The effectiveness of faecal immunochemical test (FIT)–based colorectal cancer (CRC) screening programs relies heavily on the timely completion of colonoscopy following a positive FIT result, enabling the detection and removal of early-stage cancers and precancerous lesions. However, due to limited endoscopic capacity and organizational constraints, many programs struggle to meet the recommended target of performing colonoscopy within 31 days of a positive FIT. This study aimed to evaluate the potential of a Conditional Inference Decision Tree (CIDT) model as a supportive tool to classify the risk of CRC among FIT-positive participants in the Emilia-Romagna regional screening program.

Study data were obtained from the Information System for the Surveillance of Colorectal Screening of the Regional Department of Health. The analysis included all individuals aged 50–69 years with a positive FIT result recorded between 2005 and 2024. Candidate predictors included age, sex, haemoglobin concentration at FIT, cumulative haemoglobin concentration for repeat participants, and screening round (first vs subsequent).

A Conditional Inference Decision Tree was applied to identify the most relevant predictors of CRC. Internal validation was performed using cross-validation, and model performance was assessed through the area under the ROC curve (AUC). Weighted analyses were applied to account for the low prevalence of CRC (4%).

Among 201,065 FIT-positive participants, 4% were diagnosed with colorectal cancer at colonoscopy. Although CRC was a relatively rare outcome, the Conditional Inference Decision Tree (CIDT) identified clear patterns of risk within the screened population. Haemoglobin concentration at FIT was the dominant predictor and constituted the first splitting variable in the tree, effectively distinguishing participants with very low probabilities of CRC from those at substantially higher risk. Increasing haemoglobin values showed a strong and progressive association with the likelihood of detecting colorectal cancer at colonoscopy.

Age and screening round provided additional discriminatory information, refining risk estimates within haemoglobin-based strata. Individuals undergoing their first screening round and those in older age groups had higher probabilities of CRC, whereas cumulative haemoglobin across previous FIT rounds contributed to differentiating risk among repeat participants. Sex did not substantially influence the structure of the tree.

Participants with lower haemoglobin values showed markedly lower risk across all branches of the tree, with final estimated probabilities of CRC ranging from 1.4% to 4.6%, depending on age and screening round. Within this group, the lowest observed risks were found among participants with low haemoglobin values who were younger (≤59 years) and in subsequent screening rounds, with estimated probabilities as low as 1.4%. In contrast, participants with higher FIT haemoglobin concentrations exhibited substantially greater probabilities of CRC. Within this group, age and screening round further stratified risk: older individuals (>57 years) undergoing their first screening round formed the highest-risk terminal nodes, with estimated CRC probabilities reaching 19.9%. The CIDT had an AUC of 0.71. 

The moderate discriminative performance of the tool, which is consistent with the low-prevalence nature of the endpoint and the real-world heterogeneity of FIT-positive populations, suggests that it needs to be improved with additional patient information. Our findings indicate that the CIDT has a potential to become a simple, interpretable, and data-driven approach to assist clinicians in the management of  colonoscopy workload by identifying higher-risk subgroups of FIT-positive patients, particularly in settings with constrained resources or waiting list pressures.