CERC

CERC extracts and summarizes salient information from clinical and biomedical texts to support precision medicine and clinical decision-making.

Key Features:

• Multi Indicator Text Summarization (MINTS): a sentence-ranking framework that performs extractive summarization by evaluating multiple indicators of importance.
• Random forests ranking/classification: MINTS employs a random forests algorithm to assess relevance and rank sentences.
• Weighted TF-IDF of domain-specific terms: uses weighted term frequency–inverse document frequency scores for over-represented domain-specific terms.
• Controlled vocabulary dictionary: leverages a dictionary constructed from MeSH, SNOMED-CT, and PubTator to identify domain-specific terms.
• Training corpus: trained on 35 full-text articles from the CRAFT corpus.
• Evaluation metrics and baselines: evaluated against 32 CRAFT articles and 30 clinical case reports using the ROUGE toolkit (ROUGE-1, ROUGE-2, ROUGE-SU4) and compared to single-indicator methods such as LexRank and random selection.
• Statistical performance: achieved statistically significant improvements in ROUGE scores (p < 0.01) and a random forests sentence classification accuracy of 87.5% for labeling sentences as "good" or "bad."
• Early medical risk detection: includes features aimed at early detection of medical risks from clinical text.

Scientific Applications:

• Clinical summarization: produce extractive summaries of clinical and biomedical texts to support clinical decision-making.
• Precision medicine support: extract domain-specific evidence to inform precision medicine workflows.
• Early risk identification: identify signals for early detection of medical risks from text-derived information.
• Clinical decision support integration: provide extracted and summarized evidence that can be integrated with clinical decision support systems.
• Biomedical text-mining research: serve as a benchmark and component for research in biomedical information extraction and summarization.

Methodology:

MINTS ranks sentences using multiple indicators evaluated by a random forests algorithm; domain-specific terms are weighted by TF-IDF using a controlled vocabulary from MeSH, SNOMED-CT, and PubTator; the system was trained on 35 CRAFT full-text articles and evaluated on 32 CRAFT articles plus 30 clinical case reports with ROUGE-1, ROUGE-2, and ROUGE-SU4 metrics (p < 0.01), and the random forests classifier achieved 87.5% accuracy for "good"/"bad" sentence classification.