HSC Section 8_April 2017

izziness is among themost common chief complaints

among patients presenting to frontline clinicians.

1,2

Whereas dizziness is a symptom with many causes,

evaluation by an otolaryngologist is commonly recom-

mended. Treatment of patients with vestibular disorders is

complex, requires additional clinic time, and uses greater re-

sources (eg, videonystagmography [VNG], rotary chair). Many

patients are found not to have otologic disease, leading to pa-

tient and clinician frustration and delays in diagnosis.

Patient history plays a critical role in the evaluation of ves-

tibular complaints.

3,4

The nature of the dizziness (ie, vertigo,

lightheadedness, imbalance), the temporal pattern of the

dizziness (ie, single episode, recurrent), the duration of at-

tacks (ie, seconds, hours), and associated symptoms (ie, hear-

ing loss, headache) can identify otologic vs nonotologic dis-

ease, and even a specific diagnosis.

Physical examinationmay

help establish a vestibular diagnosis but often has normal re-

sults. Similarly, vestibular testing may be useful in establish-

ing a diagnosis but requires a narrowdifferential diagnosis for

correct test selection and interpretation of test results.

A questionnaire focusing on key elements of the history

may provide adequate information for development of a nar-

row differential diagnosis prior to the office visit.

4,6

Our pro-

gram began using a vestibular disorders intake questionnaire

in September 2012. This 10-page questionnaire was designed

as a quality improvement measure to provide more efficient

and timely care to patients. The results of the questionnaire

have been used to direct appointments (eg, to physician,

vestibular therapist, nurse practitioner, neurologist) and to

inform choice of testing (eg, VNG, rotary chair, posturogra-

phy, vestibular evoked myogenic potentials [VEMPs]). Sub-

jectively this appears to have improved clinical efficiency but

places time burden on administrative and clinical staff toman-

age this system.

We performed data analyses of the triage questionnaire.

This study used414 consecutive patient questionnaires for de-

scriptive analyses andpredictivemodel building. Results of this

studymay be generalized to practicemanagement for allocat-

ing resources and improving efficiency of patient evaluation.

Methods

Approval was obtained from the Medical College of Wiscon-

sin institutional review board. Informed consent was waived

due to the retrospective nature of this study. This project ana-

lyzes a clinically used intake questionnaire specifically de-

signed to triage new patients with vestibular disorders.

Questionnaire

Thequestionnairewas developedatMayoClinic andwasmodi-

fiedslightlybeforebeing implemented inour institution. Atotal

of 162 data variables were captured from each questionnaire.

The questionnaire captures demographic information includ-

ing medical, family, and social history, and current medica-

tion use. There are sections that focus on:

1. The nature of the dizziness perception. This includes a

series of check boxes to describe the dizziness, and ques-

tions as to the onset, duration, and frequency of spells (epi-

sodes), triggers for spells, and the relationship of spells to

motion.

2. Headache, migraine, and migraine-associated symptoms.

3. Otologic problems including hearing loss, tinnitus, aural

pressure, otalgia, and otorrhea.

4. Prior tests and results including audiograms, imaging,

VEMPs, VNG, rotary chair, cardiacHoltermonitors, tilt table

testing, and so forth.

Predictive Model Development

The development of predictivemodels for the diagnosis of be-

nign paroxysmal positional vertigo (BPPV), Ménière’s disease,

andvestibularmigraine incorporatedan initial data set for iden-

tifying key variables for further data collection and a large data

set for predictive model building. The initial group consisted

of 212 consecutive newpatient intake questionnaires. All vari-

ables and fields were collected from these questionnaires for

analysis.We

initially tried todevelopmodels using all available

variables, but this resulted in complex algorithms with unsat-

isfactory sensitivity and specificity. By repetitively narrowing

the data set, and checking for improvements in sensitivity and

specificity, we identified a set of factorswith strong correlation

with specific diseases. Asubsequent 202 consecutivequestion-

naires were then interrogated for this narrow set of variables.

These variables from the combined 414 questionnaires were

then analyzed to build the statistical models for diagnosis

predictions.

Statistical Analysis

The initial data set was screened to identify variables using 3

criteria: (1) significant (

< .05) associationwith the 3 diagno-

ses, (2) sufficient number of observations (≥5 per cell after

cross-tabulation with the outcome), and (3) clinical impor-

tance and relevance. All variableswere converted intodichoto-

mous form (ie, 0 = absent; 1 = present).

The final data set had information on 414 individuals, of

which 381 were ultimately fully evaluable (see Results). Lo-

gistic regression analyses were performed to build parsimo-

nious predictive models with model variables significant at

< .05. All 2-way interactions between significantmodel vari-

ables were investigated for statistical significance. A forward

stepwise variable selectionprocedurewas used. The 3 final par-

simonious models included only variables significant at the

< .02 level, more stringent than the initially planned signifi-

cance cutoff of .05. The receiver operating characteristic (ROC)

curve, area under ROC curve, sensitivity, and specificity at se-

lected cutoffs (ie, linear predictor [LP] values) were assessed

using 10-fold cross-validation.

Thestatisticalanalysiswasperformedusingtheopen-source

software R, version 3.1.1

( http://www.r-project.org).

Two-

tailed Wald tests were used for statistical significance testing.

Results

Of the 414 questionnaires analyzed, 381 had clinical informa-

tionnecessary to define a final diagnosis (

Figure

). Of these, 183

Research

Original Investigation

Statistical Model for the Prediction of Common Vestibular Diagnoses

JAMA Otolaryngology–Head & Neck Surgery

April 2016 Volume 142, Number 4

(Reprinted)

jamaotolaryngology.com