Interested participants contacted the study administrator via email and received extensive written information about the purpose of the project; study design; length of participation and remuneration; possibility to withdraw participation at any time; and data protection, management, and storage. The study design and implementation and data collection, analysis, and storage were conducted in accordance with current literature on ethical considerations in the context of mobile and mHealth apps [34,35]. Security and privacy recommendations were also adhered to. It was clearly stated that a medical diagnostic was not provided in the study. Communication with the participants took place exclusively via email and SMS text message. A pilot study was conducted with a young (aged 23 years), healthy female participant in August 2021 to evaluate the usability and technical functionality of the mobile study.

The web-based study was conducted on the personal mobile phones of the participants to approximate the experience of using an app. Only 3.8% (7/185) of the participants used their computers due to technical difficulties with their smartphones. Data collection was carried out using formr, an open-source web-based application programming interface (API) for the R language that creates automated studies [36]. In formr, different questionnaires and tests (refer to Tables 2 and 3) were linearly chained together as modules of a so-called run. A run reproduces the desired design and can be accessed by users through a specific link. The formr software first provides a unique study link to the run, which was shared via email with enrolled participants. Upon accessing this link, participants were assigned a unique visitor session in formr and provided with a second individualized link based on web cookies. The unique visitor session prevented users from providing multiple entries for the same survey. This unique session code enabled the anonymization of the data within formr and, for the duration of the study, was stored in a written coding list, where the participants’ names and session codes were recorded. The customer communication platform Twilio (Twilio Inc) [37] was used to send the individualized study link to the participants through daily SMS text message reminders. Automated SMS text message delivery was initiated via the external link module in formr, which uses Representational State Transfer API to connect to Twilio. Representational State Transfer API allows a software program (in this case, formr) to expose functionality and data to other programs (Twilio) in a consistent and secure format, ensuring privacy and data protection. With the individualized link received via SMS text message, participants could perform the study on their own smartphone’s browser. For the 3.8% (7/185) of participants who completed the study on their computer, the daily SMS text messages were sent to their personal mobile phones as reminders. The individualized link was additionally sent via email at the beginning of the study to allow these participants to access the study via their computer browser.

A detailed list of all assessment tools and their references, as well as the derived features for analysis, is provided in Tables 2 and 3. Each assessment tool was implemented as a survey in formr. Most of the surveys included in the study were implemented following the paper-and-pencil version that was retrieved from the literature. The surveys that were developed or adapted specifically for this study can be shared upon request. The items assessed can be inferred from the features provided in Tables 2 and 3. Submission of each survey was possible only after all mandatory questions had been answered. After submission of one survey, the study advanced automatically to the following questionnaire or test planned in the formr run. Users were not allowed to go back and modify their answers after submission.

The total assessment time of 8 hours was distributed across the working days of 3 consecutive weeks, with an overall daily active participation of approximately 30 minutes. The study design is detailed in Multimedia Appendix 1. Participants could select the start date of the study to ensure that the assessment could be easily integrated into their personal schedule. The first week (baseline assessment) included 1 measurement time point per day (requiring approximately 20-30 minutes to complete), which could be performed at any time. On the first day, participants received an email with the study link and their unique participant code. After opening the study link on their browser, each participant was given the possibility to read a summary of the data protection conditions in the first page of the study again. Second, they were asked to provide a telephone number and email address, which were then stored in formr and used for the automatic SMS text message reminders. They would then receive an automatic SMS text message with the individualized link to the study, through which they could begin the assessment. From days 2 to 5, participants received an SMS text message with the link to the study at 7 AM, but they had been previously informed that they could perform the tasks at any time during the day. An email reminder was sent at 7 PM in case participants had not accessed the study link by that time.

The second and third week included 2 measurement time points per day of approximately 15 minutes each. The longitudinal assessments were prompted via SMS text message at 7 AM and 7 PM. Participants were instructed to access the study at their earliest convenience after waking up and before going to bed, thus allowing them to accommodate the study to their daily schedules. In the morning, after clicking on the link they received via SMS text message at 7 AM, each participant was first presented with some questions on baseline mood and sleep quality. They were then required to click on a second link embedded in the following survey page that redirected them to the hearing assessment. Finally, participants were asked again to report on their mood after receiving feedback on their hearing performance. In case the participant forgot to access the link and perform the study, an email reminder was sent at 1 PM. If, after the reminder, the participant still did not take part in the study, the session was established as incomplete. The study administrator had to manually allow the participant to move to the next measurement time point (in this case, the evening assessment) within formr. This same assessment scheme was repeated in the evening. In this case, the SMS text message with the link to the study was sent at 7 PM, and the email reminder was sent at 11 PM.

At the end of the study, participants were asked to provide consent to be contacted after 2 months for a voluntary (and nonremunerated) follow-up questionnaire. Those who provided their consent received an email with a link to a single formr survey that required <5 minutes to complete. The short survey consisted of 2 multiple-choice questions and was completed by 70.8% (131/185) of the participants. Individuals were asked to report again on their action to seek professional help following the feedback received during the study and were asked to indicate whether study participation improved their awareness of hearing difficulties.

This microlongitudinal assessment accounts for potential daily fluctuations in hearing performance and affect, which might depend on particular daily events and states. The affect questionnaire included 14 items in line with the circumplex model of affect (Posner et al [64]). A total of 8 items were related to negative affect, and 6 items were related to positive affect [65]. The items are listed in Multimedia Appendix 1. The affect questionnaire was presented before and after the hearing test to assess mood at baseline and after receiving feedback on the hearing test, respectively.

Hearing performance was assessed using the digit triplet test (DTT) [45,66] by Hörzentrum Oldenburg gGmbH. This widely used screening instrument [67] measures speech intelligibility in noise by means of the speech recognition threshold (SRT), which indicates the signal-to-noise ratio (SNR; difference between speech and noise level) at which the participant reaches 50% speech intelligibility. SRT measures obtained using the DTT showed high correlations (r>0.70) with pure tone average measures while being relatively robust against changes in presentation level [68]. Moreover, the DTT has shown to be robust to ambient noise levels outside of audiometric booth environments [68]. Together with its low linguistic and cognitive demands [68], the DTT appears to be suitable for mobile, remote self-test–based screening of hearing abilities in the older population. Smartphone-based DTT has also shown the potential to provide widespread access to hearing screening in low- and middle-income countries and across different socioeconomic strata [69]. After completing the hearing test, participants received feedback on their performance in the form of a traffic light color, where green indicated good performance (SRT<–7.1 dB SNR), yellow indicated intermediate performance (–7.1≥SRT<–5.1 dB SNR), and red reflected poor performance (SRT≥–5.1 dB SNR) [45,70]. The Multimedia Appendix 1 provides detailed information on the hearing test and Multimedia Appendix 2 provides information on its feedback. Participants performed the hearing test with their personal smartphone and headphones. A total of 1.6% (3/185) of the participants used loudspeakers as headphones were not available to them. Calibration of the hardware equipment was not possible due to the remote assessment. However, SRT estimation is relatively robust against changes in presentation level, and no exact calibration is needed [71]. Moreover, the use of different types or qualities of headphones has shown no impact on test reliability [67,69]. Each hearing test began with a signal adjustment trial meant to set the stimulus at approximately 65 dB sound pressure level. The participant was presented with a digit triplet in noise and asked to “adjust the volume to hear both the digits and the noise clearly.”

Data analysis was performed using the R software (R Foundation for Statistical Computing) [72]. Raw data from all questionnaires were imported from the web-platform formr to the R environment using the dedicated formr package [73]. For each questionnaire or test presented at baseline, global scores were computed and considered as features. If both global scores and scale scores were available for a given assessment tool, only the scale scores were retained if considered differentially relevant for the outcome. Hearing test results were sent via email to the investigator from the researchers of Hörzentrum Oldenburg and imported into R as .eml files. The performance feedback category (green, yellow, and red) was additionally extracted and stored for each raw SRT result. Due to the particular implementation of the study in formr, participants could perform the hearing test more than once at each measurement time point. Whenever this happened, only the last SRT result at a given time point was kept for analysis. This led to a removal of 3.9% of the raw SRT results. The longitudinal data on daily mood and hearing performance were summarized into individual means and SDs. The summarized longitudinal data were merged with the cross-sectional data, resulting in a wide-format data frame including 83 features. Tables 2 and 3 provide a complete list of the features considered for analysis.

Completion rate for the baseline questionnaire was 100% (185/185), whereas there were missing data for the longitudinal measures of hearing performance and affect. A complete set of 20 SRT results was collected for 43.8% (81/185) of the participants, whereas at least 15 SRT results were obtained in 95.1% (176/185) of cases. Missing hearing data at a specific time point were considered not available (NA). Where an SRT result was missing, the respective feedback and measures of affect at the posttest time point were established as missing as well (NA). Through visual inspection of the individual SRT distributions, some specific outlier patterns were identified. A total of 7.6% (14/185) of the participants showed a much larger SRT result at the first measurement, which qualified as an outlier following the IQR rule. These large SRT values (indicating poor performance) were considered to be caused by misunderstanding of the hearing test instructions and, therefore, were established as NA. However, the respective feedback category and measures of affect where not established as NA. This is because, despite the unreliable SRT value, participants’ mood could still have been affected by the feedback received. With respect to daily affect measures, 1.1% (2/185) of the participants provided no data at measures of posttest affect such that summary measures could not be computed. In these cases, the mean imputation technique was applied—the sample mean and sample variance for negative and positive affect at the posttest time point were imputed to replace missing values.

After identifying the best-performing machine learning algorithm, feature importance was considered. Each feature receives a coefficient of importance that indicates its contribution to model performance regardless of the type of relationship (direction and linearity) between the feature and the outcome [85]. In RF, feature importance is model dependent and indicates how much the feature contributes in reducing node impurity. Importance values were retrieved using the function getFeatureImportance() (mlr package) applied on the RF model trained using the tuned hyperparameters. These importance results have the advantage of being inherent to the model and closely tied to its performance [86]. Conversely, there are no model-specific importance metrics available for NB and SVM. For these algorithms, the importance value assigned to each feature corresponds to the area under the receiver operating characteristic curve, which is computed from sensitivity and specificity measures [86]. The function varImp() from the caret package [87] was applied on the model trained using the function train() (caret package) after ensuring comparable performance with that of the same model previously trained on the mlr package.

Features with higher importance ranking were considered for inclusion in the profiling algorithm as they represented the most relevant predictors for telling apart help seekers and nonseekers. No statistical criterion exists to determine which importance value threshold should be used to retrieve relevant features. Hence, 3 threshold values were inspected (the first 10, 15, and 20 features in their importance ranking order) and evaluated in terms of predictive accuracy and interpretability. The classification accuracy of these 3 feature sets was assessed on the outcome data obtained at follow-up. For this analysis, the data set was reduced to 131 participants who completed the follow-up questionnaire. The important features were fed into the best-performing machine learning algorithm from the previous step.

A summary of the model-specific classification accuracy for the first outcome (action to seek help) is provided in Table 5. The 3 algorithms showed similar overall performance accuracy estimates on the test set, correctly classifying approximately 65.9% (122/185) to 70.3% (130/185) of the cases in the full data set (n=185). RF was the best-performing algorithm with an accuracy of 70.3% (130/185) and an MCC of 0.28, indicating that the model’s prediction improved to approximately 20% over chance. By inspecting the confusion matrix (Figure 1), we observed that RF shows high specificity, correctly classifying 90.9% (110/121) of the cases belonging to the no action class. The NB classifier (10-fold CV repeated 50 times) showed the best sensitivity compared with the competing algorithms, with 51% (33/64) of cases in the action class being correctly classified. RF and NB were selected for feature importance analyses given the good predictive performance and high specificity of RF as well as the high sensitivity of NB.

Table 6 summarizes the classification performance with respect to the second outcome (action and motivation to seek help), which includes 3 classes (refer to the previous section). RF provided the highest accuracy with 55.1% (102/185) and an MCC of 0.30 as compared to NB (10-fold CV repeated 100 times) and SVM. However, the confusion matrix revealed that none of the 3 models was able to adequately tell apart individuals within the no action class who differed with respect to high versus low motivation. All algorithms could only correctly classify 2% (1/47) to 25% (12/47) of cases in the no action and high motivation class. Potentially, an improvement in classification accuracy could be achieved using a larger data set in which the classes are better balanced and with a more reliable and elaborate measure of the participants’ motivation to seek help. In view of these limitations, the second outcome was not considered for feature importance analysis.

Feature importance was analyzed based on the RF and NB algorithms predicting the first outcome, action to seek help at study end on the full data set (n=185). Each importance value signifies the feature’s contribution to the model’s performance. However, as detailed in the Methods section, RF and NB models calculate these coefficients differently. Consequently, ranking values were used. In both models, the 83 features were initially ranked in descending order based on their importance values. Features with higher importance rankings were mostly relevant in predicting help-seeking behavior. Three sets of features among the most important ones were taken into account for subsequent analysis: (1) the top 10 features indicated by the 2 models, resulting in a total of 12 best features; (2) the top 15 features indicated by the 2 models, resulting in a total of 19 best features; and (3) the top 20 features indicated by the 2 models, resulting in a total of 28 best features.

Figure 2 shows all 28 features with their importance ranking values originating from the RF and NB models. The specific importance values for each feature are provided in Multimedia Appendix 3. Next, the 3 sets of features (top 10, 15, and 20 features) were evaluated for their predictive performance and classification accuracy on the reduced data set of 131 participants who completed the follow-up questionnaire. NB and RF were trained on the 3 different feature sets for predicting the action to seek help at follow-up. The results are summarized in Table 7. They show that all feature sets provided good predictive performance and that the NB algorithm outperformed RF, with an overall accuracy ranging between 73.3% (96/131) and 74.8% (98/131) and an MCC between 0.43 and 0.47. Class-specific classification accuracy was comparable between NB and RF, with the action class correctly classified in 52% (27/52) to 63% (33/52) of the cases and the no action class correctly classified in 82% (65/79) to 86% (68/79) of cases. As can be seen from the confusion matrix in Figure 1, sensitivity and specificity measures were comparable for both algorithms predicting action to seek help at follow-up using the set of top 10 most important features. Sensitivity ranged from 55% (NB) to 58% (RF), and specificity ranged from 84% (RF) to 65% (NB).

Textbox 3 provides a brief description of 12 features ranked among the most important (previously named as the top 10 features for the 2 models) in the prediction of action to seek help at follow-up.

As outlined in the Methods section, ranking values up to 10 were arbitrarily chosen to identify the most important features. The following two features had slightly lower importance values (and received a ranking value of 11) but might provide further insights into targeted counseling in an mHealth hearing app: (1) neuroticism, which refers to a predisposition to experiencing negative emotions [2] and was assessed through the corresponding items of the Neuroticism-Extraversion-Openness Five-Factor Inventory questionnaire [46]; and (2) monthly income, which was categorized using 3 cutoff values (<€1500 [<US $1611.21], €1500-€2500 [US $1611.11-$2685.35], €2500-€4000 [US $2685.35-$4296.56], and >€4000 [>US $4296.56]).

This study contributes to the identification of individuals’ hearing-related, psychological, and general health–related traits that predict the readiness to seek professional help for HL. Cross-sectional and longitudinal data were collected in a comprehensive mobile study. Potential users of a future mHealth hearing app, namely, individuals with subjective hearing difficulties, were classified into help seekers and nonseekers by means of supervised machine learning algorithms. The trait measures used in this study were collected from previous literature investigating health care seeking, particularly in the audiological domain. From these, we derived a comprehensive set of 83 features to be used for prediction and profiling. The 3 algorithms taken into account (NB, RF, and SVM) accurately predicted help-seeking behavior at the end of the study in 65.9% (122/185) to 70.3% (130/185) of cases. In particular, the RF algorithm achieved high specificity, meaning that it was most successful in identifying individuals who might not intend to seek professional help. By selecting a subset of important traits revealed by our empirical feature importance analyses to predict hearing help seeking, the prediction accuracy for action to seek help at the 2-month follow-up reached 74.8% (98/131). This study identified the following features to be most important in the prediction of help-seeking behavior: perceived consequences of HL in daily life, motivation to seek help, attitudes toward HAs, sensory sensitivity, neuroticism, physical health, and income. We conclude that these individual characteristics should be assessed in a profiling module that could complement the main auditory assessment module for hearing screening of existing or future mHealth apps. The degree of HL but, importantly, also its day-to-day variability were among the most important predictors, suggesting the need to perform repeated hearing assessments, which could be prompted by the app at different times of the day. To streamline the implementation of the profiling module in a mobile app, the questionnaires and subscales used to measure these important features should undergo item selection analysis to derive simple and short yet reliable and valid scales. By incorporating a selected machine learning algorithm (RF), the app can profile users into help seekers or nonseekers based on the data collected through this short questionnaire battery. This information would complement the audiological data gathered through existing hearing screening or diagnostic tests, providing an informative user profile. The derived profile would guide the app in selecting the appropriate set of recommendations, optimizing an intervention on help-seeking behavior where needed. The results of this study will also provide suggestions for the design of such targeted treatment recommendations. Ultimately, our aim was to provide clinicians and mHealth app developers with relevant knowledge to promote hearing health by encouraging the uptake of hearing health care services and HAs when needed.

In total, 3 machine learning classifiers correctly predicted action to seek help at study end in 65.9% (122/185) to 70.3% (130/185) of cases, clearly improving over chance prediction. This is a promising result considering the complexity of the prediction outcome. As discussed previously, several individual factors can influence the decision to seek hearing health care services, and there can be discrepancy among contemplating, planning, and taking concrete action [3]. RF showed the best prediction accuracy and high specificity, whereas NB showed the highest sensitivity. When predicting action to seek help at follow-up using the selected important features, the performance of the RF and NB models improved up to 75% despite the smaller data set (n=131). NB showed higher predictive performance for this outcome. Overall, all models exhibited high specificity (ranging from 75% to 86%) and comparatively low sensitivity (31% to 58%). This could be attributed to the fact that the action class encompassed individuals who had sought professional help as well as those who were only considering taking action. RF showed high accuracy in identifying the no action class both at study end and at follow-up and, therefore, can be considered the best-performing algorithm in this framework. Accurate identification of nonseekers is the most relevant performance outcome in an mHealth app to design targeted recommendations. Indeed, the envisioned profiling algorithm should be a system with high specificity that motivates and promotes help seeking, especially in those cases in which users would not spontaneously take action.

Hearing performance appears to be one of the most important features to predict help seeking. The association between degree of HL and help seeking, as well as HA uptake, is well established in the literature [4,9,26,28,29]. These results also highlight—to our knowledge, for the first time in the literature—the predictive role of intraindividual fluctuations in hearing performance, emphasizing the need to move beyond the traditional view of hearing as a stable neurosensory process [88]. The implementation of repeated daily measurements of hearing performance provides further insights on the impact of HL on the individual’s everyday life. In line with this, feature importance findings emphasize the relevance of self-reports on the consequences of HL. The assessment should consider self-reported listening effort in different contexts as well as perceived handicap and emotional consequences of HL. Indeed, individuals who report greater negative impact of HL in their lives are more prone to seek help and later uptake HAs [9,27]. Individuals’ self-awareness of HL can be validated or improved by providing repeated feedback on hearing performance in an mHealth hearing app. As observed at the follow-up survey, 85.5% (112/131) of participants reported increased awareness of their hearing abilities after receiving repeated feedback during the study. Finally, according to these results and previous findings [26,27,29], investigating stigma, attitude, and expectations regarding HAs informs on individuals’ readiness to seek help as well as later uptake of an HA. Stigma and negative stereotypes related to HAs may deter individuals from seeking help and can represent a barrier to HA use [7,28].

Audiological factors emerged as the most important features. Nevertheless, other general health and psychological factors were also relevant in the prediction of help seeking. In this study, physical health was an important predictor for help seeking, although the evidence on this relationship is discordant [26]. While people with better self-reported health were more likely to seek help [4,27], HA uptake was predicted by poor self-reported health [6]. Other important factors were related to the personality traits of sensory sensitivity and neuroticism. Individuals characterized by high sensitivity to sensory stimuli [63] and emotionally instable personality traits seem to perceive increased psychological discomfort following HL even in the presence of effective HA treatment [2]. Finally, income emerged as another important predictive feature. This is in line with evidence suggesting that higher socioeconomic status [9], higher income or pension earnings [3,27], and access to financial support [26] promote HA uptake.

By assessing and analyzing the aforementioned hearing-related and psychological traits, the algorithm developed in this study aims to profile the user as help seeker or nonseeker. Completing this profile with complementary audiological test results, an informative picture of the user can be derived. Using this profile, clinical experts and intervention app designers could propose different sets of recommendations to assist individuals in their decision-making process in a targeted manner. We propose to first differentiate between profiled help seekers and nonseekers, where the former should receive simple and straightforward recommendations only depending on their hearing status. For nonseekers, there is a need to design more specific and targeted recommendations based on information about the relevant characteristics to predict HA seeking. Users who were profiled as determined help seekers could receive clear and concise guidance on the hearing care they need. Those among them who should uptake a hearing device (given their audiological outcome) could benefit from additional information on available hearing care services and professionals to facilitate faster HA adoption rates. This would facilitate individuals’ perceived competence and autonomy, which are important predictors of hearing health–seeking behavior [89]. On the other hand, users with HL who are profiled as nonseekers should receive more elaborate, targeted recommendations to motivate and promote access to hearing care services. Recommendations for nonseekers should be further differentiated and designed depending on their perceived consequences of HL; attitudes toward HAs; sensory sensitivity; neuroticism; and, potentially, income. These recommendations could act as an intervention on modifiable predictive features such as self-recognition of HL and attitude toward HAs. For example, users profiled as nonseekers with good awareness and self-recognition of HL but negative expectations regarding HAs and low income should receive a different set of recommendations than nonseekers with low self-awareness and high neuroticism.

Where HL is detected, the mHealth app could prompt repeated testing on different days and at different times of the day and provide individual feedback on the performance compared to normative data. More detailed feedback on daily hearing performance could improve awareness of the hearing deficit. The app could also inform the user of the risks of an untreated HL and the benefits of early intervention through HAs. Indeed, it has been shown that individuals are more likely to positively change their behavior when provided with actionable and meaningful information on their health status [90]. Where there is a need to promote positive attitudes toward HAs, information could be provided on the wide range of devices available as well as examples of successful peer cases. Knowledge of accessible financial support for HAs by insurance companies could additionally promote HA uptake given the predictive role of income. Furthermore, an implemented HA simulator in an mHealth hearing app could offer possibilities to experience improved listening conditions and promote positive expectations regarding an HA. Elaborate information on HA technologies, such as the benefits of noise control and noise reduction algorithms, could promote help seeking by fostering the knowledge of individuals who are more sensitive to environmental noise (high sensory sensitivity trait). The effectiveness of such recommendations could be further increased through targeting or tailoring communication. Targeted messages are designed for a specific population, whereas tailored communication is individualized to the person and has been shown to be most effective in promoting health behavior change [91]. Indeed, messages that are congruent with the personality traits of the audience are more positively evaluated and persuasive and generate more interest [92]. The predictive role of neuroticism for help-seeking behavior can be considered for efficient communication both in the context of an mHealth hearing app and in clinical counseling. Individuals with a high neuroticism trait are more susceptible to perceived disease [93] and are drawn to action through motives of safety and security [92]. For example, recommendations that target profiled nonseekers with low expectations regarding HAs can be differentiated depending on personality traits. To promote positive HA expectations in individuals with high sensory sensitivity traits, recommendations could focus on the benefits of HAs related to noise control and noise suppression. However, such recommendations may not be effective for people who do not have this high sensitivity trait and who, for example, score high on neuroticism. Instead, they might be more convinced by recommendations that emphasize the risks of an untreated HL and the benefits of an early intervention through HAs.