Our TITAN-SR paper is out in J Clin Epidemiol. A biomedical BERT ensemble for systematic-review title/abstract screening — developed and externally validated against active learning and LLM chatbots. doi.org/10.1016/j.jclinepi.2… #SysRev #MedTwitter #EvidenceSynthesis

What TITAN-SR is. Ensemble of PubMedBERT + BioLinkBERT, trained on ~763k human screening decisions from real reviews. One job: include/exclude. Not a rented chatbot with a clever prompt — a specialist classifier. coresr.ai

Why that matters for systematic reviews. Chatbots can flip on the same abstract across runs. TITAN-SR is deterministic: same input → same output every time (agreement 1.0). Reproducibility is the floor, not a bolt-on feature. coresr.ai

Safety margin you can set. It returns calibrated probabilities. Threshold fixed on held-out data to hold recall ≥0.99 — and that operating point transfers to new records. You set the bar once. It carries. doi.org/10.1016/j.jclinepi.2…

Headline from the paper: the BERT ensemble outperformed active learning and LLM chatbots for title/abstract screening on external validation. Screening is classification, not conversation. Build the specialist. coresr.ai #SystematicReview #AI