Diagnosing Middle Ear Dysfunction in 10- to 16-Month-Old Infants Using Wideband Absorbance: An Ordinal Prediction Model

Abstract

Purpose The aim of this study was to develop an ordinal prediction model for diagnosing middle ear dysfunction in 10- to 16-month-old infants using wideband absorbance. Method Wideband absorbance, tympanometry, and distortion product otoacoustic emissions were measured in 358 ears of 186 infants aged 10–16 months ( M age = 12 months). An ordinal reference standard (normal, mild, and severe middle ear dysfunction) was created from the tympanometry and distortion product otoacoustic emission results. Absorbance from 1000 to 5657 Hz was used to model the probability of middle ear dysfunction with ordinal logistic regression. Model performance was evaluated using measures of discrimination ( c -index) and calibration (calibration curves). Performance measures were adjusted for overfitting (bias) using bootstrap resampling. Probabilistic and simplified methods for interpreting the model are presented. The probabilistic method displays the probability of ≥ mild and ≥ severe middle ear dysfunction, and the simplified method presents the condition with the highest probability as the most likely diagnosis (normal, mild, or severe middle ear dysfunction). Results The c -index of the fitted model was 0.919 (0.914 after correction for bias), and calibration was satisfactory for both the mild and severe middle ear conditions. The model performed well for the probabilistic method of interpretation, and the simplified (most likely diagnosis) method was accurate for normal and severe cases but diagnosed some cases with mild middle ear dysfunction as normal. Conclusions The model may be clinically useful, and either the probabilistic or simplified paradigm of interpretation could be applied, depending on the context. In situations where the main goal is to identify severe middle ear dysfunction and ease of interpretation is highly valued, the simplified interpretation may be preferable (e.g., in a screening clinic that may not be concerned about missing some mild cases). In a diagnostic clinical environment, however, it may be beneficial to use the probabilistic method of interpretation.

Diagnosing conductive dysfunction in infants using wideband acoustic immittance: development and validation of predictive models

“…Seven-hundred and fifty-three infants were recruited at birth and followed through infancy (Myers et al, 2018a(Myers et al, , 2018b(Myers et al, , 2019b(Myers et al, , 2019c Table 6.1. Infants were scheduled to be followed up at around 6, 12, and 18 months of age.…”

“…For the 6-month infants, HFT results were classified in a similar way (drawing a line), but with a third category of "negative peak" for peaked traces with tympanometric peak pressure <−150 daPa. LFT results for the 12-and 18-month groups were classified as "type A" if there was a peak |Y| ≥0.3 mmho between −150 to 50 daPa, "type C" if the peak (≥0.3 mmho) occurred at tympanic peak pressure <−150 daPa, and otherwise "type B" (Myers et al, 2019b;Roush et al, 1995). HFT results for the 6-month group (peaked, negative peak and not peaked) are also referred to as type A, type C and type B, respectively in this report since data from the 6-, 12-, and 18-month groups were analysed together.…”

“…Diagnostic studies have shown that WAI can accurately diagnose conductive dysfunction in neonates (Aithal et al, 2015;Hunter et al, 2010;Keefe et al, 2003a;Myers et al, 2018a;Sanford et al, 2009), and infants Hunter et al, 2008b;Myers et al, 2018bMyers et al, , 2019bPrieve et al, 2013a). Using WAI clinically typically involves subjective interpretation of visual displays.…”

“…An issue with developing such algorithms for use with infants, however, is that there are substantial maturational effects on WAI that need to be controlled for (Aithal et al, 2014b;Hunter et al, 2010;Hunter et al, 2016;Keefe et al, 1993;Keefe et al, 2000;Kei et al, 2013;Myers et al, 2018aMyers et al, , 2019cSanford et al, 2009;Shahnaz et al, 2014;Werner et al, 2010). One solution to this problem has been to create age-specific algorithms, such as models that have been developed for neonates Myers et al, 2018a;Sanford et al, 2009) or different age groups through infancy (Myers et al, 2018b(Myers et al, , 2019b. Alternatively, models could be developed with age included as an interaction term with WAI, which would allow interpretation to vary with age (Myers et al, 2019c).…”

“…Another approach would be to only include frequency regions that are relatively unaffected by age as predictors in a model (Myers et al, 2019c;Sanford & Feeney, 2008). For example, for infants aged 0 to 6 months, Shahnaz et al (2014) found limited developmental effects on R from 600 to 1600 Hz, and Sanford and Feeney (2008) Another issue in diagnostic WAI research is that although many predictive models have been developed Keefe et al, 2003a;Keefe et al, 2003b;Myers et al, 2018aMyers et al, , 2018bMyers et al, , 2019bPiskorski et al, 1999;Sanford et al, 2009), none have been validated in an external sample of subjects. Some studies have estimated the amount of bias (overfitting) in a model with bootstrap resampling, or by developing the model on one ear of subjects and applying it to the opposite ears Keefe et al, 2003b;Myers et al, 2018aMyers et al, , 2018bMyers et al, , 2019b…”