A method for constructing a presbycusis risk prediction model based on human peripheral blood transcriptomic data

Abstract

This patent application discloses a method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data. It belongs to the technical field of using computer technology for risk prediction of age-related hearing loss, aiming to improve the accuracy of such predictions. The method employs a stratified screening strategy to select core transcripts based on aging and the degree of hearing loss. Then, a model is constructed: the selected core transcriptomes are used as features, with hearing status as a label, and a machine learning prediction model is built on a computer using a training set. The core transcriptomes selected by this method are more representative of the molecular events related to the pathogenesis of age-related hearing loss, improving the model's biological relevance and specificity.

Description

Technical Field

[0001] This invention relates to a method for constructing a model for predicting the risk of age-related hearing loss using computer technology. Background Technology

[0002] The pathogenesis of age-related hearing loss (ADR) is complex, involving systemic transcriptomic disorders. Transcriptomics provides a panoramic view of changes in the transcriptome of genes within an organism, offering a powerful tool for elucidating disease mechanisms and discovering biomarkers. However, due to the specific and difficult-to-obtain nature of human cochlear samples, current transcriptome sequencing primarily focuses on mouse cochlear sequencing, lacking human samples. Considering that ADR is a disease based on systemic aging changes, with the cochlea as a target organ of aging, we collected peripheral blood from a population cohort as our research sample. Traditional transcriptomic analyses often focus on identifying single or a few differentially expressed genes, making it difficult to comprehensively capture the complex transcriptomic network changes associated with the disease. Furthermore, transcriptomic data is characterized by high dimensionality and small sample sizes, limiting the application of traditional statistical methods in processing such data and constructing robust predictive models. Currently, a systematic method is lacking to effectively screen core transcripts related to the disease itself (rather than simply aging) from massive amounts of transcriptomic data, and to use these transcripts to construct a high-precision, interpretable predictive model for ADR risk.

[0003] Any discussion of background art throughout this specification does not imply that the background art is necessarily prior art known to a person skilled in the art; nor does the discussion of prior art throughout this specification imply that the prior art is necessarily widely known. Summary of the Invention

[0004] The present invention aims to provide a method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data, so as to improve the accuracy of risk prediction for age-related hearing loss.

[0005] The method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data in this scheme includes the following steps: S1 Sample Collection and Data Acquisition: Peripheral blood samples were collected from the study population for whole transcriptome analysis to obtain raw transcriptome gene expression profile data. At the same time, the audiological assessment results and age information of the subjects were collected to form a training set. S2 core transcript screening: Core transcripts were screened using a stratified screening strategy based on the degree of aging and hearing loss. S3 Model Construction: Using the selected core transcriptome as features and hearing status as labels, a machine learning prediction model is constructed on a computer using the training set. S4 Model Validation: The built model is evaluated for performance using an independent internal validation set and an external validation queue. Metrics include AUC, accuracy, sensitivity, and specificity to confirm core features.

[0006] Furthermore, it also includes step S5 model application: the prediction module storing the machine learning prediction model is deployed on the server side. The prediction module is used to receive the detection value of the core gene input by the user, and then generate the risk probability of the user suffering from age-related deafness as the prediction result.

[0007] Furthermore, it also includes a front-end interaction module for communication with the prediction module, which is used to input the detection values ​​of the user's core genes and output the prediction results.

[0008] Furthermore, the front-end interaction module calls the SHAP interpreter based on the user's input data, ultimately generating a SHAP force map. The SHAP force map can be dynamically generated in the form of a bar chart or waterfall chart, intuitively displaying the direction and magnitude of the contribution of each input metabolite to the final prediction result, explaining the sources of disease risk.

[0009] Furthermore, the screening criteria in S2 include: the study population includes a group of patients with age-related hearing loss and a healthy control group; significant differentially expressed genes are screened by comparing the differences in mRNA expression between patients with age-related hearing loss and the healthy control group; and differentially expressed genes that are significantly associated with hearing loss are screened by correlation analysis with the pure tone mean hearing threshold.

[0010] Furthermore, the screening criteria in S2 include: screening for gene modules significantly associated with age-related hearing loss through network-weighted co-expression, taking the intersection of differentially expressed genes significantly associated with hearing level with genes in the age-related hearing loss module, and selecting the core transcriptome set that is associated with the age-related hearing loss disease state itself.

[0011] Furthermore, S3 employs a variety of machine learning algorithms, including Lasso regression, gradient boosting linear model glmBoost, random forest, and NaiveBayes. The algorithm with the best prediction performance can be selected.

[0012] The beneficial effects of this invention are: 1. Enhanced disease specificity: Through a pioneering “stratified screening” strategy that takes into account both aging and the degree of hearing loss, the core transcriptome selected is more representative of the molecular events related to the pathogenesis of age-related hearing loss, thus improving the model’s biological relevance and specificity.

[0013] 2. Superior predictive performance: Machine learning models built based on the selected core transcriptome, such as the glmBoost+NaiveBayes model, have demonstrated good predictive accuracy and generalization ability in both internal and external validation, outperforming traditional methods.

[0014] 3. Robust and Interpretable Model: Utilizing a well-established golden process from "feature selection" to "stable prediction," it excels at handling the challenges unique to genomic data, such as high dimensionality, small sample sizes, and complex interactions. Combined with SHAP interpretability analysis, it clearly reveals the contribution of each core transcriptome to the prediction results, providing new clues for exploring the transcriptomic mechanisms of age-related hearing loss.

[0015] 4. Clinical application of cutting-edge technologies: Transforming complex metabolomics and machine learning research findings into online tools accessible to everyone, greatly promoting the application of cutting-edge technologies in clinical practice and health management.

[0016] 5. Precision and Personalization: Predictions are based on an individual's unique metabolic profile, enabling truly personalized risk assessment. The 14 core genes showcased can more accurately capture disease-related metabolic characteristics.

[0017] 6. In-depth visualization and interpretability: It not only outputs risk probabilities but also uses SHAP (Shape-Based Analysis) to provide in-depth interpretation of the "black box" model. Users can intuitively see which genes drive their risk levels, providing unprecedented information for exploring individual etiologies and developing precise intervention strategies.

[0018] This patent also provides an application of a combination of gene markers in constructing a risk prediction model for age-related hearing loss, the combination of gene markers including the following genes: SFRP5, SAXO1, GOLGA6L4, NCAPH, MED31, P2RX4, CYP2J2, ZSCAN23, S1PR3, MTHFR, DCBLD2, TNFAIP8L3, NAB2 and SLC7A3.

[0019] These genes are core genes identified based on the aforementioned age-related hearing loss risk prediction model. This combination exhibits excellent discriminative ability, providing a reliable tool for early clinical screening. Furthermore, this gene combination is derived from the peripheral blood transcriptome and can be obtained non-invasively through routine blood collection, enabling large-scale population screening.

[0020] The HGNC IDs corresponding to the genes are as follows: SFRP5: HGNC:10779 SAXO1: HGNC:28566 GOLGA6L4: HGNC:27256 NCAPH: HGNC:1112 MED31: HGNC:24260 P2RX4: HGNC: 8535 CYP2J2:HGNC:2634 ZSCAN23: HGNC:28097 S1PR3: HGNC:3166 MTHFR:HGNC:7436 DCBLD2: HGNC:24626 TNFAIP8L3: HGNC:26095 NAB2:HGNC:7363 SLC7A3: HGNC:11062. Attached Figure Description

[0021] Figure 1 A comparison diagram of the expression of 14 core genes in different groups in this embodiment of the invention.

[0022] Figure 2 Performance ranking chart of diagnostic models.

[0023] Figure 3 ROC curve and confusion matrix of the .glmBoost+NaiveBayes model.

[0024] Figure 4 Interactive web page interface.

[0025] Figure 5 Heatmap of genes with significant differential expression in the transcriptome between the Control and ARHL groups.

[0026] Figure 6 Volcano plot of differentially expressed genes: green indicates significant downregulation, and red indicates significant upregulation.

[0027] Figure 7 Clustering trees of samples with gene modules highly associated with ARHL were identified using WGCNA.

[0028] Figure 8 Gene co-expression heatmap of different samples.

[0029] Figure 9 The topology analysis evaluation diagram of soft threshold power is displayed.

[0030] Figure 10 Clustering dendrogram of gene modules and module merging.

[0031] Figure 11 A comparison chart of the correlation between the module and ARHL incidence.

[0032] Figure 12 A scatter plot of gene-related data belonging to the grey module.

[0033] Figure 13A comparison chart of machine learning results for different model configurations in the Training and Test groups.

[0034] Figure 14 ROC analysis plot of the Training group in the .glmBoost+ NaiveBayes model.

[0035] Figure 15 Confusion matrix model diagram of the Training group.

[0036] Figure 16 A heatmap showing the correlation between 14 genes and PTA.

[0037] Figure 17 Independent AUC values ​​of genes.

[0038] Figure 18 ROC analysis plot of the Test group in the .glmBoost+ NaiveBayes model.

[0039] Figure 19 Confusion matrix model diagram for the Test group. Detailed Implementation

[0040] The present invention will be further described below with reference to specific embodiments.

[0041] The examples used peripheral blood samples and hearing data from 200 community members.

[0042] 1. Data Acquisition: Peripheral blood was collected from all participants for whole transcriptome sequencing to obtain raw whole transcriptome data. Pure-tone audiometry results were also obtained. Participants aged 60 and above with a mean hearing threshold ≥35 dB were defined as the presbycusis (ARHL) group, while age-matched participants with a hearing threshold <35 dB served as the control group.

[0043] 2. Core Transcript Screening: A stratified screening strategy was employed to exclude the influence of aging alone on transcriptome genes. The screening criteria were as follows: Genes showing significant differential expression were identified by comparing mRNA expression differences between patients with age-related hearing loss and healthy controls (e.g., P < 0.05 and logFC > 0.585, i.e., a 1.5-fold difference); correlation analysis with pure-tone mean hearing thresholds was used to identify differentially expressed genes significantly associated with hearing loss. Network-weighted co-expression (WGCNA) was used to screen for gene modules significantly associated with age-related hearing loss. The intersection of differentially expressed genes significantly associated with hearing loss levels and genes in the age-related hearing loss module was used to select the core transcriptome set associated with the age-related hearing loss disease state itself.

[0044] Figure 1 This is a graph comparing the expression of the 14 core genes selected in different groups, which verifies the effectiveness of the selection strategy.

[0045] The types shown in the diagram on the right include: Yellow: ARHL (Age-Related Hearing Loss) Blue-green: Control group The projects are grouped from top to bottom as follows: Pale purple: Mild Presbycusis (mild age-related hearing loss) Pink: Severe Presbycusis (severe presbycusis) Bright pink: MiddleControl (Middle-aged comparison) Green: OldControl (for the elderly) Sky blue: YoungControl (Youth Comparison) The right side of the figure shows 14 core genes from top to bottom, arranged as follows (corresponding to the vertical axis labels): SLC7A3, MED31, SAXO1, ZSCAN23, DCBLD2, CYP2J2, SFRP5, NAB2, S1PR3, P2RX4, MTHFR, TNFAIP8L3, NCAPH, GOLGA6L4.

[0046] As shown in the figure, all 14 genes exhibited statistically significant differences in expression among at least one group (as indicated by color contrast intensity); the expression patterns were highly consistent with the clinical phenotypes (mild / severe hearing loss, age gradient).

[0047] 3. Model Construction: 14 machine learning methods were used—Selection Operator (Lasso), Regression, Stepglm [both], Survival-Support Vector Machines (SVM), glmBoost, plsRglm, Stepglm [forward], RF, Stepglm [backward], LDA, and NaiveBayes. Lasso Regression: Selects key features through L1 regularization to avoid overfitting.

[0048] Stepglm[both]: Uses a two-way stepwise selection method (forward and backward) to select the optimal feature subset.

[0049] Stepglm [forward]: Starts with an empty model and gradually adds features that most significantly improve model performance until a stopping condition is met (such as minimizing AIC or the maximum number of features).

[0050] Stepglm [backward]: Starting with the complete model containing all features, it gradually removes features that contribute the least to the model's performance until a stopping condition is met.

[0051] SVM: Used to process survival data and optimize survival time prediction.

[0052] glmBoost: Gradient boosting model, which improves prediction performance by iteratively optimizing the decision tree.

[0053] plsRglm: Suitable for high-dimensional data, it extracts latent variables through dimensionality reduction.

[0054] RF: Based on ensemble learning, it improves prediction accuracy by voting among multiple decision trees.

[0055] LDA: Used for classification tasks to maximize the separation between classes.

[0056] NaiveBayes: Based on Bayes' theorem, it assumes that features are conditionally independent.

[0057] The generated models were further refined by combining the different modeling methods described above. Finally, we constructed 41 diagnostic models for ARHL. Using a linear combination function of the machine learning model, we calculated the risk score and sample classification (whether or not the sample has ARHL) for each sample based on gene expression levels, and calculated the AUC values ​​of the generated models in the Train and Test groups. The models were ranked according to the average AUC value, and the optimal model - glmBoost+NaiveBayes - was selected. ROC curves and AUC values ​​were used to evaluate the effectiveness and accuracy of the models built based on the selected genes of glmBoost+NaiveBayes. Confusion was used to predict whether the samples in the Train and Test groups had the disease, and recall, precision, and F1 (harmonic mean of precision and recall) were calculated to evaluate the diagnostic performance of the ARHL feature gene set.

[0058] Figure 2 The constructed glmBoost+NaiveBayes model ranked first among all models; Figure 3 ROC curves and confusion matrix of the glmBoost+NaiveBayes model in internal and external validation.

[0059] 4. Interpretability analysis: SHAP analysis (SHapley Additive exPlanations) was used to quantify the contribution of each feature to a single prediction result or the overall model output. The results showed that all 14 genes contributed to the model and revealed the nonlinear dose-dependent relationship of key genes such as S1PR3, providing direction for further mechanism research.

[0060] The present invention will be further described below with reference to specific embodiments. The prediction model used in the embodiments is the glmBoost+NaiveBayes model constructed according to "A method for constructing a prediction model for age-related hearing loss based on human peripheral blood transcriptomics data".

[0061] 1. Tool Development: Develop an interactive web application based on the Shiny framework in R. The front-end design includes a form with 14 input fields, each corresponding to a core gene (such as S1PR3, NAB2, P2RX4, etc.).

[0062] 2. Backend Integration: Save the trained glmBoost+NaiveBayes model (containing coefficients for 14 genes) and other necessary model parameters as an RDS file and integrate it into the server-side logic of the Shiny application. The server-side code is responsible for receiving input, performing necessary transformations (if needed), calling the model to calculate the predicted probabilities, and calculating the SHAP value corresponding to each input (using R packages such as shapviz).

[0063] 3. Deployment and Application: Deploy the developed Shiny application to the cloud server and generate a URL. Figure 4 This is a screenshot of the web interface according to an embodiment of the present invention. After obtaining the user's genetic testing results, the user (such as a researcher or clinician) visits this website and enters the quantitative values ​​of 14 metabolites one by one in the left panel. After clicking "Submit", the right panel will immediately display the probability of age-related hearing loss based on the metabolite profile and generate a color SHAP waterfall chart, explaining in detail how each metabolite promotes or inhibits the final risk score.

[0064] Example 2: 1. Screening Strategy A cohort of 161 samples from the first community was used as the internal group. Differential expression analysis (logFC = 0.585, p < 0.05) and correlation analysis with pure tone mean hearing threshold (PTA) (Spearman r ≥ ± 0.2, p < 0.05) identified 4874 differentially expressed genes associated with PTA for use in WGCNA (Figrue 2a-b, Supplementary File 1). Highly correlated gene modules were identified using WGCNA, and their gene sets were extracted. Correlation analysis with PTA (Spearman r ≥ ± 0.4, p < 0.05) identified a core gene set for constructing genetic features to predict age-related hearing loss (ARHL). A cohort of 39 samples from the second community was used as the external cohort to independently validate the constructed genetic features.

[0065] Figure 5 The heatmaps of differentially expressed genes in the transcriptomes of the Control and ARHL groups show visualizations of 30 differentially expressed genes, which are further subdivided into five subgroups: Severe Presbycusis, Mild Presbycusis, Old Control, Middle Control, and Young Control.

[0066] The types shown in the diagram on the right include: Yellow: ARHL (Age-Related Hearing Loss) Blue-green: Control group The projects are grouped from top to bottom as follows: Light purple: 1YoungControl (Youth Control) Pink: 2MiddleControl (Middle-aged comparison) Dark pink: 3OldControl (elderly control) Bright blue: 4 Mild Presbycusis (mild hearing loss) Dark cyan / blue-green: 5 Severe Presbycusis (severe hearing loss) The right side of the image lists 30 gene names from top to bottom, arranged as follows (corresponding to the vertical axis labels): FOXJ1, ZNF69, ITC9, SALL2, ABLIM1, LEF1, ITGA6, METAP1D, EPHA1, SLCTA6, DNMT3A, NPAS2, GAL3ST4, MMP28, ZDHHC9, CYP2J2, NTSE, SFRPS, CACHD1, CD248, NRCAM, SIRT4, WAP9, ZNF880, ATF7IP2, C5orf63, PMIP2B, DPH5. NUCB2, RFESD, SLC20A1, METRNL, CASZ1, PDE1B, ZNF335, NCAPH, ARHGAP11B, MCOLNI, FITM1O, SMURF1, JLK1, GABARAPL1, DENND3, PI PRE, TBK1, ZEB2, PLSCR2, ACOT9, DUSP4, ZC3H10, DYNC1LI2, COPA, S1PR3, SMAD6, IPCN2, SH3BP2, INFRSF1B, AGAP3, RUSC2 and HOMER3.

[0067] Figure 6 Volcano plot of differentially expressed genes: green indicates significant downregulation, and red indicates significant upregulation.

[0068] 2. WCGNA identifies modules closely related to ARHL (age-related hearing loss). First, one sample was identified as an outlier and excluded based on the sample clustering tree (Figrue2a). The gene co-expression heatmap of different samples is shown in Figrue3b. A soft threshold power of 14 was set (Figrue2c), and eight gene modules were identified based on the gene dendrogram: gray, green, brown, blue, red, turquoise, black, and yellow modules (Figrue2d). The disease correlation heatmap of different modules showed that the grey module had the strongest importance (Figrue2e). The MM-GS coefficient between module membership and gene significance was 0.466, and the correlation coefficient with disease was 0.53, p < 0.001 (Figrue2f). Based on correlation analysis with PTA (Spearman r ≥ ±0.4, p < 0.05), 72 core genes belonging to the grey module were selected for further analysis (Supplementary Document 2).

[0069] Figure 7Gene modules highly associated with ARHL were identified using WGCNA. A sample clustering tree was constructed, and one outlier sample was removed. Cluster analysis was performed on multiple biological samples to assess the similarity of gene expression profiles among samples and identify potential outliers. All samples were grouped as follows: Young Control, Middle Control, Old Control, Mild Presbycusis, and Severe Presbycusis.

[0070] The clustering dendrogram shows that samples within the same group generally cluster together, indicating a high degree of consistency in gene expression patterns within the group; samples of mild and severe age-related hearing loss show a partial separation trend in the branches, reflecting molecular differences related to disease severity; the Euclidean distance of all samples in the figure is less than 250,000 units (based on the default scale of the clustering algorithm).

[0071] Figure 8 This is a heatmap of gene co-expression from different samples.

[0072] The vertical axis on the right side of the diagram, from top to bottom: TRIM44, PRKACB, APPBP2, DDX1, GPR171, VPS33A, FAM217B, BCL2L13, SOD1, FASTKD1, LPIN2, SEC62, KLHL42, DET1, ZNF260, TMEM42, BRD3OS, PLEKHB1, PRAG1, ATM, UXS1, RPS18, NOC3L, CCL28, AMY2B, MRPL34, RPL10, TMEM143, RPP21, HDAC9, GDPGP1, STPG1, CCDC15, ATXN7L3B, EIF4E, TYW1B, ZNF507, P OLR3G, OCM, AK5, PLAG1, CLDN15, ADSL, CDK20, POPDC2, C16orf74, GPRC5B, ADD2, SOAT2, H4C16, LRRC46, ASCL5, MAPRE3, SUGP2, SMARCB1, PLXNB1, CIAO3, GP R137C, RIN2, FAM110B, TNNT1, DYNLT3, DNAJA1, UBXN11, STX4, MYO1F, DYNC1LI2, CEMIP2, PSD2, NTRK2, AKNAD1, WDFY1, MAD2L1BP, CLCF1, PLIN3, RLF, TOM1.

[0073] The bottom horizontal axis in the diagram, from left to right: TOM1, DUSP8, CEBPB, PNCK, USP36, TMEM184B, FEM1C, DRAM1, ZBTB46, SEMA5B, COL12A1, ARFGAP1, TMEM104, CEMIP2, RBM23, CERS5, RXRA, HOMER3, ADAP1, KATNIP, IL17RC, TIPIN, PPP1R2, DMXL2, MTHFD2, PLE KHA3, RIN2, EIF4EBP1, ARGFX, MAN1B1, GIPR, TSEN54, DHX37, DDB2, BAZ1A, ANKRD54, RSPH6A, MUC4, IGSF11, H 4C16, ANKDD1B, KCNG1, CD22, DNASE1L3, AMN, KCNC4, CYP26C1, SPATC1L, PGAP2, CD3E, MRNIP, FAM222B, PLAG1 , RUNX2, PKIA, OXNAD1, KAT6B, NAA30, MIX23, FAM228B, CPLANE1, SCAF4, TSPYL4, ZNF491, ZNF175, STPG1, GOT 1. B4GAT1, LBR, CHCHD6, UFSP1, SHPK, RPL8, DCTPP1, CCT3, HMOX2, NEMP2, SPIN3, NOC3L, SCMH1, RPL14, PAIP2 B. PPM1K, AK9, APEX1, TMEM69, NAA20, ERG28, PIGW, NIF3L1, GPR171, CEP350, PCNX4, NCL, AASDH, DENR, IARS2, KLHDC10.

[0074] Figure 9 The content of the left and right parts in the middle determines the soft threshold power to be 14.

[0075] Figure 10 It is a clustering dendrogram of gene modules and module merging.

[0076] Figure 11 The correlation between the visible module and ARHL disease is shown.

[0077] Figure 12 This is a scatter plot of gene-related data belonging to the grey module, showing the correlation between genes and disease phenotypes.

[0078] 3. Construct and validate genetic features for predicting ARHL using machine learning. Ultimately, the gene signature constructed using glmBoost+ NaiveBayes had the optimal AUC (area under the curve).

[0079] Figure 13 The machine learning results for different model configurations in the Training and Test groups are shown. The glmBoost+NaiveBayes model has the best average AUC of 0.861.

[0080] Figure 14 It is the ROC of the Training group in the glmBoost+ NaiveBayes model.

[0081] The study included four groups: Train (AUC = 0.906), Group 1 (AUC = 0.840), Group 2 (AUC = 0.840), and Group 3 (AUC = 0.859) (Figrue3a-b, Figrue3e). Genetic characteristics included 14 genes (SFRP5, SAXO1, GOLGA6L4, NCAPH, MED31, P2RX4, CYP2J2, ZSCAN23, S1PR3, MTHFR, DCBLD2, TNFAIP8L3, NAB2, SLC7A3). Correlation analysis with hearing threshold (PTA) showed that 7 genes were significantly positively correlated with PTA (Spearman r = 0.41–0.46, p < 0.001), and 7 genes were significantly negatively correlated with PTA (Spearman r = -0.6–-0.41, p < 0.001) (Figrue3c).

[0082] Figure 16 This is a heatmap showing the correlation between 14 genes and PTA.

[0083] The accuracy of ARHL prediction was evaluated using a confusion matrix. Within the internal groups, the accuracy was 85% (Train=82.2%, Group1=75%), precision was 81.8% (Train=79.3%, Group1=83.3%), recall was 90% (Train=92%, Group1=80%), and F1 (harmonic mean of precision and recall) was 85.7% (Train=80.7%, Group1=78.9%) (Figrue3d). In the external independent validation cohort, the accuracy was 76.9% (Group3), precision was 92% (Group3), recall was 76.9% (Group3), and F1 (harmonic mean of precision and recall) was 83.8% (Figrue3f). This indicates that the 14-gene feature has good accuracy for predicting ARHL.

[0084] Figure 15 The results of the confusion matrix model for the Training group are shown in the figure.

[0085] Figure 17 The independent AUC values ​​of the genes are shown.

[0086] Figure 18 The ROC of the Test group in the glmBoost+ NaiveBayes model is shown.

[0087] Figure 19 The confusion matrix model results for the Test group are displayed.

[0088] 4. The prediction model used in the examples is the glmBoost+NaiveBayes model constructed based on "A method for constructing a prediction model for age-related hearing loss based on human peripheral blood transcriptomics data".

[0089] (1) Tool development: Develop an interactive web application based on the Shiny framework of R language. The front-end design is a form with 14 input boxes, each corresponding to a core gene (such as S1PR3, NAB2, P2RX4, etc.).

[0090] (2) Backend integration: The trained glmBoost+NaiveBayes model (containing coefficients of 14 genes) and other necessary parameters of the model are saved as an RDS file and integrated into the server-side logic of the Shiny application. The server-side code is responsible for receiving input, performing necessary transformations (if needed), calling the model to calculate the predicted probability, and using R packages such as shapviz to calculate the SHAP value corresponding to each input.

[0091] (3) Deployment and Application: Deploy the developed Shiny application to a cloud server and generate an accessible URL. After obtaining the user's genetic testing results, the user (such as a researcher or clinician) accesses the URL and enters the quantitative values ​​of 14 metabolites one by one in the left panel. After clicking "Submit", the right panel will immediately display the probability of age-related hearing loss based on the metabolite profile and generate a color SHAP waterfall chart, which explains in detail how each metabolite promotes or inhibits the final risk score.

[0092] The above descriptions are merely embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the scope of the invention, and these should also be considered within the protection scope of the present invention. These modifications and improvements will not affect the effectiveness of the invention or the practicality of the patent. The scope of protection claimed in this application shall be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data, characterized by: Includes the following steps: S1 Sample Collection and Data Acquisition: Peripheral blood samples were collected from the study population for whole transcriptome analysis to obtain raw transcriptome gene expression profile data. At the same time, the audiological assessment results and age information of the subjects were collected to form a training set. S2 core transcript screening: Core transcripts were screened using a stratified screening strategy based on the degree of aging and hearing loss. S3 Model Construction: Using the selected core transcriptome as features and hearing status as labels, a machine learning prediction model is constructed on a computer using the training set. S4 Model Validation: The constructed model was evaluated for performance using an independent internal validation set and an external validation queue. The calculated metrics included AUC, accuracy, sensitivity, and specificity to identify core genes, which included: SFRP5, SAXO1, GOLGA6L4, NCAPH, MED31, P2RX4, CYP2J2, ZSCAN23, S1PR3, MTHFR, DCBLD2, TNFAIP8L3, NAB2, and SLC7A3.

2. The method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data according to claim 1, characterized in that: It also includes step S5 model application: the prediction module storing the machine learning prediction model is deployed on the server side. The prediction module is used to receive the detection value of the core gene input by the user, and then generate the risk probability of the user suffering from age-related deafness as the prediction result.

3. The method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data according to claim 2, characterized in that: It also includes a front-end interaction module for communication with the prediction module. The front-end interaction module is used to input the detection values ​​of the user's core genes and output the prediction results.

4. The method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data according to claim 3, characterized in that: The front-end interaction module calls the SHAP interpreter based on the user's input data, and finally generates a SHAP force map.

5. The method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data according to claim 4, characterized in that: The screening criteria in S2 include: the study population includes patients with age-related hearing loss and a healthy control group; significant differentially expressed genes are screened by comparing the differences in mRNA expression between patients with age-related hearing loss and the healthy control group; and differentially expressed genes that are significantly associated with hearing loss are screened by correlation analysis with the pure tone mean hearing threshold.

6. The method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data according to claim 4, characterized in that: The screening criteria in S2 include: screening for gene modules significantly associated with age-related hearing loss through network-weighted co-expression, taking the intersection of differentially expressed genes significantly associated with hearing level with genes in the age-related hearing loss module, and selecting the core transcriptome set that is associated with the age-related hearing loss disease state itself.

7. The method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data according to claim 5 or 6, characterized in that: S3 employs a variety of machine learning algorithms, including Lasso regression, gradient boosting linear model glmBoost, random forest, and NaiveBayes.

8. The application of a combination of gene biomarkers in the method for constructing a risk prediction model for age-related hearing loss based on human peripheral blood transcriptomics data as described in claim 1, characterized in that: The combination of genetic markers includes the following genes: SFRP5, SAXO1, GOLGA6L4, NCAPH, MED31, P2RX4, CYP2J2, ZSCAN23, S1PR3, MTHFR, DCBLD2, TNFAIP8L3, NAB2 and SLC7A3.

Images