Feature-Based Audiogram Value Estimator (FAVE): Estimating Numerical Thresholds from Scanned Images of Handwritten Audiograms.

Abstract

Hearing loss is a public health concern that affects millions of people globally. Clinically, a person's hearing sensitivity is often measured using pure-tone audiometry and visualized as a pure-tone audiogram, a plot of hearing sensitivity as a function of frequency. Digital test equipment allows clinicians to store audiograms as numerical values, though some practices write audiograms by hand and store them as digital images in electronic health records systems. This leaves the numerical values inaccessible to public-health researchers unless manually interpreted. Therefore, this study developed machine-learning models for estimating numerical threshold values from scanned images of handwritten audiograms. Training data were a novel set of 556 handwritten audiograms from a longitudinal cohort study of age-related hearing loss. The models were sliding-window, single-class object detectors based on Aggregate Channel Features, altogether called Feature-based Audiogram Value Estimator or "FAVE". Model accuracy was determined using symbol location accuracy and comparing estimated numerical threshold values to known values from an electronic database. FAVE resulted in an average of 87.0% recall and 96.2% precision for symbol locations. The numerical threshold values were less accurate, with 78.3% of estimations having no error, though threshold estimates were not significantly different from true thresholds. Threshold estimation was more accurate than pre-trained deep learning approaches for the current test set. Future work should consider implementing detectors with similar image channels and identify limitations on symbol and axis tick label detection.