Early screening method and system for ear-nose-throat diseases

Abstract

The invention discloses an ear-nose-throat disease early screening method. The method comprises the following steps: a voice acquisition step: acquiring a standardized voice sample of a user through an audio acquisition device; a feature extraction step: extracting multi-level acoustic features including vocal cord vibration features, resonance features, pronunciation stability features and depth representation features from the voice sample; and a baseline establishment step. According to the ear-nose-throat disease early-stage screening method and the ear-nose-throat disease early-stage screening system provided by the embodiment of the invention, the influence of individual voice difference on the screening result is effectively overcome by establishing the individual dynamic voiceprint base line; multi-level acoustic feature analysis is adopted, so that the accuracy of disease recognition is improved; continuous monitoring and dynamic risk assessment are realized, and possibility is provided for early lesion discovery; on the premise of protecting user privacy, model optimization is realized through federal learning; an objective and quantitative auxiliary diagnosis basis is provided for clinicians, and the diagnosis and treatment efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voiceprint recognition, and in particular to an early screening system for ear-nose-throat diseases. BACKGROUND

[0002] Currently, ear-nose-throat disease detection mainly adopts a combination of visual examination, endoscopic examination and imaging examination. Doctors perform basic visual examination through instruments such as forehead mirrors, observe deep tissue structures using laryngoscopes and nasal endoscopes, and assess lesion ranges with the help of imaging technologies such as CT and MRI when necessary. Although these methods can accurately diagnose obvious lesions, they rely on the experience of doctors and are difficult to achieve early screening.

[0003] The examination process itself includes preoperative preparation, local anesthesia, instrument disinfection and other links, and a single examination takes about 15-30 minutes, which is difficult to complete screening quickly when the outpatient volume is large.

[0004] In addition, the examination results need to be interpreted and analyzed by doctors, further prolonging the diagnosis cycle. This time delay may result in early lesions not being discovered in time, especially for fast-progressing conditions such as vocal cord nodules and early inflammation, and the delay in diagnosis timing may directly affect the treatment effect. SUMMARY

[0005] The present application aims to at least partially solve one of the technical problems in the related art.

[0006] To this end, the first purpose of the present application is to provide an early screening method and system for ear-nose-throat diseases, which effectively overcomes the influence of individual voice differences on screening results by establishing a personal dynamic voiceprint baseline, improves the accuracy of disease recognition by using multi-level acoustic feature analysis, realizes continuous monitoring and dynamic risk assessment, provides the possibility for early lesion discovery, realizes model optimization through federated learning under the premise of protecting user privacy, and provides objective and quantitative auxiliary diagnosis basis for clinical doctors, and improves diagnosis and treatment efficiency.

[0007] To achieve the above purpose, the first aspect of the present application proposes an early screening method for ear-nose-throat diseases, comprising the following steps: Voice collection step: acquiring standardized voice samples of a user through an audio collection device; Feature extraction step: extracting multi-level acoustic features including vocal cord vibration features, resonance features, pronunciation stability features and deep representation features from the voice samples; Baseline establishment step: establishing a personal dynamic baseline model based on historical voice data of the user, the model learning the biological rhythm change pattern of the user; Risk assessment step: comparing and analyzing the current voice features with the personal dynamic baseline model, calculating the voiceprint health deviation degree and generating a risk assessment result; a result outputting step of generating a graded health suggestion and a screening report based on the risk assessment result.

[0008] In addition, the early screening system for ear-nose-throat diseases according to the above-mentioned application can further have the following additional technical features: Further, the feature extraction step comprises: The collected original speech signal is sequentially preprocessed by pre-emphasis, frame windowing and endpoint detection; based on the preprocessed signal, the absolute fundamental frequency perturbation formula is used to calculate the fundamental frequency perturbation, the absolute amplitude perturbation formula is used to calculate the amplitude perturbation, and the autocorrelation method is used to estimate the harmonic-to-noise ratio in the time domain; The resonance characteristics are extracted by linear predictive coding analysis, including solving the reciprocal of the roots of the LPC polynomial to estimate the resonance peak frequency and bandwidth, and calculating the spectral tilt; Based on the long-time-scale speech segment, the standard deviation and coefficient of variation of pitch and the standard deviation and coefficient of variation of intensity are calculated as the pronunciation stability features; The pre-trained model is used to extract deep representation features, and the acoustic embedding vector of fixed dimension is obtained by extracting its context representation and performing time average pooling.

[0009] Further, based on multiple speech samples collected by the user at different time periods in the initial health state, the statistical mean of each acoustic feature is calculated to establish an initial personal voiceprint baseline model; Through Gaussian process regression or periodic function fitting, the speech data continuously collected by the user in the healthy state is modeled; The baseline model is updated using a time-weighted sliding window mechanism; A data quality evaluation system is established, and the data quality score is calculated according to the signal-to-noise ratio, the recording environment stability and the pronunciation stability; Based on the data quality score and the user's medical feedback result, the baseline learning rate is dynamically adjusted, the learning rate is increased when the data quality score is higher than the threshold and the medical feedback confirms the screening accuracy, and the learning rate is decreased when a false alarm occurs.

[0010] Further, the risk assessment step uses a multi-dimensional risk assessment model, which includes: Instantaneous risk assessment: based on the deviation degree of single detection and personal baseline; Trend risk assessment: based on the change trend of continuous monitoring data; Pattern risk assessment: based on the matching degree of voiceprint features and disease feature library; A multi-level early warning fusion mechanism is established to weight and fuse the risk assessment results of the three dimensions, with the pattern risk assessment having the highest weight, the instantaneous risk assessment having the lowest weight, to generate a comprehensive risk score and a corresponding early warning level.

[0011] Further, a privacy protection step is further included, wherein, Acoustic feature extraction is completed locally at the user equipment end; Model updating is performed using federated learning technology; Shared data is protected using differential privacy technology.

[0012] Further, feature decoupling technology is used to separate identity features and pathological features in voiceprints, and the pathological feature extractor is made insensitive to speaker identity through adversarial training.

[0013] The second aspect embodiment of the present application proposes an ear-nose-throat disease early screening system for the first aspect to propose the ear-nose-throat disease early screening method, comprising: A voice collection module configured to collect standardized voice data; A feature extraction module configured to extract multi-level acoustic features; A personal baseline learning module configured to establish and update a personal dynamic baseline model; A risk assessment module configured to calculate a health risk level; A result generation module configured to generate a screening report and health advice.

[0014] In addition, the ear-nose-throat disease early screening system proposed in the above application can have the following additional technical features: Further, the personal baseline learning module comprises: A context-aware normalization unit configured to normalize acoustic features according to recording time and user state; A dynamic updating unit configured to implement progressive baseline updating based on a credit scoring system; and a credit scoring unit configured to evaluate data quality and adjust learning parameters.

[0015] Further, the risk assessment module comprises: A multi-modal evidence fusion unit configured to fuse acoustic, behavioral, and contextual evidence; A real-time warning engine configured to detect high-risk patterns and generate warnings; A treatment effect monitoring unit configured to quantitatively evaluate changes in voiceprints before and after treatment.

[0016] Beneficial effects: The ear-nose-throat disease early screening method and system of the embodiments of the present application effectively overcome the influence of individual voice differences on screening results by establishing a personal dynamic voiceprint baseline; multi-level acoustic feature analysis is used to improve the accuracy of disease recognition; continuous monitoring and dynamic risk assessment are realized to provide the possibility of early lesion detection; under the premise of protecting user privacy, model optimization is realized through federated learning; objective and quantitative auxiliary diagnosis basis is provided for clinicians to improve diagnosis and treatment efficiency.

[0017] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter in the description of the application. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description of embodiments, taken in conjunction with the accompanying drawings, in which: Figure 1 is a flow chart of the ear-nose-throat disease early screening method according to the present application; Figure 2 is a system block diagram of the ear-nose-throat disease early screening system according to the present application. DETAILED DESCRIPTION

[0019] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application. On the contrary, the embodiments of the present application include all changes, modifications and equivalents falling within the spirit and scope of the appended claims.

[0020] The ear-nose-throat disease early screening method of the embodiments of the present application is described below in conjunction with the accompanying drawings, including the following steps: Figure 1 Voice collection step: obtaining standardized voice samples of the user through an audio collection device; Feature extraction step: extracting multi-level acoustic features including vocal cord vibration features, resonance features, pronunciation stability features and depth representation features from the voice samples; Baseline establishment step: establishing a personal dynamic baseline model based on the user's historical voice data, and the model learns the user's biological rhythm change pattern; Risk assessment step: comparing and analyzing the current voice features with the personal dynamic baseline model, calculating the voiceprint health deviation degree and generating the risk assessment result; Result output step: generating a graded health suggestion and a screening report based on the risk assessment result.

[0021] In the specific implementation process:​ Voice acquisition step.

[0022] The system first acquires standardized voice samples through the voice acquisition module. This module is equipped with 24-bit precision, 48 kHz sampling rate audio acquisition hardware. In an environment with background noise below 50 dB, the user is guided to complete three standardized pronunciation tasks: 3-second sustained vowel / a / pronunciation, reading of the standard text "Today the weather is good, let's go to the park for a walk", and complete scale do-re-mi-fa-sol singing. The system monitors the signal-to-noise ratio in real time to ensure that the acquisition quality meets the requirements.

[0023] Feature extraction step.

[0024] The acquired voice data is then sent to the feature extraction module for multi-level analysis. This module runs four feature extraction units in parallel: the vocal cord vibration feature unit calculates the fundamental frequency perturbation and amplitude perturbation; the resonance feature unit extracts the formant parameters through 14-order LPC analysis; The pronunciation stability unit analyzes the pitch and intensity changes of long-time speech segments; the depth representation unit uses the Wav2Vec 2.0 model to extract 768-dimensional acoustic embedding features. After standardization, all features form a complete acoustic feature vector.

[0025] Baseline establishment step.

[0026] The personal baseline learning module is responsible for establishing a personalized reference model. The system first collects the user's health status voice data for 7 consecutive days, three times a day, and establishes an initial baseline through Gaussian distribution modeling. Further, the module divides a day into six time periods and establishes feature distribution models for each period to adapt to the user's biological rhythm changes. During continuous use, the module dynamically updates the baseline using a sliding window mechanism, with a window size of the last 30 valid recordings, and sets a decay factor to ensure that the model adapts to normal physiological changes in a timely manner.

[0027] Risk assessment step.

[0028] The risk assessment module generates evaluation results through multi-dimensional analysis. First, calculate the Mahalanobis distance between the current acoustic features and the personal baseline to quantify the feature deviation. Further, comprehensive evaluation is carried out from three dimensions: immediate risk is based on single detection results, trend risk analyzes the feature change rule in the past 7 days, and pattern risk is achieved through matching degree analysis with the disease feature library. Finally, the system integrates the scores of each dimension into a comprehensive risk value of 0-100 points, and divides it into four risk levels.

[0029] Result output step.

[0030] The result generation module outputs a personalized report according to the risk assessment result. The system generates a complete report containing a risk score, an abnormal feature analysis and specific suggestions, displays the trend of feature changes through visual charts, and provides targeted health guidance combined with the user's personal information.

[0031] In an embodiment of the present application, the feature extraction step includes: pre-processing the collected original speech signal in sequence by pre-emphasis, frame windowing and endpoint detection; based on the pre-processed signal, calculating the pitch perturbation by the absolute pitch perturbation formula, calculating the amplitude perturbation by the absolute amplitude perturbation formula, and estimating the harmonic-to-noise ratio in the time domain by the autocorrelation method; Extracting the resonance feature through linear predictive coding analysis, including solving the reciprocal of the roots of the LPC polynomial to estimate the resonance peak frequency and bandwidth, and calculating the spectral tilt; Based on the long time scale of the speech segment, the standard deviation and coefficient of variation of the pitch and the standard deviation and coefficient of variation of the intensity are calculated as the pronunciation stability features; Using a pre-trained model to extract deep representation features, by extracting its context representation and performing time average pooling, a fixed-dimensional acoustic embedding vector is obtained.

[0032] In specific embodiments, the feature extraction step realizes the complete extraction of multi-level acoustic features through a systematic signal processing procedure. First, the collected original speech signal is standardized and pre-processed, and a first-order FIR high-pass filter is used for pre-emphasis to enhance high-frequency components, with a transfer function of ; Then, frame windowing processing is performed, with a frame length of 25 milliseconds, a frame shift of 10 milliseconds, and a Hamming window function is used to reduce spectral leakage; finally, endpoint detection is completed through a double-threshold method based on short-time energy and zero-crossing rate, accurately identifying the range of valid speech segments.

[0033] Based on the pre-processed speech signal, the system performs multi-dimensional feature extraction in parallel. In terms of vocal cord vibration features, the pitch period is estimated by the autocorrelation function method, the absolute pitch perturbation formula is used to calculate the pitch perturbation parameters, and the specific formula is , where is the th pitch period; at the same time, the amplitude perturbation is calculated by the absolute amplitude perturbation formula, the formula is , where is the amplitude peak value of the th period; and the harmonic-to-noise ratio is estimated in the time domain by the autocorrelation method, quantifying the energy ratio of periodic components and non-periodic noise components.

[0034] In terms of resonance feature analysis, the system uses a 14th-order linear predictive coding analysis to extract vocal tract resonance characteristics, by solving the roots of the LPC polynomial The root of the equation is calculated, and the reciprocal of the root is calculated to obtain the frequency and bandwidth parameters of the first three resonance peaks. Meanwhile, the spectral tilt is obtained by calculating the logarithm of the energy ratio of the low frequency band of 0-1 kHz to the high frequency band of 1-2 kHz. For the evaluation of pronunciation stability, the system calculates the standard deviation and coefficient of variation of the fundamental frequency sequence as the pitch stability index based on long speech segments lasting more than 5 seconds, calculates the standard deviation and coefficient of variation of the short-time energy sequence as the intensity stability index, and calculates the pronunciation rate through syllable boundary detection.

[0035] Finally, the system uses the pre-trained Wav2Vec 2.0 model to extract high-level acoustic features, inputs the preprocessed speech into the model to extract its context representation, and performs average pooling operation in the time dimension to convert the variable-length sequence into a fixed-dimensional acoustic embedding vector of 128 dimensions as a deep representation feature.

[0036] In an embodiment of the present application, based on multiple speech samples collected by the user at different time periods in the initial health state, the statistical mean of each acoustic feature is calculated to establish an initial personal voiceprint baseline model; Through Gaussian process regression or periodic function fitting, the speech data continuously collected by the user in the healthy state is modeled to learn the biological rhythm change pattern of the acoustic features within 24 hours; A time-weighted sliding window mechanism is used to update the baseline model, wherein for the th historical data point in the window, its weight is determined by the formula , wherein is the current time, is the collection time of the data point, and is a pre-set weight decay factor; A data quality evaluation system is established, and the data quality score is calculated according to the signal signal-to-noise ratio, recording environment stability and pronunciation stability; Based on the data quality score and the user's medical feedback result, the learning rate of the baseline is dynamically adjusted, and when the data quality score is higher than the threshold and the medical feedback confirms the accuracy of the screening, the learning rate is increased, and when a false alarm occurs, the learning rate is decreased.

[0037] A data quality evaluation system is established, and the data quality score Q of each speech sample is calculated according to the signal signal-to-noise ratio, the background noise level of the recording environment and the stability index of the pronunciation; Based on the data quality score Q and the user's subsequent medical feedback result, the learning rate η of the baseline model update is dynamically adjusted, wherein when Q is higher than the quality threshold and the medical feedback confirms the accuracy of the screening, η is increased to accelerate the model adaptation; When a false alarm occurs, η is decreased to enhance the stability of the model.

[0038] In an embodiment of the present application, the risk assessment step adopts a multi-dimensional risk assessment model, including: Instant risk assessment: based on the deviation degree of single detection from the personal baseline; Trend risk assessment: based on the change trend of continuous monitoring data; Pattern risk assessment: based on the matching degree of voiceprint features and disease feature library; A multi-level early warning fusion mechanism is established to weight and fuse the risk assessment results of the three dimensions, among which the weight of pattern risk assessment is the highest and the weight of instant risk assessment is the lowest, to generate a comprehensive risk score and a corresponding early warning level.

[0039] Specifically, the risk assessment step realizes accurate disease risk judgment by constructing a multi-dimensional collaborative evaluation model. The system first performs instant risk assessment, uses Mahalanobis distance algorithm to calculate the statistical distance of the current acoustic feature vector and the personal dynamic baseline model in the multi-dimensional feature space, and marks it as abnormal when the deviation degree exceeds the 95% confidence interval established based on historical health data.

[0040] In the trend risk assessment stage, the system uses the autoregressive integrated moving average model to analyze the monitoring data of the past 30 days, calculates the change slope and persistence index of the key acoustic features, and identifies the potential pathological progression pattern. When the change slope of the fundamental frequency perturbation is more than 0.05 and the harmonic-to-noise ratio 7-day moving average change rate is less than -0.02 for more than 5 days, a trend warning is triggered. For pattern risk assessment, the system uses a deep matching network based on multi-head attention mechanism to perform multi-scale similarity calculation between the current voiceprint features and the pre-trained disease feature library, and outputs the matching probability of a specific disease pattern through the softmax function. When the matching probability of any disease pattern exceeds 0.7, it is determined that there is a pattern risk.

[0041] Finally, the system establishes a multi-level early warning fusion mechanism, adopts a dynamic weight distribution strategy based on confidence, sets the initial weight of pattern risk assessment to 0.5, the weight of trend risk assessment to 0.3, and the weight of instant risk assessment to 0.2, and dynamically adjusts the weights according to the confidence scores of the evaluation results of each dimension. A comprehensive risk score of 0-100 is generated by weighted summation, and corresponding early warning levels are divided into low risk (0-30 points), medium risk (31-60 points), high risk (61-80 points) and critical risk (81-100 points), thus constructing a progressive evaluation system from instant anomaly detection to long-term trend analysis and disease pattern recognition.

[0042] In an embodiment of the present application, it also includes a privacy protection step, wherein, Acoustic feature extraction is completed locally at the user device end; Federal learning technology is used for model updating; Shared data is protected using differential privacy techniques.

[0043] Specifically, privacy protection is achieved through a multi-level technical solution to complete the data security closed loop. The system first completes the localization extraction of all acoustic features on the user device side, and the original speech data is immediately securely erased after feature extraction, only retaining the desensitized feature vector, ensuring that sensitive speech information never leaves the user device.

[0044] In the model updating phase, a federated learning framework is adopted, each terminal device trains the model using local data and generates parameter updates, and transmits the model updates (rather than the original data) to the central server through a secure encryption channel. The server side uses a secure aggregation algorithm to aggregate the model updates from multiple users, where a single user update cannot be parsed before aggregation, thus forming a collaborative learning mode that shares knowledge without sharing data. To further enhance the protection strength, the system introduces differential privacy technology in the federated learning process, adding an accurately calibrated Laplace noise to the model updates, accurately balancing privacy protection and model utility through a privacy budget ε (controlled within the range of 1.0-3.0), while using gradient clipping technology to limit the potential impact of a single user on the global model.

[0045] In an embodiment of the present application, feature decoupling technology is used to separate the identity features and pathological features in the voiceprint, and the pathological feature extractor is made insensitive to the speaker identity through adversarial training.

[0046] Specifically, (1) a double-branch feature encoding network is constructed, including an identity feature encoder and a pathological feature encoder, wherein: The identity feature encoder uses a pre-trained model based on the speaker verification task, and is trained using a generalized end-to-end loss, specifically extracting acoustic features related to the speaker identity; The pathological feature encoder uses a deep neural network based on the medical diagnosis task, and is trained using a cross-entropy loss, specifically extracting pathological features related to ear-nose-throat diseases; (2) Implement an adversarial training mechanism to make the pathological feature encoder insensitive to the speaker identity through the following steps: Connect an identity discriminator after the pathological feature encoder, which attempts to predict the speaker identity based on the pathological features; Adversarial training is achieved through a gradient reversal layer, which multiplies the gradient of the identity discriminator by a negative coefficient and then backpropagates when training the pathological feature encoder; The loss function of the identity discriminator is cross-entropy loss, and the goal is to minimize the identity recognition error; the goal of the pathological feature encoder is to maximize the error rate of the identity discriminator; (3) Adopting a feature decoupling loss function, including: a mutual information penalty term between identity features and pathological features, with the formula being: wherein is the identity feature, is the pathological feature; a feature orthogonality constraint term, requiring the identity feature vector and the pathological feature vector to be mutually orthogonal in the feature space: a pathological feature reconstruction loss, ensuring that the pathological feature can still effectively complete the disease classification task: (4) Establishing a decoupling effect evaluation mechanism to verify the feature decoupling effect through the following indicators: identity leakage rate: the accuracy rate of speaker recognition using pathological features, which is required to be lower than the random guessing level; disease classification accuracy: the accuracy rate of disease classification using pathological features, which is required to be above 85%; feature correlation coefficient: calculate the Pearson correlation coefficient between identity features and pathological features, which is required to be below 0.1; (5) The feature decoupling technology also includes a dynamic optimization strategy, which adaptively adjusts the adversarial training intensity and loss function weight according to the decoupling effect evaluation results, to ensure that the disease recognition performance is maintained while maximizing the removal effect of identity information.

[0047] The embodiments of the present application will be described below in conjunction with the accompanying Figure 2 The present application discloses an ear-nose-throat disease early screening system for implementing any step of the above ear-nose-throat disease early screening method, comprising: a voice collection module configured to collect standardized voice data; a feature extraction module configured to extract multi-level acoustic features; a personal baseline learning module configured to establish and update a personal dynamic baseline model; a risk assessment module configured to calculate a health risk level; a result generation module configured to generate a screening report and health advice.

[0048] The personal baseline learning module includes: a context-aware normalization unit configured to normalize acoustic features according to recording time and user state; a dynamic update unit configured to implement progressive baseline update based on a credit scoring system; a credit scoring unit configured to evaluate data quality and adjust learning parameters.

[0049] The risk assessment module includes: a multi-modal evidence fusion unit configured to fuse acoustic, behavioral, and contextual evidence;​​ a real-time early warning engine configured to detect high-risk patterns and generate early warnings; and a treatment effect monitoring unit configured to quantify changes in voiceprints before and after treatment.

[0050] Specifically, the voice collection module is responsible for collecting voice data meeting quality requirements, supporting multiple audio input devices. The feature extraction module adopts a hybrid architecture, combining traditional signal processing methods and deep learning techniques, to extract multi-level acoustic features including vocal cord vibration features and resonance features.

[0051] The personal baseline learning module includes three core units: a context-aware normalization unit that calibrates features based on recording time and user state; a dynamic updating unit that adjusts baseline updating strategies based on credit scores; and a credit score unit that evaluates data quality and optimizes learning parameters. This module establishes a personal voiceprint baseline through Gaussian process regression and enables continuous optimization of the baseline.

[0052] The risk assessment module integrates acoustic features, behavioral data, and environmental information through a multi-modal evidence fusion unit, and generates a comprehensive assessment result using a weighted fusion algorithm. The real-time early warning engine monitors changes in voiceprint features and triggers corresponding levels of early warnings when abnormal patterns are detected. The treatment effect monitoring unit provides a reference for efficacy evaluation by comparing changes in voiceprint features before and after treatment.

[0053] The results generation module generates structured screening reports and personalized health recommendations based on the output data from each module. The system supports report export functions to facilitate reference and use by medical professionals.

[0054] The embodiment of the present application discloses an early screening method for ear-nose-throat diseases, which is used for screening vocal cord nodules, vocal cord polyps, vocal cord paralysis, chronic laryngitis, early lesions of laryngeal cancer and resonance abnormalities related to sinusitis.

[0055] The embodiment of the present application discloses a computer readable storage medium having a computer program stored thereon, which realizes any step of the early screening method for ear-nose-throat diseases when executed by a processor.

[0056] In summary, the early screening method for ear-nose-throat diseases and its system effectively overcome the influence of individual voice differences on screening results by establishing a personal dynamic voiceprint baseline. The use of multi-level acoustic feature analysis improves the accuracy of disease recognition. Continuous monitoring and dynamic risk assessment make early lesion detection possible. On the premise of protecting user privacy, model optimization is achieved through federated learning. Objective and quantitative auxiliary diagnostic evidence is provided for clinicians to improve diagnosis and treatment efficiency.

[0057] In the description of the specification, the terms "first", "second", "third", etc. are used only to describe various features, and are not to be construed as indicating or implying relative importance or a specific order of limiting the indicated technical features. Therefore, the features defined with "first", "second", etc. can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0058] In the description of the specification, the description referring to the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. Furthermore, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples, without contradiction.

[0059] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. An early screening method for ear-nose-throat diseases, characterized in that, The method comprises the following steps: a voice collection step: obtaining a standardized voice sample of a user through an audio collection device; a feature extraction step: extracting multi-level acoustic features including vocal cord vibration features, resonance features, pronunciation stability features, and depth representation features from the voice sample; a baseline establishment step: establishing a personal dynamic baseline model based on historical voice data of the user, which learns the biological rhythm change pattern of the user; a risk assessment step: comparing and analyzing the current voice features with the personal dynamic baseline model, calculating the voiceprint health deviation degree, and generating a risk assessment result; a result output step: generating a graded health suggestion and a screening report based on the risk assessment result.

2. The method of claim 1, wherein the method is used for early screening of an ear, nose, and throat disease. The feature extraction step comprises: performing preprocessing on the collected original voice signal in sequence, including pre-emphasis, frame windowing, and endpoint detection; based on the preprocessed signal, calculating the fundamental frequency perturbation through the absolute fundamental frequency perturbation formula, calculating the amplitude perturbation through the absolute amplitude perturbation formula, and estimating the harmonic-to-noise ratio in the time domain using the autocorrelation method; extracting resonance features through linear predictive coding analysis, including solving the reciprocal of the roots of the LPC polynomial to estimate the resonance peak frequency and bandwidth, and calculating the spectral tilt; based on long-time-scale voice segments, calculating the standard deviation and coefficient of variation of pitch and the standard deviation and coefficient of variation of intensity as pronunciation stability features; extracting depth representation features using a pre-trained model, extracting its context representation, and performing time average pooling to obtain a fixed-dimensional acoustic embedding vector.

3. The early screening method for ear-nose-throat diseases according to claim 1, wherein based on multiple voice samples collected by the user at different time periods in an initial healthy state, statistical means of each acoustic feature are calculated to establish an initial personal voiceprint baseline model; through Gaussian process regression or periodic function fitting, voice data collected continuously by the user in a healthy state are modeled; a sliding window mechanism based on time weighting is used to update the baseline model; a data quality evaluation system is established, and data quality scores are calculated based on signal-to-noise ratio, recording environment stability, and pronunciation stability; based on the data quality score and the user's medical feedback result, the baseline learning rate is dynamically adjusted; when the data quality score is higher than the threshold and the medical feedback confirms the screening accuracy, the learning rate is increased; when a false alarm occurs, the learning rate is decreased.

4. The method of claim 1, wherein the method is used for early screening of an ear, nose, and throat disease. characterized in that the risk assessment step uses a multi-dimensional risk assessment model, including: instant risk assessment: based on the deviation degree of single detection and personal baseline; trend risk assessment: based on the change trend of continuous monitoring data; pattern risk assessment: based on the matching degree of voiceprint features and disease feature library; a multi-level early warning fusion mechanism is established to weight and fuse the risk assessment results of the three dimensions, wherein the weight of pattern risk assessment is the highest, and the weight of instant risk assessment is the lowest, to generate a comprehensive risk score and a corresponding early warning level.

5. The method of claim 1, wherein the method is used for early screening of an ear, nose and throat disease. It also includes a privacy protection step, wherein acoustic feature extraction is completed locally on the user device; federated learning technology is used for model updating; difference privacy technology is used to protect shared data.

6. The method of claim 1, wherein the method is used for early screening of an ear, nose and throat disease. Identity features are decoupled from pathological features in voiceprints using feature decoupling techniques, and an adversarial training is used to make the pathological feature extractor insensitive to speaker identity.

7. An early screening system for ear-nose-throat diseases, characterized in that, A method for implementing any of claims 1-6, comprising: a voice collection module configured to collect standardized voice data; a feature extraction module configured to extract multi-level acoustic features; a personal baseline learning module configured to establish and update a personal dynamic baseline model; a risk assessment module configured to calculate a health risk level; a result generation module configured to generate a screening report and health recommendations.

8. The early screening system for otolaryngological diseases according to claim 7, characterized in that, The personal baseline learning module includes: a context-aware normalization unit configured to normalize acoustic features according to recording time and user state; a dynamic update unit configured to implement gradual baseline updates based on a credit scoring system; a credit scoring unit configured to evaluate data quality and adjust learning parameters. 9.The early screening system for otolaryngological diseases according to claim 7, characterized in that, The risk assessment module includes: a multi-modal evidence fusion unit configured to fuse acoustic, behavioral, and contextual evidence; a real-time warning engine configured to detect high-risk patterns and generate warnings; a treatment effect monitoring unit configured to quantitatively assess voiceprint changes before and after treatment.

Images