Autosomal dominant nonsyndromic hearing loss (ADNSHL) presents a significant genetic diagnostic challenge due to its underlying heterogeneity—more than 64 genes are implicated in its etiology [1]. Because of this complex genetic landscape, computational tools designed to correlate audiogram profiles (commonly called audioprofiles) with specific genes have achieved only moderate success [2,3]. One such tool, which we developed approximately 15 years ago, is AudioGene. AudioGene uses numerous machine learning (ML) approaches to improve diagnostic precision [2,4,5]. These approaches include semisupervised support vector machines (SVMs), ensemble models, and hyper-tuning methods. However, challenges such as data imbalance and class sparsity continue to restrict the accuracy of these models.

Precision medicine harnesses information about the genome of an individual, environment, and lifestyle to guide medical care. With heterogeneous disorders such as ADNSHL, genetic variant interpretation can be challenging, complicating the diagnostic process and impacting patient care. Computational tools may improve the precision and reliability of genetic assessments by capitalizing on genotype-phenotype associations [6,7].

Current diagnostic methods for ADNSHL largely rely on clinical examination and audiometry, which do not provide sufficient resolution for the complex genetic landscape of ADNSHL [2,4]. However, with the availability of ML and artificial intelligence–driven approaches, there has been a shift toward integrating computational and visualization tools with genetic diagnostics to improve accuracy and predictive power [8,9].

To address these challenges, we have developed the AudioGene Translational Dashboard with the goal of enhancing both the accuracy and interpretability of genetic predictions. A feature of the AudioGene Translational Dashboard is the “70/30” phenomenon: by integrating the results from both models on a training dataset comprising 3189 audiograms from 1445 patients, we observed that the correct disease-causing gene was predicted within the top 3 predictions 74% of the time, with incorrect predictions accounting for the remaining 26%, hence “70/30.” This observation signals to health care providers when they can have confidence in the top predictions, serving as a “green flag” or “red flag” in the diagnostic process. Having a true positive rate of 70% is beneficial from a research perspective; however, for a diagnostic tool, the remaining 30% represents some risk that necessitates additional interpretative context. By providing this context, the AudioGene Translational Dashboard enables health care providers to weigh their confidence in the predictions, supporting more informed diagnostic decisions.

The AudioGene Translational Dashboard was introduced into the AudioGene toolset to increase transparency into the “black box” underlying the models by providing explainable artificial intelligence (XAI) to enhance model interpretability and utility in clinical settings, in line with trends in precision medicine that emphasize the importance of genotype-phenotype associations in improving diagnostic outcomes [7,10].

Early attempts to map audiometric phenotypes to their underlying genotypes were spearheaded by AudioGene version 4 (AG4), a semisupervised multi-instance SVM that treats the collection of audiograms for a single patient as a “bag” and ranks loci according to pair-wise–coupled probability estimates [2,11]. Building on this foundation, AudioGene version 9.1 (AG9.1) introduced selective intraensemble data partitioning: training examples are first divided by gene-specific data volume, patient age, and audiogram shape, then modeled with a committee of k-nearest neighbor (KNN), adaptive boosting, and random forest subclassifiers, whose outputs are fused by logistic regression. AG9.1 offers a top-3 accuracy of 77.8%, with a precision of 0.51 and a recall of 0.56, at the cost of introducing a more complex model. We report top-3 accuracy rather than top-1 accuracy because, in the context of gene prioritization for validation sequencing, the cost of excluding the true causative gene is higher than the cost of evaluating a small number of candidate genes. In addition, the top-3 threshold represents a practical trade-off between high confidence in predictions and an acceptable loss of significance when selecting genes for sequence-based validation [4,5,12,13]. Both frameworks have improved locus-ranking accuracy for the 23 well-curated ADNSHL genes that account for roughly three-quarters of cases in populations of European ancestry [14]. Nonetheless, their predictions can still be difficult to interpret when class imbalance, sparse age coverage, or atypical audiogram morphologies are present.

Complementary to algorithmic advances, domain-specific visualization has been welcomed as a potentially beneficial tool for clinical use. Audioprofile surfaces (APS) plot 3D trajectories of frequency-specific threshold shift over time, revealing gene-characteristic progression patterns that are not obvious in 2D audiograms [15]. Circle-based genome views (eg, Circos [version 0.69-10; Krzywinski, Canada’s Michael Smith Genome Sciences Center] enable high-density comparison of structural variation or copy number events [16], while integrative genome browsers such as Integrative Genomics Viewer (version 2.19.7; UC San Diego and Broad Institute of MIT and Harvard) allow rapid inspection of read evidence at candidate loci [17]. More recent health care dashboards use fuzzy logic overlays and interactive filtering to expose outliers or low-confidence regions directly to end users [18]. Despite these advances, few systems combine genotype-prediction engines with audiogram-aware visual contexts; therefore, clinicians must cross-reference separate tools, a workflow that can erode trust in algorithmic suggestions and slow decision-making [19]. These limitations reflect shortcomings in how models communicate their reasoning and how results are presented to end users.

Accordingly, the literature reveals two unmet needs:

The AudioGene Translational Dashboard addresses these gaps by (1) merging the complementary strengths of AG4 and AG9.1 and (2) embedding 6 interactive visual modules—APS, 2D audioprofiles, uniform manifold approximation and projection (UMAP) cluster projections, gene count bar charts, region-of-origin pie charts, and age distribution plots—around the model output. This hybrid XAI-driven design supports clinicians in validating or questioning the algorithm’s “70/30” confidence observation and thus advances the state of practice in ADNSHL diagnostics.

The AudioGene Translational Dashboard integrates 2 ML models to increase diagnostic accuracy for ADNSHL:

The AudioGene Translational Dashboard interface provides six distinct visualization tools for health care providers and researchers to interactively assess genetic data and model predictions in real time: (1) audioprofile, a 2D plot displaying the average hearing loss (in dB) over 10 frequencies (125 Hz to 8000 Hz) for each age group, allowing comparison with a patient’s hearing loss over time (Figure 1A); (2) APS, a 3D surface plot depicting gene-specific hearing loss progression over time across frequencies, illustrating age-related changes in decibel loss (Figure 1B) [20]; (3) a region-of-origin pie chart, which displays the geographic origin distribution (eg, Dutch, German, and Chinese) for audiograms associated with each gene; (4) a count bar chart, which illustrates the audiogram count for each gene, highlighting class imbalance challenges (Figure 1C); (5) spatial analysis and clustering, which shows the cluster position of each gene and a 3D plot of audiograms (clusters are created using the k-means algorithm to partition data into 23 gene-specific groups [15]; the 3D plot compresses 11 features [age plus 10 frequencies] into 3 dimensions using the UMAP method (Figure 1D) [21]); and (2) 6) an age distribution scatter plot, which shows the age distribution for each gene within the training dataset, providing context for the predictive model outputs.

The first 3 visualization tools were developed to compare patient-specific data to average thresholds for each of the 23 ADNSHL-associated genes. This comparison allows patient audiograms to be contextualized for each gene. The audioprofile visualization shows how a patient’s audiogram compares to the expected audiograms associated with each gene.

The APS adds time as the third axis to provide a 3D rendering of gene-specific audiometric thresholds over time. Audiometric thresholds are represented as a 3D plane, depicting the expected dB loss over time (in years) at each frequency, thereby enabling comparisons between a patient’s hearing thresholds and gene-specific expectations.

The spatial analysis and clustering tool uses a bar chart to show the distribution of each prediction among 23 different clusters. These clusters are created using k-means clustering [15]. Additionally, a 3D plot visualizes the audiograms within our data that have a confirmed genetic diagnosis, compressing 11 features into 3 dimensions using UMAP [21]. This feature allows users to interact with the bar chart, highlighting corresponding clusters in the 3D plot (Figure 1). Users can compare their patient’s audiogram, represented by a large red dot, with others in the cluster, facilitating the identification of similar audiograms and associated genetic diagnoses.

By using these 3 visualizations, we aim to either enhance or reduce confidence in the predictions. For example, if the model ranks the COCH gene second among the top 3 predictions, health care providers can analyze the APS and spatial analysis tools to determine the degree of correlation with patterns typically associated with the COCH gene, potentially increasing confidence in that diagnosis.

The last 3 visualization tools provide context for the data in the AudioGene dataset. The count bar chart highlights significant class imbalance, showcasing the challenge the model encounters in predicting smaller classes due to underrepresentation. The region-of-origin pie chart and age distribution scatter plot provide additional context about data distribution, allowing health care providers to understand model limitations and adjust diagnostic strategies accordingly.

The AudioGene Translational Dashboard integrates into the workflow of a clinician as a secondary validation layer. For example, within our clinical workflow, clinicians, genetic counselors, and bioinformaticians review the results of a clinical genetic test, including patient history, family structure, audiograms, and identified variants in hearing loss genes. This team can then inspect a patient’s audiometric data relative to the landscape of audiometric data across all genes and patients, considering variance within a gene, rarity or abundance of cases within a cluster, and distance to genetically validated cases.

The system was designed using the SERN (SQL, Express.js, React.js, and Node.js) stack, which uses a client-server architecture where computationally intensive tasks are performed by the server, deployed in a Docker (version 28.5.1; Docker Inc) container [19,22]. The client is supported by React (version 18.2.0; React Foundation), a JavaScript library facilitating user interactions [23]. Data preprocessing used linear interpolation and extrapolation for missing values. The same methods for handling missing values were applied in the ML models [2,5,8].

The Pandas library in Python was used for data manipulation and analysis [8], and visualization libraries such as Plotly were used to create interactive graphs and plots [24].

For more information, please refer to the master’s thesis by DeSollar [3] and the GitLab repository.

This study was reviewed and approved by the University of Iowa Institutional Review Board (199701065). The institutional review board granted a waiver of informed consent under US federal regulation 45 CFR 46.116(f) (also known as the “Common Rule”), because, although audiograms were originally collected in clinical and research settings, the dataset used for this study was fully deidentified prior to analysis [25]. All procedures adhered to the ethical standards of the institutional and national research committees and to the 1964 Declaration of Helsinki and its later amendments. This paper does not contain any individual’s data in any form, including individual details, images, or videos. No compensation was provided to participants, as this study involved secondary analysis of a fully deidentified existing dataset and no participants were directly recruited or enrolled.

The AudioGene Translational Dashboard combines advanced ML models with several visualization tools to create a platform that facilitates the prioritization of ADNSHL-associated genes in genetic testing results. Using patient data, the AudioGene Translational Dashboard generates gene rankings and enables auditory scientists and health care providers to explore these predictions interactively through various visualization tools. Gene ranking in phenotype-genotype associations can aid in the interpretation of complex genetic data, thereby providing greater context and confidence in diagnostic decisions [5,15].

The “70/30” phenomenon serves as an indicator for health care providers, providing them with the necessary context through visualizations to assess the reliability of the predictions. When the top 3 predictions include the correct gene, health care providers can have greater confidence in proceeding with targeted genetic testing.

Several case studies have been carried out to demonstrate the effectiveness of the AudioGene Translational Dashboard in the diagnosis of specific genetic types of ADNSHL [3].

These studies demonstrate how the model can raise or lower confidence in variant interpretation based on whether the correct genetic cause of ADNSHL appears among the top 3 predicted genes. Cases 1 and 3 illustrate scenarios in which the model successfully includes the causative gene in its top predictions and closely matches the patient’s audiometric data, thereby justifying a higher level of trust in the result. In contrast, case 2 underscores how mismatched audioprofiles and inconclusive clustering can reveal when the actual gene of interest is likely missing from the top 3 predictions. The interactive visualizations of the AudioGene Translational Dashboard, such as the APS and spatial analysis tools, remain valuable in identifying gene-specific patterns that align with clinical observations [20].

However, there are important limitations of the AudioGene Translational Dashboard, especially concerning smaller gene classes. The sparsity of data and the lower accuracy of models in these categories can make the AudioGene Translational Dashboard and phenotypic predictions less reliable. However, by presenting visualizations of the data distribution and class imbalance, these limitations become more apparent, allowing data interpretation to be adjusted accordingly [2,4].

The AudioGene Translational Dashboard represents an advancement in the field of genetic diagnostics for ADNSHL. By integrating advanced ML algorithms with interactive visualization tools, the AudioGene Translational Dashboard enhances health care providers’ ability to interpret genetic data and make more informed diagnostic decisions.

A central feature of the AudioGene Translational Dashboard is the “70/30” phenomenon, which provides health care providers with critical context for confidence in genetic predictions. When the top 3 predictions are likely to contain the correct gene, the tool serves as a “green flag” for health care providers, increasing diagnostic confidence. Conversely, it alerts health care providers when predictions may be less reliable, serving as a “red flag” and prompting further investigation.

The AudioGene Translational Dashboard is an example of XAI in clinical settings, offering a context-driven method with increased transparency for the diagnosis of ADNSHL. Future developments will focus on incorporating custom model building, enhancing class imbalance functionality, and implementing user suggestions. The AudioGene Translational Dashboard not only advances genetic diagnostics for hearing loss but also serves as an example of a hybrid ML system.