Personalized data-based hearing aid intelligent fitting system and model construction method
By building a hearing aid intelligent fitting system based on personalized data, using machine learning models to predict satisfaction and gain values, the problem that fitting results in the existing technology rely on the hearing experience is solved, and more accurate and faster hearing aid fitting effect.
Patent Information
- Application Number
- CN202411421405.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2025-02-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing hearing aid fitting technology relies on the experience and subjective judgment of the listener, making it difficult to achieve objective and accurate satisfaction prediction and gain value setting, resulting in the problem of matching but not wearing it.
By building a hearing aid intelligent fitting system based on personalized data, using machine learning models (such as Random Forest, LightGBM, XGBoost, etc.) to preprocess and feature extraction of patients' personalized data, and build a satisfaction prediction model and a gain value prediction model. , automatically predict satisfaction and set gain value.
The objective and accurate prediction of hearing aid fitting satisfaction is achieved, the influence of human factors is reduced, the gain value suitable for the patient is automatically set, the fitting time is shortened, and the fitting effect is improved.
Smart Images
Figure CN119450326A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of hearing aids, and in particular to a hearing aid intelligent fitting system based on personalized data and a model building method. Background Art
[0002] Hearing aids, as a non-invasive hearing assistive device, are currently the most commonly used hearing intervention method. Not all patients who meet the indications for hearing aids will become successful hearing aid wearers after being fitted with hearing aids. Some patients are not satisfied with their hearing aids in long-term follow-up and do not wear them after fitting. There are many factors that affect the effect of hearing aid fitting, including auditory factors such as the degree of hearing loss and speech recognition ability, as well as non-auditory factors such as age, education level, chronic disease status, living conditions, and personal psychosocial factors. Existing studies have explored the relationship between individual patient factors and satisfaction, as well as the many reasons that lead to fitting but not wearing. However, the current patient satisfaction after fitting still depends on the experience and subjective judgment of the audiologist, which is not objective and accurate enough.
[0003] Determining the gain values of different channels of hearing aids is the core and most important content of hearing aid fitting. The existing prescription formula can automatically calculate the gain value based on the patient's hearing test results and basic information (gender, age), but this cannot achieve the best hearing aid effect. The audiologist needs to adjust the gain based on the patient's feedback, which is not only related to auditory factors such as the patient's hearing loss degree, but also to many aspects such as age, education level, and personal psychosocial factors. Due to the large individual differences in hearing aid fitting, there are high requirements for the clinical experience and professional skills of the audiologist, and the debugging process is usually time-consuming. Summary of the invention
[0004] In order to overcome the above technical defects, the first aspect of the present invention provides a method for constructing a hearing aid fitting satisfaction prediction model, which comprises:
[0005] Step S1: extracting personalized data of the hearing aid wearer from the collected sample data;
[0006] Step S2: Process missing values and outliers on the personalized data, and further perform standardization preprocessing. After the preprocessing, the data is divided into a training set and a test set according to a certain ratio;
[0007] Step S3: Construction of a piecewise linear regression model: daily wearing time is used as the independent variable and satisfaction is used as the dependent variable.
[0008]
[0009] Among them, Y represents satisfaction; X represents daily wearing time; β0 represents the intercept term, that is, the satisfaction when the daily wearing time is 0; β1 represents the regression coefficient, which indicates the influence of wearing time on satisfaction; Δ j Represents the increment of wearing time; I represents the indicator function I, which distinguishes the interval of daily wearing time;
[0010] Step S4: After determining the inflection point, the data is divided into two segments, and a linear regression model is fitted in each segment;
[0011] Step S5: Model training and evaluation: Use conventional modeling steps in this field to build random forest models, decision tree models, LightGBM models and XGBoost models respectively, use training set data for model training, use 5-fold cross-validation method to evaluate the model, and adjust model parameters to obtain optimal performance based on the cross-validation results;
[0012] Step S6: Use the test set data to evaluate and compare the model performance of the trained random forest model, decision tree model, LightGBM model, and XGBoost model, and finally select the model with the highest prediction accuracy as the satisfaction prediction model.
[0013] Preferably, in step S4, the inflection point is 5.1 hours, and the data is divided into two segments: a segment less than 5.1 hours and a segment greater than or equal to 5.1 hours;
[0014] A linear regression model is fitted in each segment, and the formula is as follows:
[0015] For wearing time t<5.1 hours:
[0016] Y=β0+β1X+ξ1
[0017] For wearing time t ≥ 5.1 hours:
[0018] Y=β2+β3X+ξ2
[0019] Among them, β0 represents the intercept term, which indicates the satisfaction when the daily wearing time is 0; β2 is the intercept term, which indicates the satisfaction when the daily wearing time is 5.1 hours; β1 and β3 represent regression coefficients, which indicate the influence of wearing time on satisfaction; ξ1 represents the error term when the wearing time is less than 5.1 hours; ξ2 represents the error term when the wearing time is greater than or equal to 5 hours.
[0020] Preferably, the model with the highest prediction accuracy is a random forest model, and the method for constructing the random forest model includes:
[0021] (1) The model parameters are set as follows: number of trees 100, maximum depth infinite, minimum sample split number 2, minimum sample leaf number 20;
[0022] (2) By constructing multiple decision trees, each tree randomly selects samples from the original data set during training, and randomly selects a part of the features from all the features for splitting, and finally combines the prediction results of all the trees;
[0023] (3) The model was evaluated using the 5-fold cross-validation method;
[0024] (4) Based on the cross-validation results, adjust the model parameters to obtain the best performance.
[0025] Preferably, the personalized data includes gender, age, communication language, genetic history, patient's previous medical history, smoking history, drinking history, whether there are accompanying ear symptoms, whether the patient has a history of ear diseases, noise exposure history, history of use of ototoxic drugs, whether there is a history of sudden deafness in the left ear, whether there is a history of sudden deafness in the right ear, the course of disease in the left ear, the course of disease in the right ear, whether it is the first fitting, whether the hearing aid has been tried in another hospital, the side of the hearing aid fitting, the type of hearing aid, the price of the hearing aid, the channel through which the decision to come for hearing aid fitting is made, the expectation of the hearing aid effect, living style, education level, nature of work, cognitive ability, the results of the patient's hearing test, air conduction hearing threshold, bone conduction hearing threshold, the type of hearing loss in the left and right ears, and the speech recognition rate before hearing aid.
[0026] A second aspect of the present application provides a method for constructing a hearing aid gain value prediction model, comprising:
[0027] Step S1: extracting personalized data of the hearing aid wearer from the collected sample data;
[0028] Step S2: Feature extraction: (1) Use the AIC method to select the features that contribute most to the model prediction results; (2) Use the stepwise linear regression method to gradually select features; (3) Use deep learning technology to generate high-order features;
[0029] Step S3: Data preprocessing: standardize all numerical features; perform one-hot encoding on categorical features; after data preprocessing, divide the data into training set and test set according to a specific ratio;
[0030] Step S4: Build a decision tree model, LightGBM model, XGBoost model, and random forest model respectively, use the training set data to train the model, use the 5-fold cross-validation method to evaluate the model, and adjust the model parameters to obtain the best performance based on the cross-validation results;
[0031] Step S5: Use the test set data to evaluate and compare the gain value mean absolute errors of the trained decision tree model, LightGBM model, XGBoost model, and random forest model, and finally select the model with the smallest mean absolute error as the gain value prediction model.
[0032] Preferably, in step S5, the gain value prediction model is an integrated model consisting of a LightGBM model and an XGBoost model.
[0033] Preferably, the method for constructing the LightGBM model includes:
[0034] (1) The model parameters are set as follows: learning rate 0.1, number of leaf nodes 50, maximum tree depth infinite, minimum leaf node data size 20, feature sampling rate 0.6, data sampling rate 0.6, data sampling frequency 5;
[0035] (2) Use the training set data to train the LightGBM model, set the loss function to the mean absolute error MAE of the regression task, and calculate it as:
[0036]
[0037] N represents the total average error of a certain number of patients; Error is the average gain error of each channel of the patient; n is the number of channels; y i is the expected output value of channel i, that is, the actual gain value of the patient; y′ i is the predicted output value of the model for channel i;
[0038] (3) Divide the training set data into 5-fold cross-validation, train the model on each fold and evaluate the performance, and record the mean absolute error of each validation;
[0039] (4) Use the grid search method to tune parameters, traverse multiple parameter combinations, and find the optimal parameters;
[0040] (5) Combined with the early stopping strategy, when the error on the validation set fails to decrease significantly in several consecutive iterations, the training is stopped early;
[0041] (6) The parameter combination with the smallest mean absolute error is taken as the final parameter;
[0042] The method for constructing the XGBoost model includes:
[0043] (1) The model parameters are set as follows: learning rate 0.1, number of leaf nodes 50, maximum tree depth 6, minimum leaf node data size 1, subsample ratio 0.8, feature sampling ratio 0.8, and L2 regularization term coefficient set to 1;
[0044] (2) Use the training set data to train the XGBoost model and set the loss function to the mean absolute error (MAE) of the regression task. The calculation formula is:
[0045]
[0046] N represents the total average error of a certain number of patients; Error is the average gain error of each channel of the patient; n is the number of channels; y i is the expected output value of channel i, that is, the actual gain value of the patient; y′ i is the predicted output value of the model for channel i;
[0047] (3) Divide the training set data into 5-fold cross-validation, train the model on each fold and evaluate the performance, and record the mean absolute error of each validation;
[0048] (4) Use the grid search method to tune parameters, traverse multiple parameter combinations, and find the optimal parameters;
[0049] (5) Combined with the early stopping strategy, when the error on the validation set fails to decrease significantly in several consecutive iterations, the training is stopped early;
[0050] (6) The parameter combination with the smallest mean absolute error is taken as the final parameter.
[0051] Preferably, the personalized data includes gender, age, communication language, genetic history, patient's previous medical history, smoking history, drinking history, whether there are accompanying ear symptoms, whether the patient has a history of ear diseases, noise exposure history, history of use of ototoxic drugs, whether there is a history of sudden deafness in the left ear, whether there is a history of sudden deafness in the right ear, the course of disease in the left ear, the course of disease in the right ear, whether it is the first fitting, whether the hearing aid has been tried in another hospital, the side of the hearing aid fitting, the type of hearing aid, the price of the hearing aid, the channel through which the decision to come for hearing aid fitting is made, the expectation of the hearing aid effect, living style, education level, nature of work, cognitive ability, the results of the patient's hearing test, air conduction hearing threshold, bone conduction hearing threshold, the type of hearing loss in the left and right ears, and the speech recognition rate before hearing aid.
[0052] The third aspect of the present application provides a hearing aid intelligent fitting system based on personalized data, which includes:
[0053] A data input module, the data input module is used to receive personalized data of the current hearing aid wearer;
[0054] A gain value prediction module, the gain value prediction module is used to predict a gain value suitable for the current hearing aid wearer based on the personalized data of the current hearing aid wearer obtained by the data input module and using a gain value prediction model pre-constructed by the above-mentioned hearing aid gain value prediction model construction method;
[0055] A result output module is used to output the gain value suitable for the current hearing aid wearer predicted by the gain value prediction module.
[0056] Preferably, the hearing aid intelligent fitting system based on personalized data further includes a satisfaction prediction module, which is used to predict the hearing aid wearing satisfaction of the current hearing aid wearer according to the personalized data of the current hearing aid wearer obtained by the data input module, using the satisfaction prediction model pre-constructed by the above-mentioned hearing aid fitting satisfaction prediction model construction method; the result output module is also used to output the hearing aid wearing satisfaction of the current hearing aid wearer predicted by the satisfaction prediction module. The satisfaction is measured using an international hearing aid effect questionnaire, and the qualitative standards are as follows: a total score of ≥28 points indicates satisfaction; a total score of 20 to 27 points indicates acceptable; a total score of <20 points indicates dissatisfaction.
[0057] It is worth noting that the types of hearing aids in the present application include but are not limited to 3-channel hearing aids, 4-channel hearing aids, 6-channel hearing aids, 7-channel hearing aids, 8-channel hearing aids, 10-channel hearing aids, 12-channel hearing aids, 14-channel hearing aids, 15-channel hearing aids, 16-channel hearing aids, 17-channel hearing aids or 20-channel hearing aids, etc.
[0058] Compared with the prior art, the above technical solution has the following beneficial effects:
[0059] The present invention provides a hearing aid intelligent fitting system based on personalized data. The system uses a pre-built satisfaction prediction model and a gain value output prediction model. According to the personalized data of the current patient, the system can not only automatically and accurately predict the current patient's hearing aid fitting satisfaction, excluding the influence of human factors, but also automatically and accurately output the most suitable gain value for the current patient, without the need for the audiologist to repeatedly adjust the gain value of the hearing aid, which helps to shorten the fitting time, save manpower and material resources, and improve the hearing aid fitting effect. Thus, an objective, fast, accurate, personalized, and highly satisfactory hearing aid fitting solution can be provided to different patients. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 This is a flowchart of a method for constructing a hearing aid intelligent fitting satisfaction prediction model in an embodiment of the present application;
[0061] Figure 2 A and B in the figure are screenshots of the international hearing aid effectiveness survey questionnaire used to quantify satisfaction;
[0062] Figure 3 is a flowchart of a method for constructing a hearing aid gain value prediction model in an embodiment of the present application;
[0063] Figure 4Schematic diagram of a hearing aid intelligent fitting system in one embodiment of the present application. DETAILED DESCRIPTION
[0064] The advantages of the present invention are further described below in conjunction with the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following specific description is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.
[0065] Example 1: Constructing a hearing aid intelligent fitting satisfaction prediction model
[0066] like Figure 1 As shown, this embodiment provides a method for constructing a hearing aid intelligent fitting satisfaction prediction model, including steps S1 to S7:
[0067] Step S1: extracting personalized data of the hearing aid wearer from the collected sample data.
[0068] A large amount of sample data was obtained through the survey. The inclusion criteria included: (a) age 18 years or older; (b) bilateral hearing loss, defined as pure tone audiometry (PTA, average air conduction audiometry thresholds at 0.5, 1, 2, and 4 kHz) greater than 25 dB in both ears; (c) native Chinese speakers; (d) able to cooperate with satisfaction follow-up three months after hearing aid fitting. Exclusion criteria: (a) patients with severe psychological or cognitive disorders who were unable to objectively evaluate the effect of hearing aid use. (b) patients with more than 20% missing data were excluded.The patient's audiological and other personalized characteristics were extracted from the collected sample data, as follows: gender (male, female); age (unit: years); communication language (the patient's usual language: dialect, Mandarin, both are often used); genetic history (whether there is a family history of deafness); the patient's previous medical history: hyperglycemia (yes or not), hypertension (yes or not), hyperlipidemia (yes or not), coronary heart disease (yes or not), cerebral infarction / cavity infarction (yes or not), nasopharyngeal cancer (yes or not), thyroid disease (yes or not), renal insufficiency (yes or not), cervical vertebra disease (yes or not); smoking history (yes or not); drinking history (yes or not); whether there are associated ear symptoms: dizziness (yes or not) , tinnitus, earache, and otorrhea are all divided into (no, left ear, right ear, and both ears); whether the patient has a history of ear diseases: external ear inflammation, history of otitis media, tympanic membrane perforation, and history of surgery, all of which are divided into (no, left ear, right ear, and both ears); history of noise exposure (yes or no); history of use of ototoxic drugs (yes or no); whether there is a history of sudden deafness in the left ear; whether there is a history of sudden deafness in the right ear; course of disease in the left ear (how many years has it been since the deafness was fitted with a hearing aid); course of disease in the right ear (how many years has it been since the deafness was fitted with a hearing aid); whether it is the first fitting (yes or no); whether the hearing aid has been tried in another hospital (yes or no); hearing aid fitting side (left ear, right ear, and both ears); hearing aid type (high-power behind-the-ear, standard Behind-the-ear, open behind-the-ear, in-the-ear, canal); hearing aid price (unit: 10,000 yuan); how they decided to get hearing aids (voluntary, through advertisements and other materials, doctor recommendation, family recommendation, friend recommendation); expectations of hearing aid effects: positive, negative, uncertain; living style: living alone, not living alone; education: illiterate, elementary school, junior high school, high school / technical secondary school, junior college, undergraduate, undergraduate and above; nature of work: mental, physical, unemployed; cognitive ability: normal, mild mental disorder, moderate mental disorder, severe mental disorder; the results of the patient's hearing test (pure tone audiometry), with both the left and right ears measured Results: It is divided into air conduction hearing threshold and bone conduction hearing threshold. The air conduction hearing threshold is tested at 7 frequencies, including air conduction hearing threshold (AC, unit: dB): 125Hz, 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, and 8000Hz; the bone conduction hearing threshold is tested at 5 frequencies, including bone conduction hearing threshold (BC, unit: dB): 250Hz, 500Hz, 1000Hz, 2000Hz, and 4000Hz; the hearing loss type of the left and right ears (divided into sensorineural deafness, conductive deafness, and mixed deafness); the speech recognition rate before hearing aid is divided into speech recognition rate at 30dB, 40dB, and 50dB (unit: %).
[0069] Due to too many variables, it is difficult to perform data analysis and visualization. In order to solve this problem, principal component analysis (PCA) was used to reduce the data dimension and retain the variance of the data set. Finally, 8 factors (new composite features) representing related elements were generated: Factor 1: reflects the patient's hearing test results; Factor 2: reflects the patient's speech recognition ability; Factor 3: reflects the patient's medical history (hypertension, diabetes, etc.); Factor 4: reflects the patient's ear disease history; Factor 5: reflects the patient's cognitive ability; Factor 6: reflects the patient's living habits (smoking, drinking history, etc.); Factor 7: reflects the patient's personalized background (age, gender, communication language), Factor 8: reflects the patient's hearing aid usage (hearing aid type, price, etc.). The above factors are all used as input factors of the model, and the satisfaction of deaf patients after 3 months, 6 months, 9 months, and 1 year of hearing aid fitting (satisfied, acceptable, dissatisfied) is used as the output. Satisfaction is measured using the International Outcome Inventory for Hearing Aids (IOI-HA) (see Figure 2 ), which is an internationally recognized hearing aid effect evaluation questionnaire, consisting of 7 core questions, each question includes 5 options, corresponding to a score of 1 to 5 points, and a total score of 35 points. The higher the score, the better the hearing aid effect. The qualitative criteria are as follows: a total score of ≥28 points is satisfactory; a total score of 20 to 27 points is acceptable; a total score of <20 points is unsatisfactory. A total of 305 cases of data were included.
[0070] Step S2: Preprocess the personalized data and divide it into a training set and a test set according to a certain ratio.
[0071] The personalized data is processed for missing values and outliers to ensure the integrity and accuracy of the data. The data is further preprocessed for standardization, and the preprocessed data is divided into a training set and a test set according to a certain ratio. Preferably, the training set and the test set are divided according to 8:2.
[0072] Step S3: Construction of a piecewise linear regression model: daily wearing time is used as the independent variable and satisfaction is used as the dependent variable.
[0073]
[0074] Where: Y: satisfaction; X: daily wearing time; β0: intercept term, that is, satisfaction when the daily wearing time is 0; β1: regression coefficient, indicating the influence of wearing time on satisfaction; : increment of wearing time; I: indicator function I, which distinguishes the interval of daily wearing time.
[0075] Step S4: After determining the inflection point, the data is divided into two segments, and a linear regression model is fitted in each segment;
[0076] The data were analyzed using econometric methods, and it was found that there was a positive correlation between wearing time and satisfaction, and that patients with a longer illness had a more significant impact on wearing time. Patients with different daily hearing aid wearing times had different satisfaction distribution characteristics. In order to better understand the relationship between wearing time and satisfaction, the segmented linear regression model (a segmented linear regression model is a special regression model that allows different linear relationships to be used in different segments of the data) was used to determine that the wearing time has an inflection point at 5.1 hours. According to the daily wearing time of 5.1 hours, the sample was divided into two sub-samples, and the data was divided into two segments: a segment of less than 5.1 hours and a segment of greater than or equal to 5.1 hours, and the satisfaction results were predicted for different sub-samples.
[0077] A linear regression model is fitted in each segment, and the formula is as follows:
[0078] For wearing time t<5.1 hours:
[0079] Y=β0+β1X+ξ1
[0080] For wearing time t ≥ 5.1 hours:
[0081] Y=β2+β3X+ξ2
[0082] Among them: β0, β2: intercept terms, representing the satisfaction when the daily wearing time is 0 and 5.1 hours respectively; β1, β3: regression coefficients, representing the influence of wearing time on satisfaction; ξ1, ξ2: representing the error terms when the wearing time is less than 5.1 hours and greater than or equal to 5 hours respectively.
[0083] Step S5: Model training and evaluation:
[0084] The two sub-sample data sets were divided into training set and test set according to 8:2 respectively. For the training set, four models were used, including decision tree (DT, Decision Tree), random forest (RF, Random Forest) and gradient boosting model (including LightGBM and XGBoost). The conventional modeling steps in this field were adopted, and the 5-fold cross-validation method was used to evaluate the model. According to the cross-validation results, the model parameters were adjusted to obtain the best performance.
[0085] Among them, the construction method of the random forest model is:
[0086] (1) The key parameters of the model are set as: number of trees (n_estimators) 100, maximum depth (max_depth) infinite, minimum number of sample splits (min_samples_split) 2, and minimum number of sample leaves (min_samples_leaf) 20.
[0087] (2) Model training: Random forest constructs multiple decision trees. During training, each tree randomly selects samples from the original data set (with replacement sampling) and randomly selects a portion of features from all features for splitting. Finally, the prediction results of all trees are combined.
[0088] (3) Model evaluation and validation: The model was evaluated using a 5-fold cross-validation method to prevent overfitting;
[0089] (4) Parameter adjustment and model optimization: Based on the cross-validation results, adjust the model parameters to obtain the best performance.
[0090] Step S6: Use the test set data to evaluate and compare the model performance of the trained random forest model, decision tree model, LightGBM model, and XGBoost model, and finally select the model with the highest prediction accuracy as the satisfaction prediction model.
[0091] The accuracy index of the model was calculated to comprehensively evaluate the model's prediction ability and classification effect. According to the test results, the satisfaction decision system adopted the random forest model.
[0092] The model performance of the trained random forest model, decision tree model, LightGBM model, and XGBoost model was evaluated and compared using the test set data. Finally, it was found that the random forest model had the best effect. Therefore, the random forest model with the highest prediction accuracy was finally selected as the satisfaction prediction model. Specifically, as shown in Table 1, the accuracy of the model random forest on the test set reached 91%. When running the model on the subsample data, the prediction accuracy was significantly improved, and the random forest model had the best effect. In the test set less than 5.1 hours, the random forest accuracy was 96.4%; in the test set greater than 5.1 hours, the random forest accuracy was 96.7%.
[0093] Table 1 Prediction accuracy of hearing aid fitting satisfaction of each model
[0094] Model Name Full sample Wearing time <5.1h Wearing time ≥5.1h DT 0.875 0.839 0.866 LightGBM 0.768 0.893 0.800 XGBoost 0.768 0.893 0.800 RF 0.910 0.964 0.967
[0095] Example 2: Constructing a hearing aid gain value prediction model
[0096] The present embodiment provides a method for constructing a hearing aid gain value prediction model. This embodiment takes Widex's practical basic hearing aids (4-channel and 6-channel hearing aids) as an example: by collecting and processing data from patients with 4-channel and 6-channel hearing aids, using AIC, stepwise linear regression and deep feature generation methods to enrich input features, the model is trained to accurately predict the gain values of large, medium and small sounds of each channel. Finally, an integrated model that integrates the LightGBM and XGBoost models was selected as the hearing aid gain value prediction model. The hearing aid gain value prediction model in this embodiment is significantly better than the Widex prescription formula in gain value prediction, showing a lower mean absolute error, and improving the accuracy and effect of hearing aid fitting.
[0097] Specifically, if Figure 3 As shown, the method for constructing a hearing aid gain value prediction model includes steps S1 to S5.
[0098] Step S1: extracting personalized data of the hearing aid wearer from the collected sample data.
[0099] For example, personalized data of patients with 4-channel and 6-channel hearing aids were collected from follow-up patients. After excluding the data of patients with unsatisfactory follow-up results, 300 patients with 4-channel and 6-channel hearing aids were finally collected.
[0100] The patient's audiological and other personalized characteristics were extracted from the collected sample data, as follows: gender (male, female); age (unit: years); communication language (the patient's usual language: dialect, Mandarin, both are often used); genetic history (whether there is a family history of deafness); the patient's previous medical history: hyperglycemia (yes or not), hypertension (yes or not), hyperlipidemia (yes or not), coronary heart disease (yes or not), cerebral infarction / cavity infarction (yes or not), nasopharyngeal cancer (yes or not), thyroid disease (yes or not), renal insufficiency (yes or not), cervical vertebra disease (yes or not); smoking history (yes or not); drinking history (yes or not); whether there are associated ear symptoms: dizziness (yes or not) , tinnitus, earache, and otorrhea are all divided into (no, left ear, right ear, and both ears); whether the patient has a history of ear diseases: external ear inflammation, history of otitis media, tympanic membrane perforation, and history of surgery, all of which are divided into (no, left ear, right ear, and both ears); history of noise exposure (yes or no); history of use of ototoxic drugs (yes or no); whether there is a history of sudden deafness in the left ear; whether there is a history of sudden deafness in the right ear; course of disease in the left ear (how many years has it been since the deafness was fitted with a hearing aid); course of disease in the right ear (how many years has it been since the deafness was fitted with a hearing aid); whether it is the first fitting (yes or no); whether the hearing aid has been tried in another hospital (yes or no); hearing aid fitting side (left ear, right ear, and both ears); hearing aid type (high-power behind-the-ear, standard Behind-the-ear, open behind-the-ear, in-the-ear, canal); hearing aid price (unit: 10,000 yuan); how they decided to get hearing aids (voluntary, through advertisements and other materials, doctor recommendation, family recommendation, friend recommendation); expectations of hearing aid effects: positive, negative, uncertain; living style: living alone, not living alone; education: illiterate, elementary school, junior high school, high school / technical secondary school, junior college, undergraduate, undergraduate and above; nature of work: mental, physical, unemployed; cognitive ability: normal, mild mental disorder, moderate mental disorder, severe mental disorder; the results of the patient's hearing test (pure tone audiometry), with both the left and right ears measured Results: It is divided into air conduction hearing threshold and bone conduction hearing threshold. The air conduction hearing threshold is tested at 7 frequencies, including air conduction hearing threshold (AC, unit: dB): 125Hz, 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, and 8000Hz; the bone conduction hearing threshold is tested at 5 frequencies, including bone conduction hearing threshold (BC, unit: dB): 250Hz, 500Hz, 1000Hz, 2000Hz, and 4000Hz; the hearing loss type of the left and right ears (divided into sensorineural deafness, conductive deafness, and mixed deafness); the speech recognition rate before hearing aid is divided into speech recognition rate at 30dB, 40dB, and 50dB (unit: %).
[0101] Step S2: Feature extraction:
[0102] In terms of feature extraction, we used a variety of methods to enrich input features to improve the performance of the model. Method 1, AIC (Akaike Information Criterion): The AIC method is used to screen out the features that contribute most to the model prediction results, reduce redundant features, and avoid overfitting. Method 2, stepwise linear regression method: The stepwise regression method is used to gradually screen features to make the model more interpretable and stable. Method 3, deep feature generation: Deep learning technology is used to generate high-order features to further improve the expression ability of the feature space and enhance the prediction accuracy of the model.
[0103] Step S3: Data preprocessing:
[0104] In terms of data preprocessing, all numerical features are standardized to ensure that different features have the same scale. Categorical features are one-hot encoded to facilitate model processing. After data preprocessing, the data is divided into training set and test set in an 8:2 ratio.
[0105] Step S4: Build a decision tree model, LightGBM model, XGBoost model, and random forest model respectively, use the training set data for model training, use the 5-fold cross-validation method to evaluate the model, and adjust the model parameters according to the cross-validation results to obtain the best performance.
[0106] Among them, the construction method of the LightGBM model includes:
[0107] (1) The key parameters of the model are set as follows: learning rate 0.1, number of leaf nodes 50, maximum tree depth infinite, minimum leaf node data size 20, feature sampling rate 0.6, data sampling rate 0.6, and data sampling frequency 5;
[0108] (2) Use the training set data to train the LightGBM model, set the loss function to the mean absolute error MAE of the regression task, and calculate it as:
[0109]
[0110] N represents the total average error of a certain number of patients;
[0111] The average gain error (Error) of each channel of the patient is calculated using the following formula:
[0112]
[0113] n is the number of channels (for example, when calculating the loud, medium, or quiet gain of a four-channel hearing aid, n=4; when calculating the loud, medium, or quiet gain of a six-channel hearing aid, n=6); y iis the expected output value of channel i (i.e., the patient's actual gain value); y′ i is the predicted output value of the model for channel i;
[0114] (3) Divide the training set data into 5-fold cross-validation, train the model on each fold and evaluate the performance, and record the mean absolute error (MAE) of each validation;
[0115] (4) Use the grid search method to tune parameters, traverse multiple parameter combinations, and find the optimal parameters;
[0116] (5) Combined with the early stopping strategy, when the error on the validation set fails to decrease significantly in several consecutive iterations, training is stopped early to prevent overfitting;
[0117] (6) The parameter combination with the smallest mean absolute error is taken as the final parameter.
[0118] XGBoost (Extreme Gradient Boosting) is an enhanced version of the gradient boosting framework. The model parameters are set as follows: learning rate 0.1, number of leaf nodes 50, maximum tree depth 6, minimum leaf node data size 1, subsample ratio 0.8, feature sampling ratio 0.8, and L2 regularization coefficient 1. The loss function and model training process are the same as LightGBM.
[0119] Step S5: Use the test set data to evaluate and compare the gain value mean absolute error (MAE) of the trained decision tree model, LightGBM model, XGBoost model, and random forest model, and finally select the model with the smallest mean absolute error as the gain value prediction model, see Tables 2 and 3.
[0120] Table 2 Mean absolute error of gain output of four-channel hearing aid
[0121] Model Low volume gain Mid-range gain Loud Gain DT 3.93 4.00 3.86 LightGBM 3.74 3.20 2.29 XGBoost 3.53 3.32 2.29 RF 5.44 3.77 3.49
[0122] Table 3 Mean absolute error of six-channel hearing aid gain output
[0123] Model Low volume gain Mid-range gain Loud Gain DT 3.86 3.54 2.78 LightGBM 3.94 3.25 2.69 XGBoost 3.51 2.94 2.69 RF 4.03 3.32 2.83
[0124] According to the data in Table 2 and Table 3, the average absolute error of the gain value of the test set is about 3dB. The average absolute error of the four-channel is: 2.29 for loud, 3.20 for medium, and 3.53 for quiet; the average absolute error of the six-channel is: 2.69 for loud, 2.94 for medium, and 3.51 for quiet. When selecting a model for gain value output, give priority to models with smaller average absolute errors. For four-channel hearing aids, the loud and medium gain values are output by the LightGBM model, while the quiet gain value is output by the XGBoost model. For six-channel hearing aids, the loud, medium, and quiet gain values all use the XGBoost model. This selection can ensure the accuracy and reliability of the gain value output and improve the fitting effect of the hearing aid.
[0125] In addition, the mean absolute error (the error between the prescription formula and the audiologist's final gain value) of the existing Widex fitting formula (WIDEX FITTING RATIONALE) was calculated on the test set. The 4-channel and 6-channel errors were about 6-8dB, which is larger than the mean absolute error of the model in this application. Therefore, the fitting effect of the existing Widex fitting prescription formula is worse.
[0126] Example 3: Intelligent hearing aid fitting system based on personalized data
[0127] Based on the hearing aid intelligent fitting satisfaction prediction model that has been constructed in Example 1 and the hearing aid gain value prediction model that has been constructed in Example 2, this embodiment provides a hearing aid intelligent fitting system based on personalized data, which integrates the satisfaction prediction function and the gain value output prediction function, and can quickly output the gain value data of patients with satisfactory follow-up results. The multifunctional hearing aid intelligent fitting system of the present application is applicable to but not limited to the intelligent fitting of the following types of hearing aids: 3-channel hearing aids, 4-channel hearing aids, 6-channel hearing aids, 7-channel hearing aids, 8-channel hearing aids, 10-channel hearing aids, 12-channel hearing aids, 14-channel hearing aids, 15-channel hearing aids, 16-channel hearing aids, 17-channel hearing aids or 20-channel hearing aids, etc. For example, AAV 3, 6, 7, 8, 10, 14, 17, 20-channel hearing aids; Starkey 4, 6, 8, 12, 16-channel hearing aids; Widex 4, 6, 10, 12, 15-channel hearing aid patients.
[0128] For example, the multifunctional hearing aid intelligent fitting system based on personalized data in this embodiment includes: a data input module, a satisfaction prediction module, a gain value prediction module and a result output module.
[0129] The data input module is used to receive the personalized data of the current hearing aid wearer.
[0130] For example, personalized data includes but is not limited to multiple or all of the following information: gender (male, female); age (unit: years); communication language (the patient's usual language: dialect, Mandarin, both are often used); genetic history (whether there is a family history of deafness); the patient's previous medical history: hyperglycemia (yes or not), hypertension (yes or not), hyperlipidemia (yes or not), coronary heart disease (yes or not), cerebral infarction / cavity infarction (yes or not), nasopharyngeal cancer (yes or not), thyroid disease (yes or not), renal insufficiency (yes or not), cervical vertebra disease (yes or not); smoking history (yes or not); drinking history (yes or not); whether there are accompanying ear symptoms: dizziness (yes or not), ear Tinnitus, ear pain, and otorrhea are all divided into (no, left ear, right ear, and both ears); whether the patient has a history of ear diseases: external ear inflammation, history of otitis media, tympanic membrane perforation, and history of surgery, all of which are divided into (no, left ear, right ear, and both ears); history of noise exposure (yes or no); history of use of ototoxic drugs (yes or no); whether there is a history of sudden deafness in the left ear; whether there is a history of sudden deafness in the right ear; course of disease in the left ear (how many years has it been since the deafness was fitted with a hearing aid); course of disease in the right ear (how many years has it been since the deafness was fitted with a hearing aid); whether it is the first fitting (yes or no); whether the hearing aid has been tried in another hospital (yes or no); hearing aid fitting side (left ear, right ear, and both ears); hearing aid type (high-power behind-the-ear, standard ear) behind-the-ear, open behind-the-ear, in-the-ear, canal); hearing aid price (unit: 10,000 yuan); how to decide to get hearing aids (voluntary, through advertisements and other materials, doctor recommendation, family recommendation, friend recommendation); expectation of hearing aid effect: positive, negative, uncertain; living style: living alone, not living alone; education: illiterate, elementary school, junior high school, high school / technical secondary school, junior college, undergraduate, undergraduate and above; nature of work: mental, physical, unemployed; cognitive ability: normal, mild mental disorder, moderate mental disorder, severe mental disorder; the results of the patient's hearing test (pure tone audiometry), and the left and right ear hearing loss were measured at the same time Results; divided into air conduction hearing threshold and bone conduction hearing threshold, air conduction is divided into 7 frequencies for testing, air conduction hearing threshold (AC, unit dB): 125Hz, 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, 8000Hz; bone conduction is divided into 5 frequencies for testing, bone conduction hearing threshold (BC, unit dB): 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz; hearing loss type of left and right ears (divided into: sensorineural deafness, conductive deafness, mixed deafness); speech recognition rate before hearing aid, divided into speech recognition rate at 30dB, 40dB, and 50dB (unit: %).
[0131] The satisfaction prediction module is used to predict the satisfaction of the current hearing aid wearer based on the acquired personalized data of the current hearing aid wearer using a pre-built satisfaction prediction model.
[0132] For example, for a four-channel hearing aid, the gain values of loud and medium sounds use the LightGBM model as the gain value prediction model, while the gain value of soft sounds uses the XGBoost model as the gain value prediction model; for a six-channel hearing aid, the gain values of loud, medium and soft sounds all use the XGBoost model as the gain value prediction model.
[0133] Satisfaction was measured using the International Outcome Inventory for Hearing Aids (IOI-HA) questionnaire (see Figure 2 ), which is an internationally recognized hearing aid effect evaluation questionnaire, consisting of 7 core questions, each question includes 5 options, corresponding to a score of 1 to 5 points, with a total score of 35 points. The higher the score, the better the hearing aid effect. The qualitative standards are as follows: a total score of ≥28 points is satisfactory; a total score of 20 to 27 points is acceptable; a total score of <20 points is unsatisfactory.
[0134] The gain value prediction module is used to predict the gain value suitable for the current hearing aid wearer based on the personalized data of the current hearing aid wearer, using a pre-built gain value prediction model. It can eliminate the influence of human factors, and the gain value output is highly accurate. There is no need to repeatedly debug the gain value of the hearing aid, which helps to shorten the fitting time, save manpower and material resources, and improve the hearing aid fitting effect, thereby providing objective, fast, accurate and personalized hearing aid fitting solutions for different patients.
[0135] The result output module is used to output the hearing aid wearing satisfaction of the current hearing aid wearer predicted by the satisfaction prediction module, and is also used to output the gain value suitable for the current hearing aid wearer predicted by the gain value prediction module.
[0136] For example, the use principle of the hearing aid intelligent fitting system based on personalized data is as follows: Figure 4 When a hearing aid intelligent fitting system is used to fit a patient with a hearing aid, the personalized data of the current patient is first input, and then the satisfaction prediction module uses a pre-built satisfaction prediction model to predict the satisfaction of the current hearing aid wearer, and finally the gain value prediction module uses a pre-built gain value prediction model to predict the gain value suitable for the current hearing aid wearer. Optionally, the satisfaction prediction step can be skipped and the gain value prediction can be performed directly.
[0137] For example, a hearing aid intelligent fitting system based on personalized data can be constructed as a hearing aid intelligent fitting platform based on personalized data, which integrates a pre-built satisfaction prediction model and a pre-built gain value prediction model. The platform architecture includes a front-end user interface, a back-end server and database, and a middle-layer API interface. React and Vue.js are used for front-end development, and the front-end user interface is designed to include a data input block, a satisfaction prediction result output block, and a gain value prediction result output block. The front-end user interface is connected to the back-end server and database through the middle-layer API interface for data interaction. When in use, the user enters the personalized data of the current patient in the data input block. After the satisfaction prediction model and the gain value prediction model predict the current patient's satisfaction and the optimal gain value respectively, the satisfaction prediction result output block and the gain value prediction result output block automatically display the satisfaction prediction result and the gain value prediction result respectively.
[0138] It should be noted that the embodiments of the present invention have better practicability and do not impose any form of limitation on the present invention. Any technician familiar with the field may use the technical content disclosed above to change or modify it into an equivalent effective embodiment. However, any modification or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.
Claims
1. A method for constructing a hearing aid fitting satisfaction prediction model, characterized in that: include: Step S1: extracting personalized data of the hearing aid wearer from the collected sample data; Step S2: Process missing values and outliers on the personalized data, and further perform standardization preprocessing. After the preprocessing, the data is divided into a training set and a test set according to a certain ratio; Step S3: Construction of a piecewise linear regression model: daily wearing time is used as the independent variable and satisfaction is used as the dependent variable. Among them, Y represents satisfaction; X represents daily wearing time; β0 represents the intercept term, that is, the satisfaction when the daily wearing time is 0; β1 represents the regression coefficient, which indicates the influence of wearing time on satisfaction; Δ j Represents the increment of wearing time; I represents the indicator function I, which distinguishes the interval of daily wearing time; Step S4: After the inflection point is determined, the data is divided into two segments, and a linear regression model is fitted in each segment; Step S5: Model training and evaluation: Use conventional modeling steps in this field to build random forest models, decision tree models, LightGBM models and XGBoost models respectively, use training set data for model training, use 5-fold cross-validation method to evaluate the model, and adjust model parameters to obtain optimal performance based on the cross-validation results; Step S6: Use the test set data to evaluate and compare the model performance of the trained random forest model, decision tree model, LightGBM model, and XGBoost model, and finally select the model with the highest prediction accuracy as the satisfaction prediction model.
2. The method for constructing a hearing aid fitting satisfaction prediction model according to claim 1, characterized in that: In step S4, the inflection point is 5.1 hours, and the data is divided into two segments: a segment less than 5.1 hours and a segment greater than or equal to 5.1 hours; A linear regression model is fitted in each segment, and the formula is as follows: For wearing time t<5.1 hours: Y=β0+β1X+ξ1 For wearing time t ≥ 5.1 hours: Y=β2+β3X+ξ2 Among them, β0 represents the intercept term, which indicates the satisfaction when the daily wearing time is 0; β2 is the intercept term, which indicates the satisfaction when the daily wearing time is 5.1 hours; β1 and β3 represent regression coefficients, which indicate the influence of wearing time on satisfaction; ξ1 represents the error term when the wearing time is less than 5.1 hours; ξ2 represents the error term when the wearing time is greater than or equal to 5 hours.
3. The method for constructing a hearing aid fitting satisfaction prediction model according to claim 1, characterized in that: The model with the highest prediction accuracy is the random forest model. The construction method of the random forest model includes: (1) The model parameters are set as follows: number of trees 100, maximum depth infinite, minimum sample split number 2, minimum sample leaf number 20; (2) By constructing multiple decision trees, each tree randomly selects samples from the original data set during training, and randomly selects a part of the features from all the features for splitting, and finally combines the prediction results of all the trees; (3) The model was evaluated using the 5-fold cross-validation method; (4) Based on the cross-validation results, adjust the model parameters to obtain the best performance.
4. The method for constructing a hearing aid fitting satisfaction prediction model according to any one of claim 1, characterized in that: The personalized data includes gender, age, communication language, genetic history, patient's previous medical history, smoking history, drinking history, whether there are accompanying ear symptoms, whether the patient has a history of ear diseases, noise exposure history, history of use of ototoxic drugs, whether there is a history of sudden deafness in the left ear, whether there is a history of sudden deafness in the right ear, the course of disease in the left ear, the course of disease in the right ear, whether it is the first fitting, whether the hearing aid has been tried in another hospital, the side of the hearing aid fitting, the type of hearing aid, the price of the hearing aid, the channel through which the decision to come for hearing aid fitting is made, the expectation of the hearing aid effect, living style, education level, nature of work, cognitive ability, the results of the patient's hearing test, air conduction hearing threshold, bone conduction hearing threshold, the type of hearing loss in the left and right ears, and the speech recognition rate before hearing aid.
5. A method for constructing a hearing aid gain value prediction model, characterized in that: include: Step S1: extracting personalized data of the hearing aid wearer from the collected sample data; Step S2: Feature extraction: (1) The features that contribute most to the model prediction results are selected using the AIC method; (2) Use the stepwise linear regression method to gradually screen features; (3) Generate high-level features using deep learning techniques; Step S3: Data preprocessing: standardize all numerical features; One-hot encoding of categorical features; After data preprocessing is completed, the data is divided into training set and test set according to a specific ratio; Step S4: Build a decision tree model, LightGBM model, XGBoost model, and random forest model respectively, use the training set data to train the model, use the 5-fold cross-validation method to evaluate the model, and adjust the model parameters to obtain the best performance based on the cross-validation results; Step S5: Use the test set data to evaluate and compare the gain value mean absolute errors of the trained decision tree model, LightGBM model, XGBoost model, and random forest model, and finally select the model with the smallest mean absolute error as the gain value prediction model.
6. The method for constructing a hearing aid gain value prediction model according to claim 5, characterized in that: In step S5, the gain value prediction model is an integrated model consisting of a LightGBM model and an XGBoost model.
7. The method for constructing a hearing aid gain value prediction model according to claim 6, wherein: The method for constructing the LightGBM model includes: (1) The model parameters are set as follows: learning rate 0.1, number of leaf nodes 50, maximum tree depth infinite, minimum leaf node data size 20, feature sampling rate 0.6, data sampling rate 0.6, data sampling frequency 5; (2) Use the training set data to train the LightGBM model, set the loss function to the mean absolute error MAE of the regression task, and calculate it as: N represents the total average error of a certain number of patients; Error is the average gain error of each channel of the patient; n is the number of channels; y i is the expected output value of channel i, that is, the actual gain value of the patient; y′ i is the predicted output value of the model for channel i; (3) Divide the training set data into 5-fold cross-validation, train the model on each fold and evaluate the performance, and record the mean absolute error of each validation; (4) Use the grid search method to tune parameters, traverse multiple parameter combinations, and find the optimal parameters; (5) Combined with the early stopping strategy, when the error on the validation set fails to decrease significantly in several consecutive iterations, the training is stopped early; (6) The parameter combination with the smallest mean absolute error is taken as the final parameter; The method for constructing the XGBoost model includes: (1) The model parameters are set as follows: learning rate 0.1, number of leaf nodes 50, maximum tree depth 6, minimum leaf node data size 1, subsample ratio 0.8, feature sampling ratio 0.8, and L2 regularization term coefficient set to 1; (2) Use the training set data to train the XGBoost model, set the loss function to the mean absolute error MAE of the regression task, and calculate it as: N represents the total average error of a certain number of patients; Error is the average gain error of each channel of the patient; n is the number of channels; y i is the expected output value of channel i, that is, the actual gain value of the patient; y′ i is the predicted output value of the model for channel i; (3) Divide the training set data into 5-fold cross-validation, train the model on each fold and evaluate the performance, and record the mean absolute error of each validation; (4) Use the grid search method to tune parameters, traverse multiple parameter combinations, and find the optimal parameters; (5) Combined with the early stopping strategy, when the error on the validation set fails to decrease significantly in several consecutive iterations, the training is stopped early; (6) The parameter combination with the smallest mean absolute error is taken as the final parameter.
8. The method for constructing a hearing aid gain value prediction model according to any one of claim 5, characterized in that: The personalized data includes gender, age, communication language, genetic history, patient's previous medical history, smoking history, drinking history, whether there are accompanying ear symptoms, whether the patient has a history of ear diseases, noise exposure history, history of use of ototoxic drugs, whether there is a history of sudden deafness in the left ear, whether there is a history of sudden deafness in the right ear, the course of disease in the left ear, the course of disease in the right ear, whether it is the first fitting, whether the hearing aid has been tried in another hospital, the side of the hearing aid fitting, the type of hearing aid, the price of the hearing aid, the channel through which the decision to come for hearing aid fitting is made, the expectation of the hearing aid effect, living style, education level, nature of work, cognitive ability, the results of the patient's hearing test, air conduction hearing threshold, bone conduction hearing threshold, the type of hearing loss in the left and right ears, and the speech recognition rate before hearing aid.
9. A hearing aid intelligent fitting system based on personalized data, characterized in that: include: A data input module, the data input module is used to receive personalized data of the current hearing aid wearer; A gain value prediction module, the gain value prediction module being used to predict a gain value suitable for the current hearing aid wearer based on the personalized data of the current hearing aid wearer acquired by the data input module and using a gain value prediction model pre-constructed according to the method of any one of claims 5 to 8; A result output module is used to output the gain value suitable for the current hearing aid wearer predicted by the gain value prediction module.
10. The hearing aid intelligent fitting system based on personalized data according to claim 9, characterized in that: It further includes a satisfaction prediction module, which is used to predict the hearing aid wearing satisfaction of the current hearing aid wearer based on the personalized data of the current hearing aid wearer obtained by the data input module, using a satisfaction prediction model pre-constructed by the method described in any one of claims 1 to 4; the result output module is also used to output the hearing aid wearing satisfaction of the current hearing aid wearer predicted by the satisfaction prediction module.