There are currently no colorectal cancer (CRC) screening recommendations tailored specifically for people with HIV (PWH). Guidance applied to PWH has generally mirrored that for people without HIV (PWoH), despite observations that CRC prevalence may be higher and that CRC can appear at earlier ages among PWH compared with PWoH. The complexity and potentially high dimensionality of register-derived health data make machine learning (ML) approaches attractive for risk prediction, since ML methods can capture nonlinearities and complex interactions that traditional models may not.
This study protocol describes development of an ensemble ML model intended to predict CRC risk in PWH using comprehensive nationwide datasets from Sweden. The ensemble ML model will be evaluated against a baseline Cox proportional hazards regression model; the superior approach will be used to produce a CRC risk prediction tool aiming to support personalised screening recommendations for PWH.
Primary objectives are to develop an ensemble ML model to predict incident CRC among PWH, to evaluate its predictive performance, and to compare that performance to a Cox regression baseline. If the ensemble or Cox model demonstrates superior predictive validity, that method will be implemented to derive a CRC risk prediction model intended to inform personalised screening for PWH.
The study population comprises all PWH and PWoH born between 1940 and 2008, aged 18 years or older, who lived in Sweden at any time between 1983 and 2024. Data linkage will be performed across six nationwide demographic and healthcare registers. The registers will provide the longitudinal data necessary to ascertain exposures, covariates and outcomes for participants across the study period.
The outcome of interest is incident colorectal cancer (CRC). Participants will be followed from cohort entry until the first occurrence of CRC, emigration from Sweden, or death. Outcomes will be ascertained from the linked national registers. The protocol specifies that follow-up will continue through the available registry period up to 2024.
People with HIV will be matched to negative controls at a ratio of 1:10. The matched PWoH serve as comparators for model development and evaluation. Details on exact matching variables and procedures were not reported in the source beyond the 1:10 matching ratio.
A Cox proportional hazards regression will be performed first and used as the baseline comparator for predictive performance. The Cox model will yield time-to-event estimates for CRC risk and serve as the standard against which ML-derived performance metrics will be compared. Specific covariates to be included in the Cox model and model selection procedures were not reported in the source.
An ensemble ML model will be developed by combining multiple ML methods using stacking. The protocol frames ML as potentially advantageous for handling complex or high-dimensional data that arise in register-based studies. A range of ML algorithms will be trained and then combined in a stacked ensemble to produce a final prediction. The ensemble's predictive performance will be evaluated and compared to the Cox regression baseline. Specific algorithms, tuning strategies, training-validation splits, performance metrics, and thresholds were not detailed in the source document.
The study received ethical approval from the Regional Ethical Committee in Sweden under several Dnr numbers (2024-04185-02, 2024-06783-02, 2023-00191-01, 2022-02897-02, 2022-05624-01, 2018/11-31/2). As a retrospective register-based investigation using pseudonymised data, the study is described as having minimal physical, psychological or privacy risks to included individuals. Results will be reported only at the population level, eliminating the possibility of individual identification. The investigators plan to submit study results for publication in a peer-reviewed journal.
The protocol states that the better-performing predictive method—either the ensemble ML model or the Cox regression model—will be implemented to develop a CRC risk prediction model with the goal of personalising CRC screening recommendations for PWH. The study rationale emphasises that improved risk prediction could address observed differences in CRC prevalence and age at diagnosis between PWH and PWoH, but the source does not report study results, performance metrics, or specific implementation pathways.