Background and Objective: Compared to the housed population, people experiencing homelessness (PEH) face disproportionately high cancer risks due to social exclusion, unstable housing, and limited healthcare access. The CANCERLESS project aimed to enhance cancer prevention for PEH through health navigation and empowerment models. However, progress is hindered by data scarcity, population heterogeneity, and complex socio-health dynamics. This study proposes a microsimulation model using machine learning to predict the effects of potential interventions on quality of life, healthcare utilisation, and empowerment. This approach aims to optimise resource allocation and identify at-risk subgroups for more effective cancer prevention by supporting evidence-based decision-making. Materials: We used data from the CANCERLESS project (June 2022–November 2023) involving 652 PEH from Austria, Greece, Spain, and the United Kingdom. Of these, 255 participants completed an 18‐month Health Navigator intervention. They all met the European Typology of Homelessness and Housing Exclusion criteria, with questionnaires administered at baseline, four weeks, and post-intervention. Methods: A microsimulation model was developed using modular machine learning (ML) algorithms for flexible adaptation to specific contexts. The model proceeds in three steps: (1) generating synthetic cases under defined constraints; (2) quantifying the change in outcomes at the intervention's end by comparing with similar participants who completed the intervention; and (3) aggregating results to assess the intervention's impact. Various ML methods were evaluated to maximise the synthetic population's similarity to the original and accurately predict the intervention's effect. Results: Our findings indicate that the Conditional Tabular GAN approach best fits the original data distribution with a propensity score of 0.152. Models yielded a mean absolute error of approximately 10 for Health Rating (scale 1–100) and 0.1 for EQ5D5L (scale -0.59–1), with ordinal outcomes exhibiting errors below one category, except for Social worker visits, which showed higher variability. A demonstrator (https://epione.upv.es) has been developed to present the microsimulation process in a user-friendly, simple, and accessible manner. Conclusion: A novel microsimulation model based on ML has been developed for the PEH population, capable of estimating intervention impacts and generating realistic synthetic individuals. Its flexible design supports evidence-based decision-making and may help optimise resource allocation, enhancing intervention outcomes by predicting health outcomes, use of health services, and empowerment, mainly before the intervention was implemented.
Epione microsimulation APP 3511da7ef6475195ff9f4e8ba62be12d3f5bfbc3