Feasibility of constructing a unilateral sudden sensorineural hearing loss machine learning classification model based on diffusion tensor imaging

Abstract

Objective To explore the feasibility of constructing a machine learning classification model for unilateral sudden sensorineural hearing loss (SSHL) patients and normal controls based on diffusion tensor imaging. Methods Prospective collection of 84 patients with untreated SSHL were recruited from the otolaryngology department of the Union Hospital of Tongji Medical College of Huazhong University of Science and Technology between June 2013 to May 2015 as the SSHL group. Meanwhile, a total of 63 healthy volunteers who were no any ear disease history, and the hearing function were confirmed with pure tone audiometry, were collected as the control group. All subjects underwent a brain DTI scan. The data were divided into the training set and validation set according to the ratio of 7 to 3, that was, the training set contained 58 cases of SSHL patients and 44 control groups, and the validation set included 26 cases of SSHL patients and 19 control groups. A vector which included the DTI parameters such as fractional anisotropy, mean diffusivity, axial diffusivity and radial diffusivity was constructed with the software R. The LASSO regression of machine learning method was used to perform feature dimensionality reduction and construct a classification model. The training set samples were used to map the nomogram based on the multivariate logistic analysis method, the validation set and the AUC were used to evaluate the prediction ability of the nomogram, and the calibration curve was used to evaluate the model. Results From the 200 feature vectors including the fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD) values of each brain region, after each dimension reduction process, a total of six features were retained, which were the MD of left superior corona radiate and right superior fronto-occipital fasciculus, the AD of the body of corpus callosum, and the RD of left inferior cerebellar peduncle, left superior corona radiate and right posterior limb of internal capsule. The six features of patients with unilateral SSHL were higher than the control group, and the difference was statistically significant (P<0.05). Based on this, a two-class model is constructed and a nomogram is drawn. The sensitivity, specificity, accuracy and AUC of the training set were 93.1% (54/58), 72.7% (32/44), 84.3% (86/102) and 0.854, respectively; the sensitivity, specificity, accuracy and AUC of validation set were 80.8% (21/26), 84.2% (16/19), 82.2% (37/45), 0.870, respectively. Nomogram could significantly improve the classification efficiency of the control group and patients, and the model with the LASSO method showed a higher prediction curve than other models. Conclusions The machine learning classification model based on DTI metrics can effectively distinguish patients with unilateral sudden sensorineural deafness from healthy control people. Key words: Hearing loss, sensorineural; Diffusion tensor imaging; Machine learning; Classification; Nomogram; Decision curve analysis