Towards similarity-based differential diagnostics for common diseases.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Medical Informatics

Background:

• Ontology-based phenotype profiles aid rare disease diagnosis and decision support.
• Semantic similarity is key for generating diagnostic hypotheses.
• Previous methods were limited to rare diseases and curated data, excluding common diseases and uncurated clinical text.

Purpose of the Study:

• To develop an approach for deriving patient phenotype profiles from clinical narrative text.
• To apply semantic similarity to text-derived phenotypes for primary patient diagnosis classification.
• To compare patient-patient and patient-disease similarity using literature-mined and text-mined phenotypes.

Main Methods:

• Developed an approach to extract patient phenotype profiles from clinical narrative text (MIMIC-III).
• Utilized semantic similarity for classifying primary patient diagnoses based on text-derived phenotypes.
• Compared patient-patient similarity and patient-disease similarity, including a combined approach extending literature phenotypes with text phenotypes from 500 patients.

Main Results:

• Demonstrated a powerful approach for differential diagnosis of common diseases using uncurated text phenotypes.
• Showcased the utility of information from both within and outside the clinical setting.
• Achieved effective classification of primary patient diagnoses.

Conclusions:

• Uncurated text-derived phenotypes can be effectively used for the differential diagnosis of common diseases.
• The developed methods show promise for various clinical tasks, including differential diagnosis and cohort discovery.
• Further optimization of these methods could enhance their applicability in clinical settings.