Abstract
Computational reproducibility remains a persistent weakness in infectious disease artificial intelligence (AI), where underspecified methods, unavailable code, and incomplete documentation impede independent verification and reuse of published models. This study investigated whether generative AI (GenAI) can measurably enhance the reproducibility of mpox outbreak models within a CRISP-ML(Q)–guided. A retrospective reproducibility assessment of mpox modelling studies published between 2019 and 2025 was conducted. Five representative studies were curated from peer-reviewed literature and public repositories, encompassing symptom-based tabular machine learning models and dermatological image-based deep learning models. The selected studies employed diverse modelling paradigms, including convolutional neural networks, vision transformers, and gradient-boosted decision trees. Each model was first replicated using only the originally available data, code, and documentation to establish baseline reproducibility, and was subsequently re-executed using GenAI-assisted interventions.
GenAI-assisted interventions included synthetic data generation to mitigate data scarcity, large language model–assisted code reconstruction to address missing or ambiguous implementation details, and automated documentation to standardise reporting. Prior to GenAI assistance, only one of the five studies was fully reproducible, with substantial variability observed across repeated executions. Following GenAI-assisted replication, four of the five studies achieved full computational reproducibility, reducing discrepancies between reported and reproduced performance metrics and improving execution stability. Automated documentation consistently achieved high documentation completeness, while containerised environments and deterministic seeding enabled auditable, repeatable execution. GenAI proved most effective in resolving methodological ambiguity and standardising execution in studies with partially available documentation and data. These findings indicate that GenAI can substantially strengthen reproducibility and transparency in outbreak modelling when used to augment open science practices and standardised workflows, highlighting priorities for future validation across additional diseases and modelling contexts.