SALTClass

SALTClass performs enrichment and supervised classification of sparse clinical text using clustering and machine learning to improve extraction of patient information from unstructured clinical notes.

Key Features:

• Clustering-Based Enrichment: Utilizes seven clustering algorithms—latent Dirichlet allocation, K-Means, MiniBatchK-Means, BIRCH, MeanShift, DBScan, and Gaussian Mixture Models (GMM)—to process and enrich sparse, short clinical text.
• Supervised Classification Integration: Incorporates ten different supervised classifiers that can operate on the original document-term matrix or on an enriched representation of the text.
• Background Knowledge Utilization: Leverages unlabeled data as background knowledge to address sparsity in short clinical notes and reduce classification errors.
• Clinical Corpus Application: Applied to a Dutch clinical cardiovascular text corpus from University Medical Center Utrecht to identify patient information such as family history.

Scientific Applications:

• Clinical information extraction: Enhances representation and classification of unstructured short notes in electronic health records for downstream phenotyping and information retrieval.
• Patient attribute identification: Supports identification of attributes such as family history within cardiovascular clinical narratives.

Methodology:

Two-step computational workflow: clustering with latent Dirichlet allocation, K-Means, MiniBatchK-Means, BIRCH, MeanShift, DBScan, and GMM to group and smooth cluster representations for enrichment; followed by supervised classification using ten integrated classifiers on the original document-term matrix or the enriched representations, with background knowledge incorporated from unlabeled data.