment modalities, including chemoradiotherapy (p < 0.01), this group demonstrated poorer outcomes, with a significantly higher mortality rate (p = 0.04). This contrasts with the available literature, which consistently indicates a poorer prognosis for RCC. Petrelli et al. published a meta-analysis demonstrating an absolute 19% reduction in the risk of death for LCC compared with RCC tumors, based solely on sidedness [29], and these results were corroborated by a large retrospective Korean cohort of 1,632 patients, which reported an increased risk of locoregional recurrence in RCC [30]. Meanwhile, contrasting findings have been reported in patients with early-stage tumors. Papers focusing on the initial stages of CRC (I and II) have reported improved survival among patients with RCC, potentially related to the higher prevalence of microsatellite instability–high, a feature associated with a more favorable prognosis in early-stage disease [31]. In addition, a retrospective analysis of the National Cancer Database showed that among patients who did not receive adjuvant chemotherapy, those with RCC had superior overall survival compared with those with LCC [32]. Taken together, these observations highlight the need for integrative, data-driven prognostic approaches that reconcile the stage-dependent, biologically distinct survival patterns associated with tumor laterality, thereby providing a rationale for applying decision-tree–based modeling in colorectal cancer. Current evidence indicates that decision-tree models are reliable statistical predictive models that accurately estimate the risk of clinical events by integrating key prognostic factors in oncology [33]. In many malignancies, including CRC, decision-tree models have demonstrated superior prognostic performance compared with traditional staging systems [34]. These observations prompted us to explore the feasibility of developing a decision-tree model to predict the prognostic impact of tumor laterality in CRC. The conventional TNM staging system has well-recognized limitations. Prognostic assessment based exclusively on tumor size, lymph node involvement, and distant metastases fails to account for several clinically relevant variables, including tumor laterality, surgical margin status, alcohol and tobacco exposure, sex, age, and the use of chemoradiotherapy [35]. Accordingly, in the present study, we integrated ML techniques with a comprehensive set of clinical and pathological variables to refine survival estimation and to specifically assess the prognostic relevance of tumor laterality. The analyses revealed that laterality was modulated by behavioral factors, particularly alcohol consumption, which was predominantly associated with LCC in both models (Fig. 2A and B). Importantly, the behavioral variables included in our models, particularly alcohol consumption and tobacco exposure, should be interpreted cautiously. Given the retrospective design of the present study, these variables cannot be considered causal determinants of tumor laterality or prognosis; rather, they represent contextual exposures observed within this specific cohort. Their inclusion aimed to explore whether lifestyle-related factors could contribute to the clinical heterogeneity observed between RCC and LCC. In this context, alcohol consumption functioned as a predictive stratification variable within the decision-tree models, reflecting the clinical profile of patients with left-sided tumors in our population rather than implying a direct mechanistic relationship. This interpretation is consistent with epidemiological observations suggesting that LCCare more frequently associated with modifiable behavioral exposures, whereas right-sided tumors are more often linked to aging, hereditary predispositions, and distinct molecular pathways. Therefore, our findings should be regarded as exploratory and hypothesis-generating, highlighting potential cohort-specific patterns that warrant external validation in larger, multicenter studies incorporating diverse populations. The statistical strength of these associations is corroborated by our predictive modeling, which demonstrates an LOOCV accuracy of 89.09% and an AUC-ROC of 0.83 (Table 2). Despite this strong predictive performance, these findings should be interpreted as context-dependent associations rather than evidence of causality, as lifestyle exposures may vary across populations and may interact with other clinical determinants of prognosis. In contrast, RCC was strongly associated with more favorable outcomes and lower recurrence rates, whereas LCC disease was further stratified by tumor size and perineural invasion (Fig. 2C). However, these specific findings should be interpreted with caution. While the model identified perineural invasion as a decision node, the quantitative metrics indicate signs of overfitting, with accuracy dropping from 88.9% in training to 62.96% in validation. Furthermore, the low AUC-ROC (0.46) suggests limited discriminative power for this specific outcome, indicating that this counterintuitive association may reflect sample size constraints or statistical noise rather than a definitive biological phenomenon. Despite applying rigorous 5-fold cross-validation and pruning techniques, the model yielded a poor AUC of 0.477 (data not shown). This emphasizes that standard clinical variables are insufficient to predict recurrence in small cohorts. However, from a hypothesis-generating perspective, if this finding reflects true biological behavior, it may suggest that tumors that recur without perineural invasion rely on alternative, highly aggressive mechanisms of early dissemination, such as primary hematogenous micrometastasis, thereby circumventing the need for neural tracking. This contrasts with the observed epidemiological evidence. In SEER-based cohorts comprising more than 163,000 patients, LCC has been consistently associated with lower cancer-specific mortality compared with RCC; however, this association is stage-dependent, demonstrating comparable or even more favorable outcomes in early-stage disease and no clear survival advantage in stage III tumors. Given the poorer outcomes observed in LCC, the model enabled a detailed exploration of risk factors and patient subgroups, stratifying them into hierarchical levels according to their relative risk and impact on survival. In patients without angiolymphatic invasion, radiotherapy emerged as the subsequent decision node, whereby the requirement for radiotherapy and more advanced disease stages were associated with an unfavorable prognosis. In the presence of angiolymphatic invasion, prognostic stratification was primarily determined by sex and age, with male patients exhibiting poorer outcomes. Notably, within this branch, younger women (< 62 years) were paradoxically associated with worse survival. These counterintuitive associations highlight the need for further investigation to elucidate the biological and clinical mechanisms underlying these observations. ​​​​In contrast to the recurrence analysis, the mortality risk stratification demonstrated high reliability. The decision-tree integrating angiolymphatic invasion, sex, and age achieved a validation accuracy of 72.73% and an AUC-ROC exceeding 0.80 (Table 2). These metrics validate the clinical relevance of the identified subgroups, supporting the model’s potential utility in identifying patients at high risk of mortality despite the observed complex interactions. Although tumor sidedness has emerged as a relevant prognostic factor in CRC, the available evidence remains heterogeneous and strongly dependent on disease stage and clinical context. Large meta-analyses and pooled analyses of advanced and metastatic disease consistently demonstrate superior outcomes for left-sided tumors, with improved overall survival and treatment response when compared with right-sided cancers [29, 35–37]. In contrast, population-based studies focusing on early-stage disease, particularly stage II, have reported outcomes for RCC after adequate surgical resection equal to or better than those for LCC [31, 38–40] findings suggest that tumor laterality does not serve as a uniform prognostic marker across all stages but rather reflects distinct biological behaviors that may confer a survival advantage in early disease while portending worse outcomes in advanced disease. Such stage-dependent and, at times, seemingly counterintuitive results underscore the complexity of CRC biology and highlight the need for further prospective studies that integrate molecular, clinicopathological, and treatment-related variables to better define the true prognostic role of tumor sidedness. In addition, the biological interpretation of tumor sidedness is closely linked to the molecular heterogeneity of colorectal cancer, as right- and left-sided tumors are known to harbor distinct genomic and epigenetic alterations, including differences in MSI/MMR status, CIMP phenotype, and driver mutations such as BRAF and KRAS [28]. However, comprehensive molecular profiling is not routinely available in many public healthcare settings in developing countries, including the Brazilian public health system, where the present cohort was treated. The limited availability of genomic testing in such contexts poses a significant barrier to incorporating molecular stratification into routine clinical decision-making. In this scenario, studies such as ours that mine clinicopathological data to identify prognostic patterns may provide valuable, accessible tools for risk stratification, particularly in resource-constrained healthcare systems. To address the class imbalance between left-sided (n = 61) and right-sided (n = 10) tumors, which could bias the decision-tree towards the majority class, we employed a balancing strategy during the modeling phase. Specifically, we used Cost-Sensitive Learning to improve the representation of the minority class in the training set. This approach aims to improve the J48 algorithm’s decision boundaries, ensuring that the predictive rules for right-sided tumors reflect robust patterns rather than statistical noise. The effectiveness of this weighting strategy is evidenced by the balanced sensitivity rates achieved in the laterality model (0.896 for LCC and 0.857 for RCC). This confirms that the algorithmic adjustments successfully mitigated bias toward the majority class, enabling reliable identification of right-sided tumor characteristics despite the numerical disparity. The interpretation of these findings warrants caution, given certain inherent limitations, specifically a retrospective, single-center design and the modest sample size, which resulted in an imbalance in tumor laterality. Specifically, the small number of RCC cases mathematically precludes granular stage-stratified survival and subgroup analyses, limiting our ability to definitively confirm whether the observed prognostic advantages are strictly stage-specific, as suggested by broader literature. Furthermore, the small sample size precluded direct quantitative performance benchmarking, such as comparative AUC-ROC, C-index, or calibration curves, between our AI models and traditional staging alone. Additionally, the lack of comprehensive molecular and genomic data, such as MSI/MMR status and BRAF/KRAS mutation profiles, precludes a deeper exploration of the biological mechanisms underlying the observed prognostic differences. While we employed Cost-Sensitive Learning and LOOCV to mitigate majority-class bias and rigorously test model stability, these algorithmic techniques cannot entirely eliminate the risks of statistical bias and overfitting inherent to small cohorts, as explicitly observed in our recurrence model. Consequently, these limitations preclude definitive claims regarding the absolute prognostic superiority of RCC, reinforcing that our models should be interpreted strictly as an exploratory, hypothesis-generating framework. This distribution likely reflects local referral patterns rather than a biological deviation and may have influenced survival estimates, particularly the higher mortality observed in LCC, despite the broader literature often associating right-sided disease with poorer outcomes. Importantly, these constraints do not undermine the study’s core contributions. The analysis provides a detailed characterization of CRC in a well-defined clinical setting, confirms the prognostic relevance of established markers of tumor aggressiveness, and demonstrates the capacity of ML-based decision-tree models to capture coherent, clinically meaningful risk stratification even in relatively small cohorts. Besides, a major strength of this study lies in the deliberate selection of the J48 (C4.5) decision-tree algorithm. We prioritized decision-tree models over more opaque algorithms to ensure transparency in risk stratification. The resulting logical rules (IF-THEN structures) mirror standard clinical reasoning, enabling oncologists to verify the biological coherence of the predictors, such as angiolymphatic and perineural invasion, tumor size, and staging identified by the AI. This transparency is essential for the clinical translational potential of our findings, which is further reinforced by the model’s interpretability. Together, these findings underscore their value as hypothesis-generating evidence for future larger and external multicenter validation and highlighting the need to integrate genomic data to avoid the development of spurious algorithmic rules. Such models should be interpreted as complementary tools designed to enhance clinical decision-making rather than replace established staging systems or validated molecular assays. Conclusion This study demonstrates that tumor sidedness represents a clinically meaningful and biologically grounded prognostic dimension in CCR when interpreted alongside established pathological and treatment-related variables. By integrating clinicopathological data into transparent, decision-tree–based ML models, we identified coherent, clinically interpretable risk signatures that extend beyond conventional TNM staging. Right-sided tumors were consistently associated with more favorable survival and lower recurrence in this cohort, whereas left-sided disease exhibited distinct behavioral, therapeutic, and pathological stratifiers linked to adverse outcomes, including alcohol exposure, tumor size, perineural invasion, and angiolymphatic involvement. Although limited by its retrospective, single-center design and modest sample size, this work supports the utility of explainable AI as a hypothesis-generating framework for individualized prognostic assessment. Future multicenter, prospective studies incorporating molecular and genomic data are warranted to validate and refine these models and to more precisely define the stage-dependent prognostic role of tumor laterality in CCR.