PubMed-Screener-420M

PubMed-Screener-420M is an open-weights, locally deployable biomedical decision model fine-tuned for high-recall title and abstract citation screening in systematic reviews.

• Model Name: PubMed-Screener-420M (internal checkpoint lineage: soces_pico_laya_e2e)
• Base Architecture: answerdotai/ModernBERT-large (395M encoder parameters) with a non-autoregressive System-1 binary proposition decision head adapted from convaiinnovations/laya (trained via RLCD against strictly proper scoring rules by ConvAI Innovations; 421M parameters total).
• Primary Domain: Biomedical systematic literature review screening
• Operating Threshold: $\theta^* = 0.028$ (established a priori on 4 calibration reviews)
• Deployment Profile: Fully offline inference on consumer hardware (e.g., Apple Silicon MPS) or server GPUs (NVIDIA CUDA) without external data transmission.

---

Intended Use & Clinical Decision Context

In systematic literature reviews, investigators must screen thousands of PubMed citations retrieved by broad Boolean search strings to identify the small fraction of studies meeting strict clinical eligibility criteria (PICO: Population, Intervention, Comparator, Outcome).

PubMed-Screener-420M serves as a pre-screening decision model / safety filter:

1. Evaluates candidate title + abstract pairs against a standardized review objective prompt.
2. Applies a frozen operational decision threshold ($\theta^* = 0.028$) to guarantee high recall ($\ge 95%$) while eliminating clearly non-eligible literature.
3. Operates entirely on local hardware, ensuring research protocol confidentiality and data sovereignty without third-party API dependencies.

---

Validation Performance (External Validation on 62 Systematic Reviews)

The model was evaluated across 62 unseen systematic reviews spanning diverse clinical domains (962,860 candidate evaluations, 2,431 ground-truth inclusions):

Evaluation Design	Prevalence	Pooled Recall	Specificity	Mean Workload Retained	Work Saved over Sampling (WSS@95)
Boolean-Search Deployment	16.4%	95.15% [92.69–97.12%]	28.20% [22.90–33.32%]	77.6% [72.3–82.8%]	10.4% [0.0–21.2%]
Cross-Topic Ranking	0.25%	95.15% [92.69–97.12%]	94.86% [92.78–96.40%]	5.36% [3.82–6.91%]	89.09% [85.22–92.46%]

Note: All 95% Confidence Intervals are computed via 5,000 systematic-review clustered bootstrap resamples.

Hardware Throughput

• NVIDIA T4 GPU (CUDA, batch size 256): 27.8 documents/s (~9.3 min per 15,530-document catalogue)
• Apple M3 MacBook Air (MPS, batch size 64): 2.4 documents/s (~106.9 min per 15,530-document catalogue, 2.1 GB peak RAM)

---

Input Prompt Contract

The model expects an input prompt pairing the systematic review objective with the candidate study's title and abstract:

[REVIEW OBJECTIVE]
Systematic review evaluating the efficacy and safety of direct oral anticoagulants versus low-molecular-weight heparin for the treatment of cancer-associated venous thromboembolism.

[TITLE]
Apixaban for the Treatment of Venous Thromboembolism in Patients with Cancer.

[ABSTRACT]
BACKGROUND: Direct oral anticoagulants are an alternative to low-molecular-weight heparin for the treatment of venous thromboembolism in patients with cancer. METHODS: In a randomized trial, we evaluated apixaban versus dalteparin...

---

Attribution & Acknowledgements

PubMed-Screener-420M was developed on the open-source foundations of:

• Laya (convaiinnovations/laya): Multilingual, non-autoregressive System-1 decision engine evaluating typed propositions in single forward passes, developed with reinforcement learning against strictly proper scoring rules (RLCD) by ConvAI Innovations.
• ModernBERT (answerdotai/ModernBERT-large): Modern contextual transformer encoder developed by Answer.ai and LightOn.

---

Citation & Manuscript

If you use PubMed-Screener-420M in your evidence synthesis workflow or research, please cite:

@article{pubmed_screener_420m_2026,
  title={Offline AI Decision Models for PubMed Systematic Review Screening: External Validation of a Calibrated Open-Weights Architecture Across 62 Reviews},
  author={Wisu et al.},
  journal={Journal of Biomedical Informatics},
  year={2026},
  note={Under Review}
}