Skip to main content

Young Scientist Journal

Young Scientist Journal

A Machine Learning Framework for Predicting Speech Outcomes in Cochlear Implant Recipients Using Genetic Markers

ABSTRACT

Hearing loss is a prevalent condition affecting more than 5% of the global population and remains a leading cause of speech disability. Cochlear implantation (CI) offers an effective rehabilitative solution for patients with severe sensorineural hearing loss; however, postoperative outcomes vary widely due to genetic and clinical factors. This study aimed to construct an artificial intelligence (AI)–based model to estimate CI success before the surgery using integrated genetic and demographic variables. Data were compiled from published cohorts, including WFS1, TMPRSS3, and transcriptomic expression datasets [1]. Data processing involved extensive cleaning, imputation of missing values, and standardization of variables such as age, gene type, and speech scores. Network and pathway analysis using STRING, Cytoscape and Reactome identified ten hub genes (ISG15, CXCL10, RPS19, CCL5, RPL11, among others) linked to synaptic signaling and cochlear cell resilience. Predictive modeling employed Random Forest algorithms to develop two complementary models: a classifier to predict binary CI success and a regressor to predict continuous speech scores. The classifier achieved an accuracy of 94.4%, and the regression model attained an R² value of 0.978 with minimal prediction error. Validation on unseen patient data demonstrated biologically consistent patterns, where younger age and specific gene variants were correlated with better speech outcomes. This integrative workflow highlights the potential of AI-driven predictors to personalize prognosis and guide rehabilitation strategies in cochlear implant surgeries.

INTRODUCTION.

Hearing loss is a common disorder that affects 5% of the world population [1]. Deafness and hearing loss not only impact the quality of life but also have a significant financial and social burden. Hearing loss consists of three types: sensorineural, conductive, and mixed. In conductive hearing loss, the conduction of sound waves to the inner ear is blocked or reduced. Sensorineural hearing loss (SNHL) occurs due to damage to hair cells within the inner ear, vestibulocochlear nerve, or the central processing centers of the brain [2]. Mixed hearing loss is a combination of conductive and mixed hearing loss. SNHL can present in childhood or adulthood, depending on the etiology. Hearing aids can be used as a treatment option for mild to moderate hearing loss. However, these cannot be used for severe or profound hearing loss as they rely on the presence of sufficient numbers of hair cells for stimulation of the auditory nerve [3]. In such cases, cochlear implants (CI) can be used for treatment in both children and adults.

The outcome of CI is influenced by several factors, such as age at implantation, duration of hearing loss, etiology, baseline auditory status, presence of additional disabilities, and rehabilitation intensity. CI recipients demonstrate substantial variability in speech perception post-surgery, suggesting the role of additional factors. Increasing evidence suggests the critical role of genetic factors in determining outcomes of CI [4-5]. Genetic studies have identified around 150 genes linked to non-syndromic SNHL and many genes associated with syndromic SNHL. Of these, GJB2, SLC26A4, TMPRSS3, and MYO7A are the most frequently reported [6]. One study found that a mutation in the TMPRSS3 gene, which is expressed in the spiral ganglion, was associated with poor outcomes [7].  This led to the spiral ganglion hypothesis, which states that the health of the spiral ganglion neurons affects the CI outcome. Further evidence has suggested that mutations affecting the neural component of the auditory pathway (spiral ganglion, auditory nerve, and central nervous system) negatively affect the outcome of cochlear implantation [8-9].

Various studies have identified age at implantation and duration of deafness as the key factors affecting the outcome of cochlear implantation [10-11]. A research study analyzed a comprehensive prediction model that evaluated CI outcome using known predictors and genetic testing, and found that the presence of a known genetic etiology was associated with better outcomes in infants. Moreover, a neural-type genetic variant was associated with a poor prognosis [12].

Despite these advances, genomic insights are rarely incorporated into predictive frameworks for CI outcomes. Studies completed using machine learning (ML) have analyzed genes and transcriptional profiles in age-related hearing loss [13-14]. One research group studied the effects of steroids on cochlear gene expression after noise-induced hearing loss (NIHL), suggesting pathways that may influence CI outcome [15].  A study integrated the analysis of four ARHL (age-related hearing loss)  GWAS (genome-wide association studies) datasets and found that cochlear sensory epithelial cells, not stria vascularis (specialized vascular epithelium in the cochlea that maintains the electrochemical environment in the inner ear), are the key cells for the genetic predisposition to ARHL, thus providing cell-specific insights for future modeling [16].

These studies highlight the need for integrating the genotypic information as well as the clinical and audiological profiles to generate predictive models for CI outcomes. These models can help identify candidates suitable for surgery, guide patient counselling, and choose tailored rehabilitation strategies. This study aims to build a predictive AI-based model using variants from genes to estimate CI success likelihood. By integrating genotype data with clinical outcome insights, it offers a novel approach to personalise auditory rehabilitation for maximizing speech and language outcomes in implantees.

MATERIALS AND METHODS.

Data Collection.

The data utilized in this study were compiled from three research datasets focusing on genetic and clinical correlates of auditory performance: WFS1 cohort, TMPRSS3 dataset and transcriptomic expression data. Each dataset provided insight for constructing the genotype–phenotype predictor of cochlear implant (CI) success. The WFS1 cohort dataset included detailed genotypic and phenotypic information of individuals affected by autosomal-dominant non-syndromic hearing loss linked to Wolfram syndrome type 1 gene (WFS1) variants. The dataset consisted of nucleotide and amino-acid substitutions, zygosity, and in-silico pathogenicity scores. Associated phenotype data included the age of onset, degree and configuration of hearing loss, comorbid features (diabetes, optic atrophy), and postoperative CI outcomes such as word and sentence recognition scores. The TMPRSS3 dataset comprised 127 participants with TMPRSS3-related deafness (DFNB8/10). The dataset integrated demographic variables (age, sex, ethnicity), genotypes categorized as missense, loss-of-function, or compound heterozygotes, and auditory phenotypes (severity, slope, and audiogram configuration). Device-related data included the type of prosthesis (hearing aid versus cochlear implant), laterality, age at implantation, and aided speech perception scores were also extracted. The transcriptomic expression dataset provided transcriptomic profiles from cochlear tissues comparing aged and control groups of mice. 18 transcriptional clusters representing 10 major cochlear cell types were identified. Fifty-six differentially expressed genes (DEGs) were screened, and immunohistochemical verification was done. KEGG pathway enrichment analysis identified key pathways for auditory gene networks.

Data processing.

The dataset was cleaned, and values were imputed using a script to find the missing values. The cleaning of the dataset involved standardizing it and ensuring all textual inconsistencies were eliminated. In some places, numeric conversions were applied where required. For example, age fields were converted from text ranges such as “20–39” to numeric averages. Apart from cleaning, missing data across several variables in the cochlear implant patient data set were imputed using the script. The imputation strategy of each variable was according to its nature. Variables such as age and speech scores were imputed using gene and data sources. On the other hand, gene and implant type values were imputed using modal values from the cohort.

Network and Pathway analysis.

The network and pathway analysis were performed using the final gene list. A gene- interaction network was constructed using the STRING database for the finalized gene list and was visualised using Cytoscape software. Topological parameters, including degree, betweenness centrality, closeness centrality, and clustering coefficient were computed using Cytoscape’s Network Analyzer tool. Based on topological parameters, the top 10 genes with the highest number of interactions were identified as the major hubs within the network. They exhibited strong interconnectivity, suggesting their significance in the network. Pathway analysis of the identified genes was performed using the Reactome database. The finalised gene list was analysed under default parameters, and an interaction network was generated. The top pathways were identified in this way. Furthermore, the top 10 genes identified were cross-referenced with the top 25 pathways. Genes found to be present in the top 25 pathways were identified.

Model Building and Analysis.

The modeling process aimed to predict outcomes in cochlear implant patients post-implantation using genetic and clinical data. This was done by building two predictive models using the Random Forest algorithm. A classifier was trained to predict the success or failure of cochlear implantation in patients, and a regressor to predict the numerical speech score. A Python script was used to build these predictive models. Apart from achieving this, the script was also used to load and clean the dataset containing the raw genetic and clinical data, and to numerically encode the gene data. The dataset was cleaned to ensure the entries for age and speech score were valid, and to exclude records with missing variables. A ‘success label’ was also created, where positive scores indicated successful outcomes after implantation. After creating the two predictive models, they were also evaluated and applied to new patients for prediction and validation. The binary outcome (predicted success) and continuous outcome (predicted speech scores) were generated for each patient to test the efficiency of the models.

RESULTS.

Dataset characteristics and imputation outcomes.

During data processing, it was found that age was the category of data that was 100% imputed, with an imputed count of 162. This was because all the age fields were converted to numeric averages to make it more standardised and comprehensive. The age estimates were produced realistically by aligning them with gene and cohort patterns. Genetic information was imputed for nearly 69% of the values in the dataset, significantly making it more reliable. The most frequent gene type within each cohort was used to impute the values. About 40% of the implant type data was imputed. Values were imputed using cohort-level modal values to make the dataset complete.

Identification of Hub genes.

Using the STRING (Search Tool for the Retrieval of Interacting Genes/Proteins) database, the seed genes were analzsed, including their interaction with other genes in the body, as shown in Fig. 1. A network of 36 nodes and 75 edges was visualized using the Cytoscape software.  Each circle or node represents a protein encoded by the genes in the dataset with different colors indicating clusters of functionally related proteins. An edge (or line) is a connection between two nodes representing a biological or functional association between the corresponding proteins.

Figure 1. A gene interaction network with 36 nodes and 75 edges generated using the STRING database. Each circle or node represents a protein encoded by the genes in the dataset with different colors indicating clusters of functionally related proteins. Lines represent protein-protein interactions.

After computing topological parameters, hub genes were identified on the basis of 4 topological parameters. The top 10 genes are listed in Table 1. The 4 chosen parameters were degree, betweenness centrality, closeness centrality, and clustering coefficient. Degree refers to the number of edges connected to a node, indicating the node’s connectivity. Betweenness centrality measures how frequently a node lies on the shortest path between other pairs of nodes in a network. Closeness centrality is a parameter that measures how efficiently a node can reach all the other nodes in a network, based on the distances of the shortest paths. The clustering coefficient measures local cohesiveness in a network by quantifying how connected a node’s neighbors are to each other. Thus, these four parameters establish a strong basis for identifying the hub genes of a biological network.

Table 1. The top 10 hub genes along with their topological parameters, including degree, betweenness centrality, closeness centrality, and clustering coefficient.
S. No. Name Degree Betweenness centrality Closeness centrality
1 ISG15 10 0.4232804233 0.4516129032
2 CXCL10 8 0.4545855379 0.5283018868
3 RPS19 8 0.07142857143 0.3456790123
4 CCL5 7 0.200617284 0.4912280702
5 RPS15A 7 0 0.3414634146
6 RPS26 7 0 0.3414634146
7 RPSA 7 0 0.3414634146
8 RPL32 7 0 0.3414634146
9 RPLP0 7 0 0.3414634146
10 RPL11 7 0 0.3414634146

After performing pathway analysis through Reactome database, the top 5 over represented pathways identified were Influenza Infection (R-HSA-168255), Peptide chain elongation (R-HSA-156902), PELO: HBS1L and ABCE1 dissociate a ribosome on a non-stop mRNA (R-HSA-9954714), Nonsense Mediated Decay (NMD) independent of the Exon Junction Complex (EJC) (R-HSA-975956), and Eukaryotic Translation Elongation (R-HSA-156842). Amongst the top 10 hub genes, ISG15, RPS19, RPS15A, RPS26, RPSA, RPL32, RPLP0, and RPL11 were found to be present in the top 25 pathways.

Predictive performance of the Random Forest Models.

Both classification and regression models were successfully built and executed using Random Forest algorithms. The classifier model aimed to predict the success of cochlear implantation, determined by whether the speech score was more than 0. The dataset was split into a ratio of 80:20 for training and testing of the model. The Random Forest classifier model achieved a high accuracy of 94.4%, with a precision of 100% for class 0 (unsuccessful outcome) and 93% for class 1 (successful outcome). There was a slightly lower recall percentage for Class 0, though: 75% as compared to 100% for Class 1. The Random Forest regression model predicted continuous speech score values based on gene and age. The regressor model achieved a high R2 value of 0.978 (97.8%) and a minimal mean squared error of 0.0049.

Validation on unseen patient data.

For validation of the models, 13 unseen patient samples were used, and predicted success as well as predicted speech score were generated. The model’s predicted scores suggested constant variance with age, supporting the model’s ability to predict based on unseen patient data. The majority of the data aligned with preestablished biological patterns, where certain genes and younger age are associated with higher predicted speech scores.

DISCUSSION.

This study incorporated data preprocessing, model development, and validation to predict cochlear implant outcomes in patients. Data cleaning and imputation were carried out to impute missing values in the dataset and standardize it. Gene and age were selected as the key predictive features for the models for their biological significance in influencing speech and auditory performance. Two models were built: A Random Forest Classifier and a Random Forest Regressor to predict binary speech success and continuous speech scores, respectively. These allowed both qualitative and quantitative interpretation of outcomes of cochlear implantation. The classifier showed an accuracy of 94.4% indicating a strong performance in predicting speech outcomes after implantation. The regressor achieved an R2 score of 0.978 and an MSE of 0.0049, suggesting a near-perfect fit and very low prediction error. The high R2 value suggests that there was a strong correlation between gene type, patient age, and speech performance. The models captured significant nonlinear relationships between genetic factors, age, and post-implant speech development. The Random Forest models were chosen due to their reliability and ability to manage heterogeneous data (categorical and numerical). The production of both classification and regression models allowed for complementary insights into the prediction of the success of cochlear implants in patients, with one producing categorical predictions of success and the other quantifying expected speech scores. Hence, the models justify the methodological design and support the reliability of the workflow.

The predictive patterns observed in this study are consistent with the genotype-phenotype reported in previous studies. Favorable auditory outcomes have been reported with variants like CDH23 (hair cells of the cochlea), SLC26A4 (outer sulcus cells), GJB2 (gap junction protein in hair cells in the cochlea), and MYO15 (stereocilia), as seen in this study [3]. However, the variant MYO7A (stereocilia) in the cochlea had a negative predictive outcome that was in contrast to the study done by a research groupl, who reported improved CI results in patients of Usher syndrome with this variant [17]. However, in this study, it has been found that age is an important prognostic factor in CI performance, with early CI showing positive speech outcomes. Our work supports this finding, as patients with early age at implantation are predicted to have better speech outcomes, reinforcing the view that the success of CI is linked to the age at which the patient undergoes surgery, with early implantation showing better outcomes.

The present study demonstrated strong genotype-phenotype prediction using Random Forest models. However, there were a few limitations. The dataset used is moderate in size, which can affect the generalizability of results across various population groups. Variants such as EYA1 and MYO7A were underrepresented, which can affect the accuracy of the models for the prediction of outcomes for rare etiologies. Moreover, other important variables like duration of deafness, residual hearing, associated morbidities, duration of rehabilitation, and speech therapy were not considered. Future research could further refine the model precision by expanding the dataset, including infrequent mutations as well as multi-model predictors like imaging and auditory training metrics. Longitudinal modelling can be devised to study the course of post-implant effects rather than static outcomes. This can help in the dynamic understanding of auditory adaptation.

CONCLUSION.

This prediction model offers a precise audiological tool that can be used for personalized preoperative counseling of parents regarding speech and language outcomes post cochlear implantation. With this model, the families of children undergoing implantation can be mentally prepared and prognosticated about the speech and language outcome of their child, and a better strategy can be devised for superior outcomes. This can also help in prioritizing early intervention for patients who are likely to benefit from the surgery. Overall, the AI prediction model bridges the molecular pathways with functional rehabilitation and speech outcomes by integrating genomic and clinical datasets.

ACKNOWLEDGMENTS.

I would like to acknowledge Aashna Saraf, for providing valuable feedback and guidance throughout the project. I would also like to acknowledge Sankalp ENT & Cochlear Implant Centre, Delhi, India, for providing the data for validation.

REFERENCES

  1. Sappington, J.M., Hearing Loss: The Silent Pandemic That is Finally Being Heard. Mo Med, 121(5), 350-354 (2024).
  2. Cunningham, L.L. and D.L. Tucci, Hearing Loss in Adults. N Engl J Med, 377(25), 2465-2473 (2017).
  3. Eshraghi, A.A., et al., Genotype-Phenotype Correlation for Predicting Cochlear Implant Outcome: Current Challenges and Opportunities. Front Genet, 11, 678 (2020).
  4. Carlson, R.J., et al., Association of Genetic Diagnoses for Childhood-Onset Hearing Loss With Cochlear Implant Outcomes. JAMA Otolaryngol Head Neck Surg, 149(3), 212-222 (2023).
  5. Fehrmann, M.L.A., et al., Cochlear Implantation Outcomes in Genotyped Subjects with Sensorineural Hearing Loss. J Assoc Res Otolaryngol, 26(3), 331-348 (2025).
  6. Nishio, S.Y. and S.I. Usami, Outcomes of cochlear implantation for the patients with specific genetic etiologies: a systematic literature review. Acta Otolaryngol, 137(7), 730-742 (2017).
  7. Eppsteiner, R.W., et al., Prediction of cochlear implant performance by genetic mutation: the spiral ganglion hypothesis. Hear Res, 292(1-2), 51-8 (2012).
  8. Shearer, A.E., et al., Genetic variants in the peripheral auditory system significantly affect adult cochlear implant performance. Hear Res, 348, 138-142 (2017).
  9. Tropitzsch, A., et al., Variability in Cochlear Implantation Outcomes in a Large German Cohort With a Genetic Etiology of Hearing Loss. Ear Hear, 44(6),1464-1484 (2023).
  10. Wu, C.M., et al., Long-Term Cochlear Implant Outcomes in Children with GJB2 and SLC26A4 Mutations. PLoS One, 10(9),e0138575 (2015).
  11. Lee, S.Y., et al., The molecular etiology of deafness and auditory performance in the postlingually deafened cochlear implantees. Sci Rep, 10(1), 5768 (2020).
  12. Han, J.H., et al., Comprehensive Prediction Model, Including Genetic Testing, for the Outcomes of Cochlear Implantation. Ear Hear, 44(1), 223-231 (2023).
  13. Liu, W., et al., Bioinformatics and experiments reveal the hub genes of age-related hearing loss and the mechanism of PLK1 silencing in the protection of aging cochlear hair cells. Gene, 963, 149632 (2025).
  14. Lee, S.J., et al., Structural analysis of pathogenic TMPRSS3 variants and their cochlear implantation outcomes of sensorineural hearing loss. Gene, 865, 147335 (2023).
  15. Maeda, Y., et al., Cochlear Transcriptome Following Acoustic Trauma and Dexamethasone Administration Identified by a Combination of RNA-seq and DNA Microarray. Otol Neurotol, 38(7), 1032-1042 (2017).
  16. Eshel, M., et al., The cells of the sensory epithelium, and not the stria vascularis, are the main cochlear cells related to the genetic pathogenesis of age-related hearing loss. Am J Hum Genet, 111(3),614-617 (2024).
  17. Usami, S.I., et al., Cochlear Implantation From the Perspective of Genetic Background. Anat Rec (Hoboken), 303(3), 563-593 (2020).

Posted by on Friday, May 15, 2026 in May 2026.

Tags: , , , ,