Title:

BLENS: Biomedical Literature Extraction & Scoring System

Systematic reviews require comprehensive retrieval across multiple databases, yet no publicly available tool exists for automated cross-database literature quality scoring. BLENS retrieves records from PubMed, ClinicalTrials.gov, medRxiv, and bioRxiv, scoring all output via a ML pipeline validated on Third Molar Extraction (TME) literature and keywords. Text is encoded into dense vectors using Sentence-BERT and scored by cosine similarity to a domain-specific query. A semi-supervised anchoring layer uses 5 expert-validated gold-standard studies as labeled seeds to recalibrate scores across all records. A rubric then scores against 5 weighted inclusion criteria gathered from a published TME meta-analysis, combining these scores into three bins tiered High, Medium, or Low. Applied to 1,795 records, Blens achieves 100% recall on gold-standard studies, with composite scores improving from 0.59-0.72 at baseline to 0.80-0.99 after anchoring. HDBSCAN clustering independently confirmed results, separating preprints into off-topic clusters due to a vague retrieval from a weak API. Anomaly rates of 25% (bioRxiv) and 15.6% (medRxiv) versus 0.8% (PubMed) and 0% (ClinicalTrials) reflects the poor retrieval of preprints due to API query limits. BLENS is a proof of concept, where precision measurement via manual audit of top-rank records is the next critical step.

Name

Comment*

Comments are viewable only by submitter

Captcha*