AEbert-ICI
Overview
AEbert-ICI is a fine-tuned biomedical language model developed to learn adverse event co-occurrence patterns from large-scale adverse event reporting data in the context of immune checkpoint inhibitor (ICI) therapy. The model is trained using publicly available reports from the U.S. FDA Adverse Event Reporting System (FAERS) and is intended to support methodological research on adverse event representation learning and imputation in immuno-oncology.
Data and model descritpon
The training data were derived from FAERS reports collected between 2015 Q1 and 2025 Q3. After preprocessing and deduplication, the cohort included 72,083 patients treated with ICIs (with 17,329 lung cancer patients and 10,872 patients melanoma or skin cancer patients). Adverse events were represented using MedDRA preferred terms (pts). To focus on treatment-associated toxicities, a curated set of 207 preferred terms was retained after removing events obviously related to disease progression or non-treatment processes.
The model is based on the BioBERT encoder and is trained using a masked adverse event completion objective, in which a subset of observed adverse events is removed and the model learns to recover the missing events from the remaining context. This formulation allows the model to capture higher-order co-occurrence structure among adverse events without relying on patient-level outcomes or time-to-event information. The resulting representations are intended for research applications such as adverse event prioritization, pattern discovery, and benchmarking of computational methods in pharmacovigilance and immunotherapy safety analysis.
Notes:
This model is not intended for clinical decision-making or individual patient risk assessment.
Developer and Contact:
ICON_Lab @ Moffitt (Xuefeng.Wang@moffitt.org)