Predicting the Cochlear Dead Regions Using a Machine Learning-Based Approach with Oversampling Techniques

Abstract

Background and Objectives: Determining the presence or absence of cochlear dead regions (DRs) is essential in clinical practice. This study proposes a machine learning (ML)-based model that applies oversampling techniques for predicting DRs in patients. Materials and Methods: We used recursive partitioning and regression for classification tree (CT) and logistic regression (LR) as prediction models. To overcome the imbalanced nature of the dataset, oversampling techniques to duplicate examples in the minority class or to synthesize new examples from existing examples in the minority class were adopted, namely the synthetic minority oversampling technique (SMOTE). Results: The accuracy results of the 10-fold cross-validation of the LR and CT with the original data were 0.82 (±0.02) and 0.93 (±0.01), respectively. The accuracy results of the 10-fold cross-validation of the LR and CT with the oversampled data were 0.66 (±0.02) and 0.86 (±0.01), respectively. Conclusions: This study is the first to adopt the SMOTE method to assess the role of oversampling methods on audiological datasets and to develop an ML-based model. Considering that the SMOTE method did not improve the model's performance, a more flexible model or more clinical features may be needed.

Keywords: cochlear dead region; machine learning; oversampling method; prediction model; synthetic minority oversampling technique.

Figure 1

The distribution of cochlear dead regions according to the hearing thresholds at each frequency in the original data (a) and the oversampled data (b). The overall frequency-specific prevalence of cochlear dead regions was 6.7% in the original data and 18.14% in the oversampled data, respectively.

Figure 2

Classification tree model of the original data (a) and the oversampled data (b). DR, cochlear dead region; WRS, word recognition score; dB, decibel of at each audiometric test frequency; PTA, pure tone average of four frequencies (0.5 kHz, 1 kHz, 2 kHz, and 4 kHz); MD, Ménière’s disease; SNHL, sensorineural hearing loss; NIHL, noise-induced hearing loss; ARHL, age-related hearing loss.