Abstract
<title>Abstract</title> <p>Background: Early outbreak line lists are high-value but difficult machine-learning data. Their schemas, missingness patterns, case definitions, and confirmation processes change across pathogens and jurisdictions. A model pretrained on one outbreak can therefore transfer harmful shortcuts rather than reusable epidemiological structure. Methods: We developed OBFM-SafeEnsemble, a guarded cross-outbreak learning framework. A historical 2022 mpox line list (69,595 records) supplied source self-supervision and source coefficient priors; a live 2026 Bundibugyo virus disease Ebola line list (2,900 records) was the primary target; and a 24-record hantavirus line list was used only as a pipeline stress test. The endpoint was confirmation status. Direct laboratory, confirmation, outcome, post-confirmation care, provenance, and identifier fields were excluded to reduce leakage. OBFM-Lite first evaluated masked-token pretraining and continual unlabeled adaptation. OBFM-SafePrior then regularized a new Ebola head toward source-learned token effects and activated transfer only when validation AUROC exceeded a prespecified gain threshold. OBFM-SafeEnsemble trained a small library of disease-only and source-prior heads and selected either the best head, a top-N average, or a convex blend, always retaining a disease-only fallback. Results: Source masked-token pretraining improved in-domain mpox Transformer AUROC from 0.839 to 0.851. Continual unlabeled adaptation improved Ebola zero-shot AUROC from 0.487 to 0.523, but absolute transfer remained weak. At 64 labeled Ebola records, SafePrior improved AUROC from 0.524 to 0.664 and AUPRC from 0.660 to 0.731. SafeEnsemble achieved AUROC values of 0.662, 0.656, 0.687, 0.682, and 0.667 at 32, 48, 64, 96, and 128 labels, respectively. Relative to matched disease-only models, AUROC gains were +0.068, +0.082, +0.235, +0.064, and +0.015; AUPRC gains were +0.042, +0.067, +0.160, +0.074, and +0.033. Interpretation: Positive cross-outbreak transfer was achievable under severe label scarcity only after source knowledge was treated as optional, target-validated regularization and averaged across low-variance candidate heads.</p>