Unique Clinical Language Patterns Among Expert Vestibular Providers Can Predict Vestibular Diagnoses

To identify novel language usage by expert providers predictive of specific vestibular conditions.

Retrospective chart review and natural language processing. Level IV.

Tertiary referral center.

Patients seen for vestibular complaint.

Natural language processing and machine learning analyses of semantic and syntactic patterns in clinical documentation from vestibular patients.

Accuracy of Naïve Bayes predictive models correlating language usage with clinical diagnoses.

Natural language analyses on 866 physician-generated histories from vestibular patients found 3,286 unique examples of language usage of which 614 were used 10 or greater times. The top 15 semantic types represented only 11% of all Unified Medical Language System semantic types but covered 86% of language used in vestibular patient histories. Naïve Bayes machine learning algorithms on a subset of 255 notes representing benign paroxysmal positional vertigo, vestibular migraine, anxiety-related dizziness and central dizziness generated strong predictive models showing an average sensitivity rate of 93.4% and a specificity rate of 98.2%. A binary model for assessing whether a subject had a specific diagnosis or not had an average AUC for the receiver operating characteristic curves of .995 across all conditions.

These results indicate that expert providers utilize unique language patterns in vestibular notes that are highly conserved. These patterns have strong predictive power toward specific vestibular diagnoses. Such language elements can provide a simple vocabulary to aid nonexpert providers in formulating a differential diagnosis. They can also be incorporated into clinical decision support systems to facilitate accurate vestibular diagnosis in ambulatory settings.