Background: Adult hearing screening can help increase awareness, identify the early signs of hearing impairment and therefore trigger timely intervention, thus preventing or delaying the progression of hearing impairment and limiting its impact on communication and psychosocial functioning. Speech-in-noise tests are valuable measures of hearing ability in real-life conditions and can be helpful to promote awareness and detect age-related hearing impairment at early stage. In the area of hearing loss prevention, a still largely unexplored area is related to the assessment of risk factors, including modifiable (e.g., cigarette smoking, noise exposure, medications) and non-modifiable risk factors (e.g., genetic predisposition, age, and co-morbidities such as diabetes and hypertension). Some risk factors are specific to hearing impairment whereas some others are common to hearing impairment and cognitive decline or dementia. To date, no standardized methodology is available to assess the risk of developing hearing impairment and the associated cognitive decline, and to help understand the potentially modifiable risk factors that determine the individual risk. The aim of this study was to develop and evaluate a novel user-operated platform to support widespread screening and prevention of hearing impairment and the associated cognitive decline using a combination of speech-in-noise testing, cognitive testing, and risk factors assessment. Methods: The proposed WHISPER (Widespread Hearing Impairment Screening and PrEvention of Risk) system is implemented on an easy-to-use graphical user interface and combines: (i) an adaptive speech-in-noise test based on multiple-choice recognition of vowel-consonant-vowel stimuli; (ii) an adaptive cognitive test, specifically the visual digit span working memory test (forward modality); and (iii) an icon-based interface to assess the main modifiable and non-modifiable risk factors for developing hearing impairment and cognitive decline. We have tested the platform on a total of >250 participants and we have extracted >30 features from the speech-in-noise test, the cognitive test, and the risk factors questionnaire. We have characterized the features using correlation and regression analysis and we have developed machine learning algorithms, including eXplainable AI techniques, to assess the relationships between the variables and the variables that are associated with hearing loss and cognitive decline. Results: The accuracy of machine learning models to predict hearing loss (mild degree or higher) was higher than 90% using different combinations of input features. The highest performance was observed using a subset of features (i.e., age, speech recognition threshold, number of correct responses, number of trials, presence of cardiovascular diseases, and level of education), specifically sensitivity = 95%, specificity = 95%. The cognitive performance, as measured using several features extracted from the digit span test (e.g., digit span score, average reaction time per digit) was associated with speech recognition performance, with the pure-tone average, with age, and with the level of education. The analysis of psychometric functions of vowel-consonant-vowel stimuli, as estimated using the individual responses in the adaptive speech-in-noise test, revealed that additional features, in addition to the speech recognition threshold, could be extracted to further improve classification ability (e.g., slope, dispersion). Further research in this direction will be helpful to understand the optimal set of features for hearing loss classification. Results from user experience surveys (specifically, the user engagement questionnaire, UEQ, and net promoter score, NPS) showed that all the items of the UEQ were, on average, higher than 0 (neutrality score) and that 22 of the 26 items were higher than 0.8. The top ranked factors were perspicuity, efficiency, and novelty. The NPS showed that none of the tested subjects were classified as detractors and that they were nearly evenly distributed between neutrals and promoters. Conclusions: This study showed that the WHISPER platform can accurately identify hearing loss in adults and that the platform is perceived as clear, efficient, and original by the tested subjects. Further research is needed to expand the set of features and develop models able to characterize the risk of developing hearing loss and the related cognitive decline on a larger population.
WHISPER (Widespread Hearing Impairment Screening and PrEvention of Risk): a New Platform for Early Identification of Hearing Impairment and Cognitive Decline
Paglialonga A;
2022
Abstract
Background: Adult hearing screening can help increase awareness, identify the early signs of hearing impairment and therefore trigger timely intervention, thus preventing or delaying the progression of hearing impairment and limiting its impact on communication and psychosocial functioning. Speech-in-noise tests are valuable measures of hearing ability in real-life conditions and can be helpful to promote awareness and detect age-related hearing impairment at early stage. In the area of hearing loss prevention, a still largely unexplored area is related to the assessment of risk factors, including modifiable (e.g., cigarette smoking, noise exposure, medications) and non-modifiable risk factors (e.g., genetic predisposition, age, and co-morbidities such as diabetes and hypertension). Some risk factors are specific to hearing impairment whereas some others are common to hearing impairment and cognitive decline or dementia. To date, no standardized methodology is available to assess the risk of developing hearing impairment and the associated cognitive decline, and to help understand the potentially modifiable risk factors that determine the individual risk. The aim of this study was to develop and evaluate a novel user-operated platform to support widespread screening and prevention of hearing impairment and the associated cognitive decline using a combination of speech-in-noise testing, cognitive testing, and risk factors assessment. Methods: The proposed WHISPER (Widespread Hearing Impairment Screening and PrEvention of Risk) system is implemented on an easy-to-use graphical user interface and combines: (i) an adaptive speech-in-noise test based on multiple-choice recognition of vowel-consonant-vowel stimuli; (ii) an adaptive cognitive test, specifically the visual digit span working memory test (forward modality); and (iii) an icon-based interface to assess the main modifiable and non-modifiable risk factors for developing hearing impairment and cognitive decline. We have tested the platform on a total of >250 participants and we have extracted >30 features from the speech-in-noise test, the cognitive test, and the risk factors questionnaire. We have characterized the features using correlation and regression analysis and we have developed machine learning algorithms, including eXplainable AI techniques, to assess the relationships between the variables and the variables that are associated with hearing loss and cognitive decline. Results: The accuracy of machine learning models to predict hearing loss (mild degree or higher) was higher than 90% using different combinations of input features. The highest performance was observed using a subset of features (i.e., age, speech recognition threshold, number of correct responses, number of trials, presence of cardiovascular diseases, and level of education), specifically sensitivity = 95%, specificity = 95%. The cognitive performance, as measured using several features extracted from the digit span test (e.g., digit span score, average reaction time per digit) was associated with speech recognition performance, with the pure-tone average, with age, and with the level of education. The analysis of psychometric functions of vowel-consonant-vowel stimuli, as estimated using the individual responses in the adaptive speech-in-noise test, revealed that additional features, in addition to the speech recognition threshold, could be extracted to further improve classification ability (e.g., slope, dispersion). Further research in this direction will be helpful to understand the optimal set of features for hearing loss classification. Results from user experience surveys (specifically, the user engagement questionnaire, UEQ, and net promoter score, NPS) showed that all the items of the UEQ were, on average, higher than 0 (neutrality score) and that 22 of the 26 items were higher than 0.8. The top ranked factors were perspicuity, efficiency, and novelty. The NPS showed that none of the tested subjects were classified as detractors and that they were nearly evenly distributed between neutrals and promoters. Conclusions: This study showed that the WHISPER platform can accurately identify hearing loss in adults and that the platform is perceived as clear, efficient, and original by the tested subjects. Further research is needed to expand the set of features and develop models able to characterize the risk of developing hearing loss and the related cognitive decline on a larger population.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
