Metric detection method, metric detection device, metric detection system, trained model, and program

Abstract

The present invention aims to provide a method, apparatus, system, etc., for detecting indicators of acoustic neuroma by standardizing audiogram images or their constituent data and inputting them into an AI model. [Solution] An index detection method comprising: a storage step in which a computer stores images of audiogram-related information or their constituent data, which are information relating to an audiogram obtained about a target person, in a database; a standardization step in which data obtained from the audiogram images or the constituent data are standardized; an index detection step in which the standardized audiogram-related information is used as input to a trained model obtained by performing a learning process using data which is a combination of previously acquired and standardized audiogram-related information and an index of acoustic neuroma for that person, thereby detecting an index of acoustic neuroma for the target person; and a detection step in which the data is input to an AI model to detect an index of acoustic neuroma.

Description

Technical Field

[0001] The present invention relates to a method, apparatus, system, or program for detecting an index of acoustic neuroma from an audiogram, a learned model for detecting an acoustic neuroma index, and a learned model for discriminating acoustic neuroma sensorineural hearing loss, etc.

Background Art

[0002] An audiogram is typically a display of an audiometry (hearing test device) output using various audiometers. For example, the hearing threshold of a subject over the entire frequency spectrum is visually represented as an inverted graph, and a plot of the threshold of hearing in the left and right ears as a function of frequency is used. These plots usually include data points for air conduction and bone conduction of the right ear and data points for air conduction and bone conduction of the left ear, and each data point may be connected by a line segment arbitrarily. The shape (curve shape), symmetry (relationship between curves), and severity (position on the Y-axis) of hearing loss are very valuable information regarding the potential cause of hearing loss and can be important in virtually all hearing tests. Since the shape, severity, symmetry, etc. of the hearing curve are characteristic of a specific pathological condition, they may be useful for diagnosis.

[0003] In order to assist in the diagnosis of hearing diseases, an audiogram classification system has been developed. The classification system of Patent Document 1 includes categories of the configuration, severity, lesion site, and / or symmetry of the audiogram, and provides a series of rules for selecting categories.

[0004] Non-Patent Document 1 discloses a data-driven audiogram classification system for mobile hearing measurement. Focusing on the configuration (shape), symmetry, and severity of the audiogram, machine learning techniques (decision trees) are applied to the classification for audiogram interpretation.

[0005] Patent Document 2 provides "an automatic identification and determination method for pure tone audiograms, comprising the steps of: acquiring a pure tone audiogram and corresponding hearing data; labeling the categories and positions corresponding to the symbols in the pure tone audiogram to obtain a labeled audiogram; inputting the labeled audiogram into a Faster-RCNN for learning to identify the categories and positions of the symbols in the pure tone audiogram to be determined; and generating an audiogram recognition report that combines the category of each symbol and the position of each symbol based on the pure tone audiogram."

[0006] Patent Document 2 describes an AI model, Faster-RCNN, which is trained by annotating (labeling) an audiogram. The target audiogram is then fed through the trained Faster-RCNN to create an audiogram recognition report that combines the category and position of each symbol. The report includes the output audiogram at a fixed size and a judgment on whether hearing is normal or not at each frequency. This judgment is made by comparing the hearing data points at each frequency with standard hearing levels such as those set by the World Health Organization (WHO).

[0007] Diagnosing acoustic neuromas (vestibular schwannomas) from audiograms is difficult, and MRI is used for definitive diagnosis. Non-patent document 2 focuses on the sensitivity, specificity, and cost-effectiveness analysis of audiometry criteria for asymmetric sensorineural hearing loss (ASNHL) when used to determine when to perform MRI. It uses a decision tree to make judgments based on the structure (shape) of the audiogram.

[0008] Non-patent document 3 describes a protocol that may be used in current clinical practice. In particular, it predicts acoustic neuroma by focusing on the asymmetry of the audiogram.

[0009] Non-patent document 4 describes the detection of tumors from audiograms using rule-based and machine learning models that distinguish between the general patient population and patients with acoustic neuroma.

[0010] Currently, even ENT specialists find it very difficult to diagnose acoustic neuroma based solely on audiograms, and MRI scans are performed on suspected patients before a diagnosis is made. [Prior art documents] [Patent Documents]

[0011] [Patent Document 1] U.S. Patent No. 8075494 [Patent Document 2] Chinese Patent Application Publication No. 115223704 Specification [Non-patent literature]

[0012] [Non-Patent Document 1] Charih F, Bromwich M, Mark AE, Lefrancois R, Green JR. Data-Driven Audiogram Classification for Mobile Audiometry. Sci Rep. 2020 Mar 3;10(1):3962. [Non-Patent Document 2] Conley M, Diaz RC. Asymmetric sensorineural hearing loss and vestibular schwannoma: when to image? Curr Opin Otolaryngol Head Neck Surg. 2020 Oct;28(5):335-339. [Non-Patent Document 3] Gimsing, S. Vestibular schwannoma: when to look for it? J Laryngol Otol 124,258-64 (2010). [Non-Patent Document 4] Carey, GE et al. Machine Learning for Vestibular Schwannoma Diagnosis Using Audiometrie Data Alone. Otol Neurotol 43, e530-e534 (2022). [Overview of the project] [Problems that the invention aims to solve]

[0013] However, Patent Document 1 and Non-Patent Documents 1-4 do not read the data points of the original audiogram to detect indicators of hearing disorders, but rather classify them according to categories such as the composition (shape), symmetry, and severity of the audiogram. Furthermore, they do not standardize the audiogram data and provide it to an AI model to classify the likelihood of hearing disorders.

[0014] Patent Document 2 describes an AI model that reads data points from the original audiogram, displays an audiogram of a fixed size in the report, and determines whether hearing is normal at each frequency. However, it does not involve inputting an image or numerical value into the AI ​​model in which "the audiogram frame is fixed at a predetermined value, and each hearing data point is shown relative to the frame" to detect indicators of hearing diseases, including acoustic neuromas. Furthermore, it does not suggest or disclose that indicators or the possibility of acoustic neuromas can be detected.

[0015] Non-patent document 2 applies the perspective of asymmetric sensorineural hearing loss to a protocol for considering MRI examination for acoustic neuroma and uses a decision tree to determine the necessity of MRI. However, it does not suggest or describe using an AI model that focuses on the "absolute value of the difference between each air conduction data point on the left and right sides" from an audiogram to detect indicators of acoustic neuroma. Furthermore, it does not suggest or disclose inputting data points into the AI ​​model such that each air conduction data point includes at least one data point at 2,000 Hz and 3,000 Hz.

[0016] Furthermore, Patent Documents 1 to 2 and Non-Patent Documents 1 to 4 do not describe or suggest performing a sensorineural hearing loss discrimination step of discriminating sensorineural hearing loss caused by acoustic neuroma from sensorineural hearing loss caused by other than acoustic neuroma from an audiogram using an AI model. Also, there is no suggestion or disclosure that the standardized data includes at least one of the air conduction data points at 125 Hz and 250 Hz.

[0017] Furthermore, there is no suggestion or disclosure that the performance of the AI model is improved by the standardized data including air conduction data points and bone conduction data points.

[0018] As described above, even an otolaryngologist cannot diagnose acoustic neuroma only from an audiogram. It is not efficient for a human to detect the possibility of acoustic neuroma from an audiogram with high accuracy using an AI model or to identify cases of sensorineural hearing loss caused by a tumor from sensorineural hearing loss cases. It shows that the present AI method, device, and system can perform qualitatively different operations.

[0019] An object of the present invention is to provide, for example, a method, device, system, or program for detecting an index of acoustic neuroma from an audiogram, a learned hearing disorder index detection model, a learned AI acoustic neuroma sensorineural hearing loss discrimination model, and the like.

Means for Solving the Problems

[0020] In order to achieve the above object, embodiments of the present invention provide the following (1) to (28).

[0021] (1) One aspect of the present invention is an index detection method including a normalization step of normalizing audiogram-related information, which is information regarding an audiogram obtained for a target person, and an index detection step of detecting an index of acoustic neuroma of the target person by using the normalized audiogram-related information as an input to a learned model obtained by performing learning processing using data that is a combination of pre-acquired normalized audiogram-related information and an index of acoustic neuroma of the person.

[0022] (2) One aspect of the present invention is the index detection method according to (1), wherein in the normalization step, the frame of the audiogram is fixed at a predetermined value, and each hearing data point is normalized by indicating it with a numerical value or a figure of a relative position with respect to the frame.

[0023] (3) One aspect of the present invention is the index detection method according to (1), including a sensorineural hearing loss discrimination step of using, as an input to a learned model obtained by performing a learning process using data that is a combination of pre-acquired and normalized audiogram-related information and information indicating whether the person has sensorineural hearing loss caused by a vestibular schwannoma or sensorineural hearing loss caused by a cause other than a vestibular schwannoma, the normalized audiogram-related information obtained for the target person, to discriminate whether the target person has sensorineural hearing loss caused by a vestibular schwannoma or sensorineural hearing loss caused by a cause other than a vestibular schwannoma.

[0024] (4) One aspect of the present invention is the index detection method according to (3), wherein in the sensorineural hearing loss discrimination step, the normalized audiogram-related information includes at least one of the air-conduction data points at 125 Hz and 250 Hz.

[0025] (5) One aspect of the present invention is the index detection method according to (4), wherein the reproducibility of the discrimination in the sensorineural hearing loss discrimination step is 0.78 or more.

[0026] (6) One aspect of the present invention is the index detection method according to (4), wherein the reproducibility of the discrimination in the sensorineural hearing loss discrimination step is greater than 0.80.

[0027] (7) One aspect of the present invention is the index detection method according to (4), wherein the reproducibility of the discrimination in the sensorineural hearing loss discrimination step is 0.89 or more.

[0028] (8) One aspect of the present invention is the index detection method according to (4), wherein the normalized audiogram-related information includes air-conduction data points and bone-conduction data points.

[0029] (9) One aspect of the present invention is the method for detecting an indicator of acoustic neuroma in the subject person, wherein the recall rate for detecting the indicator is 0.70 or higher and the specificity is greater than 0.22, as described in (1).

[0030] (10) One aspect of the present invention is the method for detecting an indicator of acoustic neuroma in the subject person, wherein the recall rate for detecting the indicator is 0.80 or higher and the specificity is greater than 0.22.

[0031] (11) One aspect of the present invention is the method for detecting an indicator of acoustic neuroma in the subject person, wherein the recall rate for detecting the indicator is 0.90 or higher and the specificity is greater than 0.22.

[0032] (12) One aspect of the present invention is the index detection method described in (9), wherein the standardized audiogram-related information includes the absolute value of the difference between each left and right air conduction data point.

[0033] (13) One aspect of the present invention is the index detection method according to (12), wherein each air conduction data point includes at least one data point of about 2,000 Hz and about 3,000 Hz.

[0034] (14) One aspect of the present invention is an index detection device comprising: a standardization unit that standardizes audiogram-related information, which is information relating to an audiogram obtained for a target person; and a detection unit that detects an index of an acoustic neuroma of the target person by using the standardized audiogram-related information as input to a trained model obtained by performing a learning process using data which is a combination of previously acquired and standardized audiogram-related information and an index of an acoustic neuroma of that person.

[0035] (15) One aspect of the present invention is an index detection device according to (14) in which the detection unit performs a learning process using data which is a combination of pre-acquired and standardized audiogram-related information and information indicating whether the person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma, and uses the standardized audiogram-related information obtained for the target person as input to the trained model, thereby identifying whether the target person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma.

[0036] (16) One aspect of the present invention is the index detection device according to (15), wherein the sensorineural hearing loss identification unit includes at least one of the standardized audiogram-related information of air conduction data points at 125 Hz and 250 Hz.

[0037] (17) One aspect of the present invention is the index detection device described in (16), wherein the recall rate of the identification in the sensorineural hearing loss identification unit is 0.78 or higher.

[0038] (18) One aspect of the present invention is the index detection device described in (17), wherein the standardized audiogram-related information includes air conduction data points and bone conduction data points.

[0039] (19) One aspect of the present invention is an index detection system comprising: a standardization unit that standardizes audiogram-related information, which is information relating to an audiogram obtained for a target person; and a detection unit that detects an index of an acoustic neuroma in the target person by using the standardized audiogram-related information as input to a trained model obtained by performing a learning process using data which is a combination of previously acquired and standardized audiogram-related information and an index of an acoustic neuroma in that person.

[0040] (20) One aspect of the present invention is an index detection system according to (19) in which the detection unit performs a learning process using data which is a combination of pre-acquired and standardized audiogram-related information and information indicating whether the person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma, and uses the standardized audiogram-related information obtained for the target person as input to the trained model, thereby identifying whether the target person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma.

[0041] (21) One aspect of the present invention is the index detection system according to (20), wherein the sensorineural hearing loss identification unit includes at least one of the standardized audiogram-related information of air conduction data points at 125 Hz and 250 Hz.

[0042] (22) One aspect of the present invention is the index detection system described in (21), wherein the recall rate of the identification in the sensorineural hearing loss identification unit is 0.78 or higher.

[0043] (23) One aspect of the present invention is the index detection system described in (22), wherein the standardized audiogram-related information includes air conduction data points and bone conduction data points.

[0044] (24) One aspect of the present invention is a trained model used in an index detection method, index detection device, or index detection system described in (1), (14), or (19), comprising an input layer for inputting an audiogram image or its constituent data, an intermediate layer consisting of a neural network for processing the input data, and an output layer for outputting the possibility of an acoustic neuroma.

[0045] (25) One aspect of the present invention is a trained model as described in (24) for causing a computer to function to output an acoustic neuroma index in response to an input audiogram image or its constituent data.

[0046] (26) One aspect of the present invention is a trained model used in the index detection method, index detection device, or index detection system described in (1), (14), or (19), comprising: an input layer for inputting an audiogram image or its constituent data; an intermediate layer consisting of a neural network for processing the input data; and an output layer for outputting the result of distinguishing between sensorineural hearing loss caused by an acoustic neuroma and sensorineural hearing loss caused by a cause other than an acoustic neuroma.

[0047] (27) One aspect of the present invention is a trained model as described in (26) for causing a computer to output a result that identifies sensorineural hearing loss caused by an acoustic neuroma and sensorineural hearing loss caused by a cause other than an acoustic neuroma, in response to an input audiogram image or its constituent data.

[0048] (28) One aspect of the present invention is a program for causing a computer to function as an indicator detection device, comprising: a standardization step of standardizing audiogram-related information, which is information relating to an audiogram obtained for a target person; and a detection step of detecting an indicator of an acoustic neuroma in the target person by using the standardized audiogram-related information as input to a trained model obtained by performing a learning process using data which is a combination of previously acquired and standardized audiogram-related information and an indicator of an acoustic neuroma in that person. [Effects of the Invention]

[0049] According to one aspect of the present invention, by standardizing the original audiogram and inputting it into an AI model, the effect is achieved that indicators of acoustic neuroma can be detected. [Brief explanation of the drawing]

[0050] [Figure 1] This is a schematic diagram of a system or apparatus according to one embodiment of the present invention. [Figure 2] This is a flowchart showing the steps of the method according to one embodiment of the present invention. [Figure 3]Figure 3 shows the mean and standard deviation (SD) of hearing thresholds (dB) at various frequencies (Hz) for different patient groups. [Figure 4] This figure shows that, according to the steps of a method according to one embodiment of the present invention, the original audiogram or a photograph thereof is fed into an AI model to detect the frame of the audiogram and each hearing data point, and that this information can be used to redraw a standardized audiogram. [Figure 5] This figure shows the training results (training / validation accuracy and training / validation loss) and the detection confusion matrix (actual value (True) and detected value (Predicted)) of an image-based deep learning model that detects indicators of acoustic neuroma by inputting a standardized audiogram in a method according to one embodiment of the present invention. [Figure 6] This figure shows the training results (training loss, validation accuracy), ROC curve, and detection confusion matrix of a multilayer perceptron (MLP) model that detects indicators of acoustic neuroma by inputting a standardized audiogram, according to one embodiment of the present invention. [Figure 7] The results of training an image AI model by inputting the original audiograms (Figures 7D and 7E) without standardization (Figures 7A and 7B), and the resulting confusion matrix for detection (Figure 7C) are shown (comparative example). [Figure 8] This figure shows the confusion matrix (Figure 8C) of the detection results of an acoustic neuroma index detection AI model in a method according to one embodiment of the present invention, which takes as input an audiogram (Figures 8A and 8B) obtained by standardizing the original audiogram by focusing on the absolute value of the difference between the left and right air conduction data points. [Figure 9] This figure shows the confusion matrix (Figure 8C) obtained by training an image AI model using plots (Figures 9A and 9B) that were created by standardizing the original audiogram and showing bone conduction (□) and air conduction data points (〇) and connecting line segments for the left and right sides, respectively, as input to an image AI model. [Figure 10] This diagram shows a confusion matrix resulting from a machine learning model in which each hearing data point was extracted from the original audiogram using an AI model, standardized, and then used to input the air conduction data points for the left and right ears as numerical values. [Figure 11]Figure 11A shows the confusion matrices for the validation dataset (Figure 11A) and test dataset (Figure 11B) obtained by training a numerical AI machine learning model using a method according to one embodiment of the present invention, in which the original audiogram was standardized numerically by focusing on the absolute value of the difference between the left and right air conduction data points and input as input. Figure 11C is a graph showing the importance of features within the optimized machine learning model. [Figure 12] This figure shows the results of implementing three types of acoustic neuroma index detection AI models according to embodiments of the present invention on a web application, inputting a target audiogram, and displaying the results detected by the AI ​​model on the web. [Figure 13] This figure shows the training results (training loss, validation accuracy) (Figure 13A), ROC curve (Figure 13B), and detection confusion matrix (Figure 13C) of a multilayer perceptron (MLP) model that identifies acoustic neuromas by inputting standardized audiogram air conduction data points from cases of sensorineural hearing loss due to tumors and cases with non-tumor causes, according to one embodiment of the present invention. [Figure 14] Figure 14A shows the confusion matrix in a test dataset of a numerical AI machine learning model that identifies acoustic neuromas by inputting standardized audiogram air conduction data points from cases of sensorineural hearing loss due to tumors and cases of non-tumor-related causes, according to one embodiment of the present invention. Figure 14B is a graph showing the importance of features within the machine learning model. [Figure 15] Figure 15A shows the confusion matrix results for a multilayer perceptron (MLP) model trained on a test dataset, using the remaining air conduction data points after excluding the low-frequency data points (125Hz, 250Hz) of standardized audiograms for cases of sensorineural hearing loss due to tumors and cases of hearing loss due to non-tumor causes. Figure 15B shows the confusion matrix results for an AI model on a test dataset, using standardized audiograms (Figures 8A and 8B) that focus on the absolute difference between the left and right air conduction data points of standardized audiograms for cases of sensorineural hearing loss due to tumors and cases of hearing loss due to non-tumor causes. [Figure 16]This figure shows the training results (training loss, validation accuracy) (Figure 16A), ROC curve (Figure 16B), and detection confusion matrix (Figure 16C) of a multilayer perceptron (MLP) model that identifies acoustic neuromas by inputting standardized air conduction data points and bone conduction data points from audiograms of cases of sensorineural hearing loss due to tumors and cases of hearing loss due to non-tumor causes, according to one embodiment of the present invention. [Figure 17] Figure 17A shows the confusion matrix results on a test dataset for a numerical AI machine learning model that identifies acoustic neuromas by inputting standardized air conduction and bone conduction data points from audiograms of cases of sensorineural hearing loss caused by tumors and cases of non-tumor-induced hearing loss, according to one embodiment of the present invention. Figure 17B is a graph showing the importance of features within the machine learning model. [Figure 18] This figure shows a schematic example of the hardware configuration of the information processing device 90 applied to this embodiment. [Modes for carrying out the invention]

[0051] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0052] Terms and Definitions In this specification, "audiogram" may refer to a graph showing the audible threshold of frequencies measured by an audiometer. The Y-axis represents intensity measured in decibels (dB), and the X-axis represents frequency measured in Hertz (Hz). The frequencies displayed in the audiogram are in octaves, and may be expressed by doubling the frequency (e.g., 125Hz, 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, 8000Hz). Commonly tested "interoctave" frequencies (e.g., 3000Hz) may also be displayed. These measured frequencies may contain an error of ±10%. The intensity displayed in the audiogram is expressed in linear 10dB steps. However, since decibels are a logarithmic scale, the volume increases significantly in 10dB increments. Pure-tone audiometry / audiogram (PTA) is usually performed, but is not limited to this method. The frame, grid lines, and values ​​at the air conduction and bone conduction hearing data points on the left and right sides are often displayed as follows: for example, red indicates measurements in the right ear, blue indicates measurements in the left ear, and data points for the right air conduction are displayed as a red "○", data points for the right bone conduction are displayed as a red "[", data points for the left air conduction are displayed as a blue "×", and data points for the left bone conduction are displayed as a blue "[]", although this is not a requirement. In this specification, "each number" and "approximately" include ±10%. ±10% may be ±0, 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10%, and may be a value between any two of the above numbers. Also, the displayed numbers may have been rounded down or rounded to the nearest decimal place.

[0053] In this specification, "audiogram frame" may refer to, but is not limited to, the outer frame of a typical audiogram. Any frame can be set. The frame may or may not include internal vertical and horizontal axis lines. The number of lines is also arbitrary and not particularly limited. Certain lines (e.g., 0dB line) may be highlighted (e.g., in bold).

[0054] In this specification, "auditory data points" may refer to, but are not limited to, the intensity or threshold of hearing at any frequency measured by any audiometric device (e.g., online or smartphone-based testing systems). They may include data points from both air conduction and bone conduction, although only air conduction or bone conduction data may be measured or displayed. Air conduction, also known as air conduction hearing, is performed using headphones, but the equipment used is not particularly limited. Bone conduction, also known as bone conduction hearing, can be performed, for example, by placing a transducer on the bone behind the ear.

[0055] In this specification, "composition data" may refer to raw data showing numerical values, etc., at each hearing data point for the left and right ears obtained from various audiometrics, but is not particularly limited to any data that constitutes such numerical values, etc.

[0056] In this specification, "image" can refer to any image format and includes printed materials, digital displays, photographs, etc. In the case of digital displays, it includes images saved in any format, such as jpeg, jpg, png, pdf, etc., which can be saved and used in their corresponding formats.

[0057] In this specification, "standardization" may be used interchangeably with "standardization," and may also be referred to as "standardization." Standardization may refer to the process of establishing a "standard" or "specification" among numerous specifications available on the market. In this specification, it may refer to extracting data from an original audiogram and providing it to an AI model as image or numerical data based on certain rules. Alternatively, it may include providing data obtained directly from a hearing measurement device, etc., to an AI model based on certain rules. An example of a rule may include, but is not limited to, the case where "standardization fixes the frame of the audiogram at a predetermined value and shows each hearing data point relative to the said frame." Here, the predetermined value may be anything that can be used during standardization (see Example 1), and is not particularly limited.

[0058] In this specification, "hearing disorders" include hearing loss. Hearing loss is broadly divided into two types based on the cause and location of the impairment: conductive hearing loss and sensorineural hearing loss. Conductive hearing loss is a type of hearing impairment that occurs when the transmission characteristics of audible sound (said to be 20Hz-20,000Hz), which is vibration of air from the outside world, are altered due to damage to one or all of the outer ear, middle ear, cochlear window, or vestibular window. Sensorineural hearing loss is a type of hearing impairment in which the energy of sound input from the outer ear or skull is converted into vibrations of the inner ear lymph fluid, but there is thought to be an organic lesion in the inner ear or the area from the inner ear to the auditory center. Examples of typical hearing disorders that cause hearing loss include, but are not limited to, acoustic trauma, perilymphatic fistula, malformations, functional hearing loss, tympanic membrane damage, Eustachian tube dysfunction, otosclerosis, ossicular transection, tumors (e.g., acoustic neuroma, brain tumor), skull fracture, noise-induced hearing loss, otitis media, toxic cochlear dysfunction, sudden hearing loss, cochlea, Meniere's disease, mumps, and age-related hearing loss.

[0059] In this specification, "acoustic neuroma," also known as vestibular schwannoma or acoustic neuroma, refers to a benign tumor that develops in the vestibulocochlear nerve, which extends from the inner ear to the brain. It is known to originate from Schwann cells that form the myelin sheath surrounding the vestibulocochlear nerve. Symptoms include tinnitus, hearing loss, dizziness, and lightheadedness. If the tumor grows and becomes severe, it can compress other nerves such as the facial nerve and trigeminal nerve, potentially causing symptoms such as facial paralysis or facial spasms. In severe cases, it can also cause gait disturbances and impaired consciousness. Because the tumor grows slowly, its impact on the brain is delayed, and symptoms are often subtle and difficult to notice.

[0060] In this specification, "patient" is interchangeable with "subject," and specifically includes "patients suspected of having an acoustic neuroma" and "patients who have an acoustic neuroma."

[0061] In this specification, “indicator” includes, in the context of disease probability as used herein, the probability or indicator used when a computer interprets an audiogram to predict the probability of disease occurring. “Method for detecting indicators” may be, for example, a method for testing or detecting the possibility of acoustic neuroma, a method for obtaining or detecting indicators of acoustic neuroma, a method for analyzing the presence or absence of acoustic neuroma, a method for supporting or assisting in the prediction of acoustic neuroma, a method for processing an audiogram image or its constituent data to detect indicators of acoustic neuroma from an audiogram, or a method for operating a system or apparatus for detecting indicators of acoustic neuroma from an audiogram. Indicators may be quantitative or qualitative.

[0062] In this specification, examples of "AI models" include, but are not limited to, machine learning models, deep learning models, and generative AI models. Examples of machine learning algorithms include, but are not limited to, k-nearest neighbors, decision trees, random forests, XGBoost, support vector machines (SVMs), logistic regression, naive Bayes, linear regression, neural networks (e.g., multilayer perceptrons (MLPs), deep learning), k-means algorithms, principal component analysis, and GANs (generative adversarial networks). AI models can be trained using supervised learning, semi-supervised learning, unsupervised learning, or reinforcement learning.

[0063] In this specification, “detection” includes the act of a computer performing detection. Detection includes, but is not limited to, the results of inferring whether or not a disease exists, or, as appropriate, the likelihood or probability of a disease existing. In this specification, “identification” may be used to mean distinguishing one disease from another, and “detection” and “identification” may be used interchangeably.

[0064] In this specification, "display" includes, but is not limited to, visual displays such as text and graphics, and auditory displays such as sound. It also includes, but is not limited to, displays on paper, displays on computer, smartphone, or tablet screens, displays on websites on the internet, and displays on screens of various devices.

[0065] In this specification, "re-depicting" includes depicting original figures, letters, or other symbols on a different medium, but does not specifically refer only to reproducing the same or substantially the same figure; it also includes depicting figures and symbols that have undergone various modifications or standardizations.

[0066] In this specification, "website" may refer to a collection of web pages located on the World Wide Web (WWW) and generally under a specific domain name. It may also simply be called a "site." However, it can also be deployed as a Single Page Application (SPA). It is not particularly limited to any site that can be viewed and operated on the internet. Furthermore, in this specification, applications (including web applications and native applications) are also included in the definition of a website.

[0067] In this specification, "subject" includes "things" and / or "concepts" such as images, people or human characteristics, objects or characteristics of objects, events, matters, phenomena, etc. Unless otherwise specified, "subject" includes data of either an "object" or a "person" or their "image" or "characteristics / features." In this specification, "subject" may include categories, etc., but may also include multiple specific content data related to the subject (e.g., individual image data).

[0068] In this specification, "statistical processing" includes, but is not limited to, any statistical processing used by a person skilled in the art (e.g., processing to obtain the mean, maximum, minimum, median, mode, variance, mean deviation, mean error, sum, recall, specificity, accuracy, precision (PPV (positive predictive value)), NPV (negative predictive value), etc.).

[0069] Embodiment 1: Method for detecting indicators of acoustic neuroma from an audiogram Embodiment 1 is, A method for detecting indicators of acoustic neuroma from an audiogram, A storage step in which a computer stores an audiogram image or its constituent data in a database, A standardization step of standardizing the data obtained from the audiogram image or the constituent data, The method includes a detection step of inputting the standardized data into an AI model to detect indicators of acoustic neuroma.

[0070] The database used is not particularly limited; any usable database can be used. Examples of database types include hierarchical, NoSQL, relational, and network databases. Specifically, examples include MySQL®, SQLite, Amazon Aurora, PostgreSQL, Oracle Database, MongoDB, MariaDB, and Microsoft SQL Server, but the invention is not limited to these. Furthermore, the database used by the display control unit to store audiogram images or their constituent data, the database to store standardized data and detection of potential hearing disorders, and the database used by the display control unit may be the same (e.g., database 20 in Figure 1), or they may be different (e.g., database 20 in Figure 1). In either case, specific tables are prepared separately, so the data from both can be managed separately. In addition, in this invention, the database also includes the storage location when data is temporarily or permanently stored in memory, etc. For example, database 20 may be located in the storage device or client-side storage on a smartphone, or it may be located in the server-side storage.

[0071] The languages ​​used for input, storage, reading, standardization, detection, identification, and display may depend on the framework, but examples include Python, Java, JavaScript, TypeScript, and PHP. The processing units (e.g., storage unit, reading unit, standardization unit, detection unit, identification unit, display control unit) may be centrally located on a cloud platform such as a Virtual Private Cloud (VPC) (AWS), GCP, Google Cloud® (Google®), Azure® (Microsoft®), etc., or they may be distributed. The storage process may also be a transmission process, for example.

[0072] This method may be implemented as software running on real or virtual hardware resources.

[0073] For example, this method is executed through a storage step of storing an audiogram image or its constituent data in a database; a standardization step of standardizing the data obtained from the audiogram image or the constituent data; a detection step of inputting the standardized data into an AI model to detect indicators of acoustic neuroma; and an output display control step of outputting or displaying the detection results. Optionally, the computer may have a sensorineural hearing loss identification step that uses the AI ​​model to distinguish between sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by causes other than acoustic neuroma from the audiogram. Each step element will be described in detail below with reference to Figure 2.

[0074] (I) A storage step in which the computer stores the audiogram image or its constituent data in a database (S100: Input and storage of audiogram) The "audiogram image or its constituent data" may be input by a human or directly from an audiometer or hearing measurement device. The input "audiogram image or its constituent data" is stored in a database by a computer (S100).

[0075] Audiogram images are not particularly limited, but examples include images used by doctors during examinations and diagnoses, or images of audiograms provided by medical institutions that patients have photographed with their smartphone cameras. Doctors may use images displayed on the screen from electronic medical records directly, or image data obtained secondarily from other devices or smartphones may also be used.

[0076] While there are no particular limitations on the configuration data, you can use raw data obtained from various audiogeometry methods directly, or you can use processed data that has been further processed and modified.

[0077] The number of audiogram images or their constituent data is not particularly limited; for example, 1, 2, 3, 4, 5, 10, 100, 1000, or more than 10,000 may be entered. The number may also include any two numbers between the above ranges.

[0078] Input may be submitted via a website or through an input device in a device or system. The means of input are not particularly limited, and any method can be used.

[0079] Websites are preferably deployed on a server as web applications. Web applications consist of arbitrary frontends and backends. Input / storage processes are preferably implemented as frontend / client-side (Figure 12A). Input / storage processes may be within an app on a smartphone or used for input / storage on a website. In the case of native apps, processing is performed using, for example, Swift, Java, Kotlin, or Flutter. When displayed on the web, the frontend consists of, but is not limited to, HTML, CSS, JavaScript, TypeScript, etc. Frontend frameworks may also be used, and examples include React, Vue.js, Angular, Ember.js, Backbone.js, Next.js, Nuxt.js, Svelte, Elm, Preact, Inferno, Flutter, etc. Backend frameworks (e.g., Django, Laravel, Ruby on Rails, Flask, Node.js) are used to process in conjunction with the server side.

[0080] The databases listed above can be used, as can databases within the system or device and / or server-side databases. Preferably, a database located on the cloud is used.

[0081] (II) Process of reading the audiogram frame and data points (S200: reading process) If the original audiogram is input in S100, the process of reading its frame and each hearing data point is executed. The reading of the frame and each symbol in the figure is not particularly limited, and the reading process is performed using any reading means. For example, it is preferable to use an AI model, more preferable to use a machine learning model, and even more preferable to use a deep learning model.

[0082] Examples of elements to be read include, but are not limited to, the outer frame, each vertical and horizontal line, right air conduction data point (e.g., red circle), right bone conduction data point (e.g., red [), left air conduction data point (e.g., blue ×), left bone conduction data point (e.g., blue []), red line, and blue dashed line. Each element to be read may be recognized by the AI ​​as an individual class, or they may be arbitrarily combined and recognized as a single class. Preferably, the AI ​​model is trained to recognize overlapping right and left air conduction data points as a single class (aa6). Table 1 shows that when a deep learning model was trained to recognize overlapping left and right air conduction data points as a single class (aa6), the recognition accuracy (mAP50) was very high at 0.995.

[0083] Table 1 of Example 1 shows that the recognition accuracy of the frame, each air conduction data point on the left and right, and each bone conduction data point was high, and that this information was extracted with high precision, demonstrating a remarkable effect.

[0084] (III) Standardization step (S300: Audiogram standardization) to standardize the data obtained from the audiogram image or the constituent data. In this step, the data obtained from the audiogram image or the constituent data is standardized. The method of standardization is not limited as long as it is achieved by processing according to certain rules. When standardizing as image data, for example, the outer frame of the audiogram may be fixed to a certain size (number of pixels). The certain size (predetermined value) is arbitrary and is not limited to, for example, matching the input size of the audiographic disorder index detection image AI model. Specifically, examples include squares of 224, 240, 260, 300, 380, 384, 416, 456, 528, 600 pixels, but are not particularly limited. The number of pixels in the vertical and horizontal directions may be different.

[0085] When standardizing numerical data, upper and lower limits may be set as constraints. For numerical processing, libraries such as NumPy or Pandas, which are Python libraries, may be used. Preferably, numerical processing is performed using a Pandas DataFrame.

[0086] Each air conduction data point and each bone conduction data point on the left and right sides are standardized by processing them according to certain rules. For example, when standardizing to image data, each hearing data point can be shown relative to the frame. The symbols representing each data point may also be arbitrary, but preferably they are the corresponding symbols used in the original audiogram (e.g., right air conduction data point (e.g., red circle), right bone conduction data point (e.g., red [), left air conduction data point (e.g., blue ×), left bone conduction data point (e.g., blue [))) (Example 1: Details of Standardization (I)). Air conduction or bone conduction points may also be connected by line segments, and the shape and color of the line segments are not limited, but preferably, air conduction data points are connected by solid red lines and bone conduction data points are connected by dashed blue lines (Example 1: Details of Standardization (I)). Alternatively, the numerical values ​​(relative values) of each data point may be standardized (Example 1: Details of Standardization (II)).

[0087] Preferably, the standardization involves fixing the frame of the audiogram at a predetermined value and representing each hearing data point with a numerical value or graphic representation of its relative position to the frame.

[0088] When standardizing image data, audiograms can also be plotted according to certain features. For example, one can focus on the absolute difference between each air conduction data point on the left and right sides, and highlight the area of ​​that difference by filling it with black (Example 4).

[0089] Preferably, the standardized data is the absolute difference between the left and right air conduction data points. More preferably, each air conduction data point includes at least one data point at 2,000 Hz and 3,000 Hz, where the 2,000 Hz and 3,000 Hz data points may be approximately 2,000 Hz and approximately 3,000 Hz data points.

[0090] When only the image size was standardized without data standardization (Comparative Example 1), phenomena that appeared to be caused by image manipulation bias occurred, and the AI ​​model failed to correctly detect the possibility of acoustic neuroma based on image features (conditions for extracting audiograms from electronic medical records). This demonstrated that standardization was significantly effective in cases where it was implemented (e.g., Examples 2, 3, 4, and 7). However, even with standardization, there were cases where the accuracy of detecting the possibility of auditory disorders was not optimal (e.g., Examples 5 and 6), indicating that selecting an appropriate AI model could yield better results.

[0091] When standardizing numerical data, each hearing data point can be shown relatively within an upper and lower limit. Alternatively, the numerical data may be standardized overall to values ​​between 0 and 1 before being input into the AI ​​model. While not limited to this, it is also preferable to process the numerical data as a Pandas DataFrame.

[0092] After the standardization described above, the data is converted into numerical data, which can then be fed into a machine learning model. For example, a machine learning model that focuses on the "absolute value of the difference between each air conduction data point on the left and right sides" and uses the absolute value of the aforementioned difference at each frequency as a feature was found to have high accuracy in detecting the possibility of hearing disorders and to be highly effective (Example 7). However, even after standardization, the accuracy of detecting the possibility of hearing disorders was not optimal in some cases (Examples 5 and 6), and it was found that selecting the optimal AI model would yield better results.

[0093] (IV) Detection step (S400: AI model detection) in which the standardized data is input into the AI ​​model to detect indicators of acoustic neuroma. In this process, the standardized data is input into an AI model to detect indicators of hearing disorders. The standardized data may be image data or numerical data.

[0094] Specifically, the recall rates for detecting acoustic neuroma indicators were 0.90 for Example 2: Acoustic neuroma indicator detection (image-based) AI model (1) EfficientNet; 0.80 for Example 3: Acoustic neuroma indicator detection (numerical-based) AI model (2) MLP; 0.70 for Example 4: Acoustic neuroma indicator detection (image-based) AI model (3) Left / Right absolute values ​​EfficientNet; 0.40 for Example 5: Acoustic neuroma indicator detection image AI model left / right separate air conduction / bone conduction EfficientNet; 0.24 for Example 6: Acoustic neuroma indicator detection (numerical-based) AI model left / right separate air conduction XGBoost; and 0.80 for Example 7: Acoustic neuroma indicator detection (numerical-based) AI model (4) Left / Right air conduction difference absolute values ​​XGBoost. Therefore, the recall rate for detecting acoustic neuroma indicators is preferably 0.70 or higher, more preferably 0.80 or higher, and even more preferably 0.90 or higher.

[0095] When the standardized data is the absolute value of the difference between the left and right air conduction data points, it has been demonstrated that both the image AI model (Example 4) and the numerical AI model (Example 7) can construct AI models that are excellent at detecting auditory disease indicators, demonstrating a remarkable effect.

[0096] Table 2 compares the performance of this model with existing technology (Comparative Example 2) (Non-Patent Literature 4). Comparative Example 2 is described in detail. The recall rate of Rule-Based 2 is 0.94, which is higher than that of this model, but its specificity is 0.22, which is the lowest and is problematic. Low specificity can lead to a higher false positive rate, resulting in too many false positives during detection, which can be problematic. Therefore, by limiting this model to those with "detection recall of 0.70 or higher and specificity greater than 0.22", "detection recall of 0.80 or higher and specificity greater than 0.22", and "detection recall of 0.90 or higher and specificity greater than 0.22", it can be shown to have superior performance compared to existing technology.

[0097] While there are no particular limitations on how the AI ​​model is trained, it is preferable that the AI ​​model is trained using the standardized data mentioned above. Training the AI ​​model with the standardized data increases the efficiency of the training and ensures accuracy.

[0098] Examples of AI models include, but are not limited to, machine learning models and deep learning models. Examples of supervised machine learning include linear regression, logistic regression, support vector machines (SVM), decision trees, neural networks (NN), and naive Bayes. Decision trees can also utilize boosting models such as random forests and XGBoost. AI models include, for example, AI models trained or pre-trained on audiogram images or their constituent data from patients with or without tumors.

[0099] Examples of neural networks include MLP (Multilayer Perceptron), CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), and GAN (Generative Adversarial Network). These deep learning models are not particularly limited; any model that can be trained can be used and trained.

[0100] Training is preferably performed in a separate cloud or on-premises environment. Examples of cloud environments include Google Colaboratory, EC2 instances with GPUs on AWS, or SageMaker. GPUs, TPUs, and CPUs can be used for training, but are not limited to these. Scikit-learn can be used for machine learning. In addition, deep learning frameworks such as TensorFlow, Keras, PyTorch, MXNet, and Caffe can be used.

[0101] The training unit is preferably located on a separate cloud VPC or device. For example, training can be performed using Google Colaboratory, an EC2 instance with a GPU on AWS, or SageMaker.

[0102] Furthermore, the environment in which the methods, apparatus, or systems described in (1) to (23) are executed / deployed and the environment in which the training is conducted may be the same or different. Preferably, they are executed / deployed in different environments.

[0103] Training can be done from scratch or using transfer learning. Transfer learning has the effect of improving learning efficiency by training the model again with data in a time series.

[0104] Methods for detecting indicators of acoustic neuroma from audiograms may be included in the following methods for detecting indicators of hearing disorders from audiograms.

[0105] (1) A method for detecting indicators of hearing disorder from an audiogram, A storage step in which a computer stores an audiogram image or its constituent data in a database, A standardization step of standardizing the data obtained from the audiogram image or the constituent data, A method comprising a detection step of inputting the standardized data into an AI model to detect indicators of hearing disorders; (2)' The method according to (1)', wherein the standardization is to fix the frame of the audiogram at a predetermined value and to indicate each hearing data point with a numerical value or figure of its relative position to the frame.

[0106] The hearing disorder is preferably at least one hearing disorder selected from the group consisting of acoustic trauma, perilymphatic fistula, malformation, functional hearing loss, tympanic membrane injury, Eustachian tube dysfunction, otosclerosis, ossicular transection, tumor (e.g., acoustic neuroma, brain tumor), skull fracture, noise-induced hearing loss, otitis media, toxic inner ear injury, sudden hearing loss, cochlea, Meniere's disease, mumps, and age-related hearing loss. More preferably, the hearing disorder is at least one hearing disorder selected from the group consisting of acoustic neuroma, Meniere's disease, otosclerosis, and noise-induced hearing loss. Even more preferably, the hearing disorder is an acoustic neuroma.

[0107] (3)' The method according to (1), wherein the auditory disorder is at least one auditory disorder selected from the group consisting of acoustic trauma, perilymphatic fistula, malformation, functional hearing loss, tympanic membrane injury, Eustachian tube dysfunction, otosclerosis, ossicular transection, tumor (e.g., acoustic neuroma, brain tumor), skull fracture, noise-induced hearing loss, otitis media, toxic cochlear dysfunction, sudden hearing loss, cochlear inflammation, Meniere's disease, mumps, and age-related hearing loss; (4)' The method according to (3), wherein the auditory disorder is at least one auditory disorder selected from the group consisting of acoustic neuroma, Meniere's disease, otosclerosis, and noise-induced hearing loss; (5)' The method according to (4)', wherein the auditory disorder is an acoustic neuroma.

[0108] The features of the standardization process, detection process, and display control process for detecting indicators of acoustic neuroma from an audiogram can also be applied to methods, devices, and systems for detecting indicators of hearing disorders from an audiogram.

[0109] (V) Acoustic neuroma identification process (S500: Tumor identification) This identification step is optional, and method (1) may further include an acoustic neuroma identification step (S500: tumor identification), or it may consist only of the acoustic neuroma identification step (S500: tumor identification). The identification step may function in conjunction with the detection step (IV), and the AI ​​model used in the detection step and the AI ​​model used in the identification step may be the same or different. The total number of AI models may be one, or each detection step or each identification step may consist of one or more AI models.

[0110] The information described in the AI ​​model detection section of S400 will be applied as appropriate to the AI ​​model used for tumor identification in S500.

[0111] Preferably, the process further includes a sensorineural hearing loss identification step in which a computer uses an AI model to distinguish from the audiogram between sensorineural hearing loss caused by an acoustic neuroma and sensorineural hearing loss caused by a cause other than an acoustic neuroma. More preferably, in the sensorineural hearing loss identification step, the standardized data includes at least one air conduction data point at 125 Hz and 250 Hz, where the 125 Hz and 250 Hz air conduction data point may be approximately 125 Hz and approximately 250 Hz air conduction data point. Even more preferably, the standardized data includes air conduction data points and bone conduction data points.

[0112] Specifically, the recall rate for distinguishing between sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by other factors from the audiogram was 0.89 for Example 10: Acoustic neuroma sensorineural hearing loss discrimination (numerical-based) AI(MLP) model (D1); 0.78 for Example 11: Acoustic neuroma sensorineural hearing loss discrimination (numerical-based) AI(XGBoost) model (D2); and 0 for Example 12: Acoustic neuroma sensorineural hearing loss discrimination (numerical-based) AI(MLP) model (no 125, 250 Hz air conduction data points). .67;Example 13: The recall, precision, and specificity of the acoustic neuroma sensorineural hearing loss discrimination (image-based) AI (EfficientNEt) model (absolute value of left-right air conduction difference) were 0.78, 0.54, and 0.25, respectively;Example 14: The recall of the acoustic neuroma sensorineural hearing loss discrimination (numerical-based) AI (MLP) model (D3) (all left-right air conduction and bone conduction) was 1.0;Example 15: The recall of the acoustic neuroma sensorineural hearing loss discrimination (numerical-based) AI (XGBoost) model (D4) (all left-right air conduction and bone conduction) was 0.78. Therefore, the recall of the discrimination is preferably 0.78 or higher, more preferably 0.89 or higher, and most preferably 1.0.

[0113] Here, we compared the performance of this model with the GIMSING protocol (Non-Patent Literature 3), which is sometimes used in clinical practice, using the same dataset used to construct and evaluate this model (Comparative Example 3). Details are described in Comparative Example 3. This model was superior to the GIMSING protocol in terms of recall, specificity, accuracy, precision, and NPV (Table 4). However, the recall of the air conduction XGBoost model in Example 11 was 0.78, which is slightly lower than the GIMSING protocol (0.80), but the other indicators were superior.

[0114] Therefore, by limiting the criteria to "recall of recognition greater than 0.80" and "recall of recognition of 0.89 or higher," the AI ​​model in question performs better than the GIMSING protocol in all of the evaluated metrics (Table 4).

[0115] The frame and data point reading process (S200), audiogram standardization process (S300), AI model detection process (S400), and acoustic neuroma identification process (S500) may constitute an API (Application Programming Interface), although this is not limited to them. The API is preferably implemented as a backend / server-side. The API may be a REST API, and examples of frameworks, though not limited to them, include Django REST framework (Python), FastAPI (Python), Micronaut (Java), Light (Java), and API Platform (PHP). Furthermore, since processing by the AI ​​model takes time after receiving a request for input images or data, an asynchronous processing mechanism may be provided to return a response within an appropriate time. The asynchronous processing mechanism is not limited to these, but in the case of Python, it can utilize the functions of Python itself, or it can be implemented by implementing Celery and Redis in a specialized container. The API may be implemented using Docker container technology, and orchestration architectures such as Docker-Compose or Kubernetes can be used for the implementation of that container.

[0116] (VI) Output display control step (S600: Detection result output) This step is also optional and involves outputting or controlling the display of the detection results. The output / display control means is not particularly limited and can be output to or displayed on a display area on the device or system, or it may be output to or displayed on a display device on a smartphone, tablet, or PC as a website.

[0117] Websites are preferably deployed on a server as web applications. Web applications consist of arbitrary frontends and backends. The detection result output / display control process is preferably implemented as a frontend / client-side process (Figures 12B and C). The detection result output / display control process may be within an app on a smartphone, or it may be used for outputting / displaying detection results on a website. In the case of native apps, processing is performed using, for example, Swift, Java, Kotlin, or Flutter. When displayed on the web, the frontend consists of, but is not limited to, HTML, CSS, JavaScript, TypeScript, etc. Frontend frameworks may also be used, and examples include React, Vue.js, Angular, Ember.js, Backbone.js, Next.js, Nuxt.js, Svelte, Elm, Preact, Inferno, Flutter, etc. Backend frameworks (e.g., Django, Laravel, Ruby on Rails, Flask, Node.js) are used to process in conjunction with the server side.

[0118] The objects to be displayed are not particularly limited as long as they can output / display detection results. Any output / display means can be used, such as text display, graphic display (Figures 12B and C), and audio display.

[0119] The order of processing steps in the methods described herein is not particularly limited. For example, steps S100 (audiogram input storage), S200 (frame and data point reading), and S300 (audiogram standardization), S400 (AI model detection), S500 (acoustic neuroma identification), and S600 (detection result output display) may be repeated, or only S100 and / or S200 may be repeated, or S300, S400, S500, and S600 may be performed intermediately or finally, but are not particularly limited.

[0120] Embodiments 2 and 3: Apparatus or system for detecting indicators of acoustic neuroma from an audiogram Embodiment 2 is, A computer includes a storage unit that stores audiogram images or their constituent data in a database, A standardization unit that standardizes the data obtained from the audiogram image or the configuration data, The device includes a detection unit that inputs the standardized data into an AI model to detect indicators of auditory disorders, and an audiogram that detects indicators of acoustic neuroma.

[0121] Embodiment 3 is, A computer includes a storage unit that stores audiogram images or their constituent data in a database, A standardization unit that standardizes the data obtained from the audiogram image or the configuration data, This system includes a detection unit that inputs the standardized data into an AI model to detect indicators of auditory disorders, and a system that detects indicators of acoustic neuroma from an audiogram.

[0122] A device or system for detecting indicators of acoustic neuroma from second and third-view audiograms is provided, which performs the methods (1) to (13) of the first view. This device or system may be implemented as an application, preferably a web application or a native application (on a smartphone or tablet). More preferably, it is implemented as a web application. The device or system can also be used independently or can be used by being integrated into or communicating with an audiometric device (hearing test device).

[0123] A device for detecting indicators of acoustic neuroma from an audiogram (Figure 1) may comprise a storage unit 10, a database 20, a standardization unit 30, a detection unit 40, an identification unit 50, and an (output) display control unit 60, but it may also have other functional units. Furthermore, the detection unit 40 and the identification unit 50 may be provided as the same unit or as separate units. This device can efficiently detect indicators of auditory diseases from an audiogram. Specifically, the storage unit 10 stores the audiogram image or its constituent data in the database 20 via a computer. The standardization unit 30 optionally reads the audiogram and then performs the standardization described above. The detection unit 40 then inputs the standardized data into an AI model to detect indicators of acoustic neuroma. In addition, it may optionally have an identification unit 50 that distinguishes acoustic neuroma from other diseases. Various standardized data and their detection results may be stored in the database 20 as appropriate. The display control unit 60 then outputs / displays the original data, standardized data, detection results, etc., stored in the database 20.

[0124] The methods (1) to (13) and the devices (14) to (18) or the systems (19) to (23) were demonstrated in Example 8 to be actually deployable as a web application.

[0125] In the apparatus or system for detecting indicators of hearing disorders from an audiogram of Embodiment 2 or 3, each component (e.g., storage unit, storage unit, database, standardization unit, detection unit, identification unit, display control unit) may or may not reside on the same cloud or device. The elements and preferred embodiments specified in the method of Embodiment 1 are applied equally and produce significant effects. Each component is described in detail below.

[0126] (A) Storage unit for storing the audiogram image or its constituent data in a database (S100: Audiogram input storage) In the storage unit 10, the "audiogram image or its constituent data" is the same as or replaceable with the one used in Embodiment 1. For example, if the object used in Embodiment 1 is an "audiogram image," it is an "audiogram image" of the user themselves or another person. If it is "constituent data," it may be raw hearing measurement data recorded in electronic medical records, etc., or data measured by various audiometric devices (hearing measurement devices) that is directly or indirectly (after some processing) input / stored in the database 20.

[0127] Input can include, for example, image data taken with a smartphone or camera. The input section may be within a smartphone app or used to upload data to a website. In the case of native apps, processing is performed using, for example, Swift, Java, Kotlin, or Flutter. When inputting data on the web, the frontend is not limited to, but is composed of HTML, CSS, JavaScript, TypeScript, etc. Frontend frameworks may also be used, and examples include React, Vue.js, Angular, Ember.js, and Backbone.js. Backend frameworks (e.g., Django, Laravel, Ruby on Rails, Flask, Node.js) are used to process data in conjunction with the server side or to store it in a database.

[0128] (B) Standardization unit that standardizes data or constituent data obtained from the audiogram image (S200: frame and data point reading, S300: audiogram standardization) In the case of an audiogram image, the process described in (II) Reading the audiogram frame and data points (S200: Reading process) is performed. Specifically, an object detection algorithm (identifying the type and location of an object) can be used. Examples of such algorithms include Fast R-CNN, Faster R-CNN, YOLO (You Only Look Once), SSD (Single Shot Detector), and DETR (DEtection Transformer). In Example 1, it was demonstrated that the type and location of each audiogram frame and hearing data point can be identified using YOLO. This reading process may be performed as part of the standardization unit 30.

[0129] The standardization unit 30 standardizes the audiogram information and audiogram configuration data that have been read. The standardization process is carried out by executing the process described in (III) Standardization step (S300: Audiogram standardization) for standardizing the data obtained from the audiogram image or the configuration data.

[0130] (C) Detection unit (S400: AI model detection) that inputs standardized data into an AI model to detect indicators of hearing disorders. The detection unit 40 executes the process described in (IV) above, the detection step (S400: AI model detection) in which the standardized data is input to the AI ​​model to detect indicators of hearing disorders. Preferably, the trained AI hearing disorder indicator detection model described in Embodiment 4 is used. Training is preferably performed using the standardized audiogram data.

[0131] The detection unit may use the trained model described above and be provided from within the system or from an external source as an API (Application Programming Interface). The API is preferably a RESTful API.

[0132] (D) Identification unit for identifying tumors from sensorineural hearing loss (S500: Tumor identification)

[0133] The identification unit 50 inputs the standardized data into the AI ​​model and executes the process described in the identification step (S500: tumor identification) for identifying tumors from sensorineural hearing loss. Preferably, the trained AI acoustic neuroma sensorineural hearing loss identification model described in Embodiment 5 is used. Training is preferably performed using the standardized audiogram data.

[0134] The identification unit may use the trained model described above and be provided from within the system or other internal components, or it may be provided externally as an API (Application Programming Interface). The API is preferably a RESTful API.

[0135] (E) Output / display control unit (S500: Detection result output) that outputs or displays the detection result. The display control unit 60 that displays the aforementioned indicators (tumor indicators and / or tumor identification results) to the user may be a display control unit of various augmentation devices, an application on a smartphone, or one used to display data on a website. In the case of a native application, processing is performed using, for example, Swift, Java, Kotlin, or Flutter. When displaying on the web, the frontend is composed of, but is not limited to, HTML, CSS, JavaScript, TypeScript, etc. Frontend frameworks may also be used, and examples include React, Vue.js, Angular, Ember.js, and Backbone.js. A backend framework (e.g., Django, Laravel, Ruby on Rails, Flask, Node.js) is used to process in cooperation with the server side.

[0136] Embodiment 4: Trained AI Acoustic Neuroma Index Detection Model Embodiment 4 is, (25) An input layer which takes an audiogram image or its constituent data as input, It consists of a neural network with an intermediate layer that processes input data, It includes an output layer that outputs the possibility of an acoustic neuroma, A trained AI acoustic neuroma index detection model used in the method, apparatus, or system described in (1), (14), or (19).

[0137] Preferably, the trained AI acoustic neuroma index detection model described in (25) is used to cause a computer to function to output auditory disorder index detections corresponding to an input audiogram image or its constituent data.

[0138] The preferred embodiments and limitations described in the methods, apparatus, and systems of Embodiments 1 to 3 are also applicable to this embodiment.

[0139] The trained AI acoustic neuroma index detection model is preferably a machine learning model, more preferably a neural network model (e.g., a multilayer perceptron model), and even more preferably a deep learning model. Any of the above-mentioned machine learning models, neural network models, and / or deep learning models can be used.

[0140] An AI model includes an input layer, an intermediate layer, and an output layer, but may also include one or more other layers.

[0141] The input layer may accept image data, numerical data, text data, or a combination of these. Multiple features can also be input as data. Categorical variables may be encoded using one-hot encoding.

[0142] The hidden layer (intermediate layer) is a layer that processes data input from the input layer. There is no limit to the number of hidden layers; any number of hidden layers can be used. Layers such as convolutional neural networks (CNNs), pooling layers, fully connected layers, and dropout layers can be used in any number and combination.

[0143] The output layer can either regress values ​​or classify categories. It can also have multiple outputs.

[0144] Preferably, an AI model is trained in which the standardized data is the absolute difference between each left and right air conduction data point. More preferably, the air conduction data point includes at least one data point between 2,000 Hz and 3,000 Hz.

[0145] These trained AI auditory disorder index detection models may function as part of a web application or as part of a device or system. Example 8 shows an example of a web application equipped with an AI model that detects acoustic neuroma indexes, demonstrating that the resulting AI model can actually function as a web application.

[0146] Examples 1-7 demonstrated that it is possible to construct actual AI-based acoustic neuroma index detection models. Furthermore, Example 8 showed that the resulting models can be implemented as web applications.

[0147] Embodiment 5: Trained AI Acoustic Neuroma Sensorineural Hearing Loss Identification Model Embodiment 5 is, (26) An input layer which takes an audiogram image or its constituent data as input, It consists of a neural network with an intermediate layer that processes input data, It includes an output layer that outputs the results of distinguishing between sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by causes other than acoustic neuroma, A trained AI acoustic neuroma sensorineural hearing loss identification model used in the method, apparatus, or system described in (1), (14), or (19); (27) The trained AI acoustic neuroma sensorineural hearing loss discrimination model described in (26) for causing a computer to output results that identify sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by causes other than acoustic neuroma, in response to an input audiogram image or its constituent data; That is the case. The information described in the trained AI acoustic neuroma index detection model can be appropriately applied to the trained AI acoustic neuroma sensorineural hearing loss discrimination model.

[0148] Preferably, standardized data is used to train an AI model that includes low-frequency data points (e.g., 125 Hz, 250 Hz) for each air conduction on the left and right sides. More preferably, the AI ​​model is trained with standardized data that includes both air conduction and bone conduction data points. It has been shown that including these features enables a better AI model for identifying acoustic neuroma sensorineural hearing loss (Examples 12 and 14).

[0149] Examples 10-15 demonstrated that it is possible to construct an actual AI-based sensorineural hearing loss identification model for acoustic neuromas.

[0150] Embodiment 6: Program for detecting indicators of acoustic neuroma from an audiogram The program according to the present invention is (28) A program for detecting indicators of acoustic neuroma from an audiogram, A storage step involves storing the audiogram image or its constituent data in a database on a computer. A standardization step of standardizing the data obtained from the audiogram image or the constituent data, This program performs a detection process in which the standardized data is input into an AI model to detect indicators of acoustic neuroma. The preferred embodiments and limitations described in the methods, apparatus, and systems of Embodiments 1 to 3 are also applicable to this program.

[0151] This program includes the whole program or parts thereof, to the extent that it enables the implementation of the methods, apparatus, systems, and models of the present invention. Examples of languages ​​include, but are not limited to, Python, Java, Kotlin, Flutter, Swift, C, C#, C++, PHP, Ruby, JavaScript, TypeScript, Scala, Go, R, Perl, Unity, COBOL, and others.

[0152] Embodiment 7: Recording medium including a program for detecting indicators of acoustic neuroma from an audiogram The recording medium according to the present invention is This recording medium includes a program for detecting indicators of acoustic neuroma from an audiogram as described in Embodiment 6.

[0153] Examples of recording media and computing devices that implement the present invention include, but are not limited to, RAM, ROM, cache, SSD, and hard disk. They also include any form of computing device, such as those on the cloud, on servers, or on on-premises computers. Recording media include, for example, non-temporary recording media that are readable by a computer.

[0154] If this method, apparatus, or system is provided as a native app on a smartphone or tablet, languages ​​such as Swift, Java, Kotlin, and Flutter may be used.

[0155] In this specification, the term "A to B" includes A and B. Furthermore, although the processes and other aspects of the present invention have been described in each embodiment, the invention is not limited to these descriptions, and various modifications can be made.

[0156] Furthermore, the order in which each part of the process or apparatus used in the method of the present invention is performed is not limited.

[0157] The present invention will be described in more detail below with reference to examples, but the present invention is not limited to the following examples. [Examples]

[0158] Example 1: Audiogram Standardization Audiogram The raw audiogram data was collected from patients who visited the Department of Otolaryngology at Tokyo Jikei University School of Medicine between 2011 and 2023. Audiograms were collected from 100 cases of pure-tone audiometry (tumor-free PTA) in patients without tumors, 89 cases of PTA in patients with acoustic neuromas, and 77 cases of sensorineural hearing loss caused by factors other than acoustic neuromas. These audiograms were used.

[0159] Characteristics of an audiogram Figure 3 shows the mean and standard deviation (SD) of hearing thresholds (dB) at various frequencies (Hz) for different patient groups (Figure 3). Error bars extending from each point represent the standard deviation. The graph in Figure 3A shows the mean and SD of PTA values ​​for 100 outpatients complaining of hearing loss. Hearing thresholds for the right ear (○) and left ear (□) are compared at different frequencies. It is observed that both ears show similar hearing thresholds. Figure 3B shows the mean and SD of PTA values ​​for 77 cases of SNHL without tumors. The thresholds for the right ear (○) and left ear (□) are compared. From this data, it can be seen that there is a consistent trend in the hearing thresholds of both ears. Figure 3C shows the mean and SD of PTA values ​​for 89 cases of acoustic neuroma. A comparison is made between the tumor side (○) and the non-tumor side (□). The tumor side shows a significantly higher hearing threshold compared to the non-tumor side, demonstrating the effect of tumors on hearing thresholds. Finally, Figure 3D shows a comparison of the data from graphs (A), (B), and (C) overlaid on top of each other. This shows the overall trend in hearing thresholds across all patient groups, highlighting the differences between outpatients, SNHL cases without tumors, and acoustic neuroma cases.

[0160] Audiogram reading PTAs from 100 tumor-free cases were directly imaged as JPG files from the Jikei Hospital electronic medical record system and grouped into 90 training / validation PTAs and 10 test PTAs. These JPG files were printed on paper, photographed with a smartphone, and duplicated as photo JPG files. The original JPG files and the photo JPG files were used to train and evaluate the YOLOv5x model (Ultralytics.YOLOv5.Git code, https: / / github.com / ultralytics / yolov5 (2020)) for mapping audiogram frames to auditory data points. Each audiogram frame (aa0) and each auditory data point (aa2, aa3, aa4, aa5, aa6) (see Table 1) were annotated using labelImg software (Tzutalin.LabelImg.Git code, https: / / github.com / tzutalin / labelImg (2015)). Training was performed using an NVIDIA A100-SXM4-40GB and 300 epochs in Google Colaboratory PRO+. Each class in each test data was detected with a confidence level of 0.30. This YOLOv5x model was saved and used for standardizing and redrawing various audiograms, and each auditory data point was provided as numerical data.

[0161] To standardize audiogram data and reduce noise and image bias, frames and auditory data points were first extracted from audiogram images and photographs. For this purpose, a YOLOv5x deep learning model was trained to map audiogram frames to auditory data points. Ninety tumor-free PTA and their photographic images were annotated according to the classes specified in Table 1 and used for training and validation. The YOLOv5x model was trained using 300 epochs of data. The mAP50 scores were 0.992 for aa0, 0.521 for aa1, 0.991 for aa2, 0.850 for aa3, 0.993 for aa4, 0.801 for aa6, and 0.995 for aa6. The frame (aa0) and each air conduction data point (aa2, aa4, aa6) were well recognized. Bone conduction data points (aa3 and aa5) were slightly less recognized, but comparable feasible results were obtained. The 0dB line (aa1) indicates a low mAP50, suggesting poor recognition.

[0162] [Table 1]

[0163] When the audiogram image of the test photograph derived from the original audiogram image (Figure 4A) (Figure 4B) was used to evaluate the YOLOv5x model with a confidence level of 0.30, the frame (aa0) and each auditory data point (aa2, aa3, aa4, aa5, aa6) were recognized and mapped perfectly, except for one extra aa2 (around 4000Hz) (Figure 4C). The 0dB line (aa1) was also detected. Similar detection accuracy was confirmed with other test images. Next, a new, noise-free, standardized audiogram was redrawn using information about the frame and each auditory data point (Figure 4D). This redrawn audiogram represented the original audiogram well.

[0164] Standardization Details (I) Information regarding the frame (aa0) and each hearing data point (aa2, aa3, aa4, aa5, aa6) was detected by the YOLOv5x model and used to redraw the original audiogram. First, a standardized frame (416 x 416 pixels) was provided using the frame information. Next, intermediate horizontal lines, including the 0dB line (bold), were drawn at equal intervals. The x, y positions of each hearing data point were converted to the x, y positions corresponding to the standardized frame (416 x 416 pixels) and plotted using the corresponding symbols (red "○" for right air conduction, red "[" for right bone conduction, blue "×" for left air conduction, blue "]" for left bone conduction, and overlapping red "○" and blue "×" for overlapping right and left air conductions). Then, the data points for the right air conduction were connected with red solid lines, and the data points for the left air conduction were connected with blue dashed lines. Finally, vertical lines were drawn at the relative x-position of each right air conduction data point (aa2). The redrawn audiogram was saved as a JPG file. In other words, fixing the frame to a constant value and displaying each hearing data point in its corresponding relative coordinate is one embodiment of standardization (in which the frame of the audiogram is fixed at a predetermined value, and each hearing data point is shown relative to the frame).

[0165] The effects of standardization (I) (1) First, as described above, a standardized audiogram created by fixing the frame values ​​and redrawing the corresponding hearing data points using their relative coordinates with various symbols (○, ×, [, ]) is free from image rendering noise and traces of image manipulation (e.g., enlargement, reduction), so the data points and the entire graph consist solely of hearing data (each frame, dB line, and frequency line are common). Therefore, it has the remarkable effect of providing an audiogram free from noise and bias from image manipulation. Furthermore, it is different from the audiograms that doctors and patients see, and in that the input for the audiogram used for judgment is different, it has the characteristic that the input for detection using AI is different from the input that a normal human would use for judgment.

[0166] (2) Next, the images used for training were images derived from electronic medical records (e.g., jpg) and images of those images printed on paper and photographed with a smartphone. Therefore, since the images used by doctors during examinations and diagnoses, and the images of audiograms provided by medical institutions to patients that are photographed with a smartphone camera are converted into standardized audiograms, this has the remarkable effect of being able to target a variety of audiograms.

[0167] (3) Raw data (constituent data) measured by various audiometric devices (hearing test devices) can also be plotted according to the above standardization, so there is a significant effect that not only printed audiograms but also raw data measured by the devices can be used as standardized audiograms.

[0168] Details of standardization (II) The above standardization details (I) are an example of standardization in an embodiment plotted as a figure, but it is also possible to standardize as the numerical value (relative value) of each data point. In this case as well, a remarkable effect similar to that of standardization effect (I) can be obtained. Furthermore, even if the original data is actual numerical values ​​from a measuring device, it can be plotted as a figure as described above, or standardized as a numerical value relative to the original data. In this case as well, a remarkable effect similar to that of standardization effect (I) can be obtained.

[0169] Example 2: Acoustic neuroma indicator detection (image-based) AI model (TD1) PTAs without tumors (100 cases) and PTAs with tumors (89 cases) were redrawn using the information obtained from the YOLOv5x model described above. The PTAs were divided into a PyTorch training / validation dataset (90 PTAs without tumors + 79 PTAs with tumors = 169 jpg images; training:validation=0.8:0.2) and a PyTorch test dataset (10 PTAs without tumors + 10 PTAs with tumors = 20 test jpg images). Images with tumors and images without tumors were labeled 1 (level 1) and 0 (level 0), respectively.

[0170] The EfficientNetV2_s model (Tan,ML,QVEfficientNetV2: Smaller Models and Faster Training. Proceedings of the 38th International Conference on Machine Learning.139(2021)) was pre-trained on Google Colaboratory PRO+ using PyTorch settings, importing tf_efficientnetv2_s from the timm (=0.9.2) library. For training, validation, and testing, image data was resized to (384×384 pixels) and normalized as in Example 1. Figures 5A and 5B show the results during training (A: training and validation accuracy, B: training and validation loss as CrossEntropyLoss). Figure 5A shows that the validation accuracy slightly increases as training progresses, and Figure 5B shows that the validation loss also gradually decreases slightly, suggesting that the model learned some auditory features, including tumors.

[0171] When the trained model was evaluated using a test dataset consisting of 10 PTA images without tumors and 10 PTA images with tumors, the resulting confusion matrix showed a recall (an indicator of the false negative rate) of 0.90, precision of 0.82, and specificity (an indicator of the false positive rate) of 0.80, demonstrating that it is a viable acoustic neuroma detection model (Figure 5C).

[0172] Example 3: Acoustic neuroma index detection (numerical basis) AI model (TD2) 100 PTAs without tumors were divided into 90 training / validation datasets and 10 test datasets. Additionally, 89 PTAs containing tumors were divided into 79 training / validation datasets and 10 test datasets. The 16 left and right air conduction data points correspond to rp0 (125Hz), rp1 (250Hz), rp2 (500Hz), rp3 (1,000Hz), rp4 (2,000Hz), rp5 (3,000Hz), rp6 (4,000Hz), rp7 (8,000Hz), lp0 (125Hz), and lp1 (250Hz), respectively. p2 (500Hz), lp3 (1,000Hz), lp4 (2,000Hz), lp5 (3,000Hz), lp6 (4,000Hz), lp7 (8,000Hz) (rp indicates the right air conduction data point, lp indicates the left air conduction data point), assigned to a standardized frame (416px × 416px), and tumor label (0: no tumor, 1: tumor). The data (including tumor data) was placed into a Python Pandas DataFrame (df). The resulting df was normalized using Scikit-learn's StandardScaler and used to train and validate a multilayer perceptron (MLP) model. This PyTorch multilayer perceptron (MLP) model includes an input layer (input dimensions: 16, 50), a hidden layer (50, 50), and an output layer (50, output dimension: 2). The ReLU function was applied to the outputs of the input layer and the hidden layer, respectively. The Softmax function was applied to the output of the output layer to detect the tumor level. The MLP model was trained for 100 epochs, with CrossEntropyLoss as the loss function, Adam as the optimizer, and a learning rate of 0.001.

[0173] Figure 6A shows the training loss and validation accuracy, indicating that the model has learned some tumor features. The AUC (Area Under Curve) of the ROC curve (Figure 6B), an indicator of tumor discrimination power, is 0.85, indicating high tumor detection accuracy. Furthermore, analysis of the confusion matrix using the same test dataset as in Example 2 (Figure 6C) showed a recall of 0.80, precision of 0.89, and specificity (an indicator of the false positive rate) of 0.90, demonstrating that it is a viable acoustic neuroma detection model.

[0174] Comparative Example 1: AI Comparative Example Model for Acoustic Neuroma Indicator Detection (Image-Based) Samples were prepared by simply scaling the original audiograms of the same dataset (training / validation dataset, test dataset) to 384 x 384 pixels (the input size for EfficientNetV2), without extracting and redrawing data points using YOLO. The same model was then trained using the same settings as in the example. Figures 7A and 7B show the training results. In this case, the accuracy became almost 1 during both training and validation, and the loss also decreased similarly, indicating that the model accurately learned some feature of the image. Figure 7C shows the confusion matrix obtained by the resulting model on the test dataset. In this case, the recall and specificity were 1, indicating that the model completely distinguished and detected the test dataset. Analysis of the cause revealed that the image density and the scaling ratio from the original image differed between PTA without tumors (e.g., Figure 7D) and PTA with tumors (e.g., Figure 7E) in the training / validation dataset and the test dataset. This is attributed to differences in image manipulation bias when adjusting the image data from the electronic medical record system. It is presumed that the model learned this difference, resulting in an AI image model that distinguishes and detects PTA without tumors (e.g., Figure 7D) from PTA with tumors (e.g., Figure 7E). In other words, it is suggested that the model is created that reads the characteristics of the provided image format rather than distinguishing the characteristics of auditory disorders (e.g., acoustic neuroma). Therefore, the need to remove image manipulation bias was revealed, and the remarkable effect of standardizing the original audiogram, as in Example 2, was demonstrated.

[0175] Example 4: Acoustic neuroma index detection (image-based) AI model (TD3) Next, we focused on the features that otolaryngologists pay attention to when interpreting audiograms. These features include the difference in hearing threshold levels (dB) between the left and right air conductions at specific wavelength frequencies (Hz). Therefore, we created audiogram plots using information obtained from YOLOv5x, filling in the areas between left and right air conduction data points in black (Figure 8). Figure 8A shows an example of a plot for PTA without tumors, and Figure 8B shows an example of a plot for PTA with tumors. In the plotted PTA without tumors, there was variability in the size and shape of the blacked-out areas. Similarly, when plotting PTA with tumors, there was variability in the size and shape of the blacked-out areas. However, the variation in the plotted PTA without tumors appeared to be more widespread.

[0176] Using the same training / validation and test datasets as in Example 1, the EfficientNetV2_s model was trained using standardized, redrawn audiograms in essentially the same way as the EfficientNetV2_s model described above. Figure 8C shows the detection results using the test dataset. The recall was 0.70, which is lower than that of the EfficientNetV2_s model in Example 2, indicating a slightly higher false negative rate. The precision was 0.88 and the specificity was 0.90, indicating a slightly lower false positive rate. This shows that when an image AI model (EfficientNetV2_s) is trained using data obtained by extracting each hearing data point from the original audiogram using YOLOv5, standardizing it by focusing on a certain viewpoint (the difference in hearing threshold levels (dB) between the left and right air conductions at a specific wavelength frequency (Hz)), and plotting (redrawing) the data as input, the trained model becomes a model that can detect indicators of the presence of auditory tumors, demonstrating excellent effectiveness.

[0177] Example 5: AI model for detecting acoustic neuroma indicators Using the same training / validation and test datasets as in Example 1, this time the air conduction and bone conduction data points for the left and right sides were standardized using YOLOv5 as in Example 1, and each air conduction data point (red circle) and each bone conduction data point (blue rectangle) were displayed, with each data point connected by a line segment. Figure 9A shows the plot without tumors, and Figure 9B shows the plot with tumors. Using the standardized and re-plotted dataset in this way, the EfficientNetV2_s model was trained in essentially the same way as the EfficientNetV2_s model described above. Figure 9C shows the detection results using the test dataset. It was found that a model with a low recall of 0.4 and a high false negative rate was created. Comparing Example 5 and Example 4, it was found that it is important to focus on what features to standardize and plot the audiogram. This demonstrated that the "difference in hearing threshold levels (dB) of left and right air conduction at a specific wavelength frequency (Hz)" in Example 4 had a particularly significant effect.

[0178] Example 6: AI model for detecting acoustic neuroma indicators (numerical basis) Numerical-based XGBoost model for detecting acoustic neuromas from simple audiogram air conduction data First, we investigated whether acoustic neuromas could be detected using a numerically-based XGBoost model based on simple left and right air conduction data points. The same training / validation dataset as in Example 1 was used to train this XGBoost model. This time, the amount of data doubled because we created separate data for the left and right sides from the information on each hearing data point obtained with YOKOv5x. Tumor levels (0 or 1) were assigned to the corresponding data points. The data points used correspond to the left and right air conduction points p0 (125Hz), p1 (250Hz), p2 (500Hz), p3 (1,000Hz), p4 (2,000Hz), p5 (3,000Hz), p6 (4,000Hz), and p7 (8,000Hz).

[0179] The XGBoost model was trained with standard settings. Figure 10 shows the confusion matrix results using the validation dataset. The recall was 0.24, the precision was 0.29, and the specificity was 0.85. Here, the recall is low (i.e., the false negative rate is high).

[0180] Example 7: Acoustic neuroma index detection (numerical basis) AI model (TD4) XGBoost model using the absolute value of the difference between each specific data point of left and right air conduction Next, we again focused on the absolute value of the difference between each specific data point (d0 (125Hz), d1 (250Hz), d2 (500Hz), d3 (1,000Hz), d4 (2,000Hz), d5 (3,000Hz), d6 (4,000Hz), d7 (8,000Hz)). Using the same training / validation and test datasets as in Example 1, we used the numerically based machine learning model XGBoost for training and evaluation.

[0181] In Example 7, information on the frame and each hearing data point was extracted from the original audiogram using YOLOv5, and p0 to p7 and d0 to d7 were calculated based on this information by standardization. This standardization based on numerical values ​​is also included in the description of "an audiogram in which the frame is fixed at a predetermined value, and each hearing data point is shown relative to the frame" in this specification.

[0182] Figure 11A shows the confusion matrix results using the validation dataset. Note that one true negative sample was obtained from the df (Pandas dataframe) negative control, so this df negative control needs to be subtracted from the number of true negative samples. The resulting XGBoost model had a recall of 0.77, precision of 0.71, and specificity of 0.65. This XGBoost was then optimized with learning_rate:0.2, n_estimators:100, objective:binary:logistic, max_depth:1, min_child_weight:4, max_delta_step:1, subsample:1, and threshold for prediction:0.4. Figure 11B shows the results of this optimized XGBoost using the test dataset. The recall was 0.80, precision was 0.67, and specificity was 0.64. The resulting model showed a slight improvement in recall, but decreased precision and specificity.

[0183] Next, we analyzed which data point differences were important for distinguishing between PTA without tumors and PTA with tumors. Figure 11C shows that d5 (3,000 Hz) and d4 (2,000 Hz) were particularly important, followed by d7 (8,000 Hz). This indicates that the difference in the values ​​of the hearing data points at 3,000 Hz and 2,000 Hz particularly contributes to the detection performance of this AI model. This result confirms that d5 (3,000 Hz) and d4 (2,000 Hz) are particularly important among the absolute values ​​of the difference between each specific data point of left and right air conduction, and that the "difference in hearing threshold levels (dB) of left and right air conduction at a specific wavelength frequency (Hz)" in Example 4 is important. It demonstrates that a model for detecting indicators of acoustic neuroma can be constructed even with a numerically based machine learning AI model, and it has shown remarkable effectiveness. In addition, a previous report (Non-Patent Literature 3) has shown that asymmetry at two frequencies between 2 and 8 kHz is an indicator of acoustic neuroma symptoms. Therefore, it can be seen that the AI ​​model created captured known features.

[0184] Comparative Example 2: Performance Comparison between Existing Technology and This AI Model (I) Table 2 compares the various AI models created so far with current state-of-the-art rule-based protocols and machine learning models (Non-Patent Literature 4). Rule-based protocols are defined as (1) asymmetry of 20 dB or more, or (2) asymmetry of at least 10 dB at two adjacent frequencies, or 15 dB at one frequency. Protocols (1) or (2) had recall of 0.63 or 0.94, specificity of 0.81 or 0.22, and precision of 0.62 or 0.37. Current machine learning-based models, (1) artificial neural networks or (2) random forests, showed recall of 0.56 or 0.68, specificity of 0.86 or 0.83, and precision of 0.01 or 0.01, respectively. Although the ratio of tumor-positive to tumor-negative cases differs (affecting accuracy), our model (see rows 2-5 in Table 2) was comparable to or better than current rule-based protocols and machine learning-based models (see rows 6-9 in Table 2). [Table 2]

[0185] Example 8: Web application equipped with an AI model for detecting acoustic neuroma indicators We built a web application equipped with three types of AI models: Example 2 (Model B: Image AI Model), Example 4 (Model A: Image AI Model), and Example 5 (Model C: Numerical AI Model). The web application used the Python framework Django 3.2. An API (Application Programming Interface) that takes audiogram images as input was built as a Docker container using the Django REST framework (3.13.1). Each service was implemented as a Docker container; specifically, Nginx was used as the proxy server, an API container including the UI (user interface) view was built, and Celery and Redis were implemented as auxiliary containers for asynchronous AI model inference. MySQL® 5.7 was used as the database container. These containers were orchestrated using Docker-Compose to implement the web application.

[0186] Figure 12A shows the UI for audiogram image input. A target audiogram for detecting acoustic neuroma indicators was input as a JPG image. Simultaneously, a check flag indicating whether an acoustic neuroma was diagnosed and comments regarding current symptoms were also entered. These elements may potentially improve acoustic neuroma detection in the future. Then, a password was entered, and the image and information were POSTed to the API. At this time, JWT (Jason Web Token) was used for authorization to enhance security.

[0187] When an audiogram without a tumor was posted, models A, B, and C all detected no possibility of a tumor and displayed this on the UI (Figure 12B). On the other hand, when an audiogram with a tumor was posted, models A, B, and C all detected a possibility of a tumor and displayed this on the UI (Figure 12C). Therefore, Example 8 demonstrated that the created acoustic neuroma detection AI models (Examples 2, 4, and 7) can actually function as web applications.

[0188] Example 9: Distinguishing between sensorineural hearing loss caused by tumors and non-tumor-induced causes. Sensorineural hearing loss can be caused by various factors, and even specialists currently find it difficult to distinguish between cases caused by tumors and those caused by other factors from audiograms alone. A definitive diagnosis requires the results of an MRI scan. In Example 9, after examining 89 cases of auditory neuroma-induced sensorineural hearing loss and 77 cases of auditory neuroma-induced sensorineural hearing loss using MRI scans, the presence or absence of tumors was detected from the audiograms using various acoustic neuroma index detection AI models. For cases of sensorineural hearing loss caused by other factors, 83.1% were detected as tumor-positive using the EfficientNetV2_s model (Model A: Example 4) which used a plot with the absolute air conduction difference between the left and right sides colored separately, 76.6% using the EfficientNetV2_s model (Model B: Example 2) which used a plot of a standardized redrawn audiogram, and 72.7% using the XGBoost model (Model C: Example 7) which used the absolute air conduction difference between the left and right sides respectively (Table 2). Therefore, the tumor detection model described above was unable to distinguish between sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by other factors. However, this result does not negate the usefulness of AI models for detecting indicators of acoustic neuroma from general cases. This is because acoustic neuroma is a rare disease in all cases, and its detection provides an incentive to perform further MRI examinations, thus having a significant and useful effect in itself.

[0189] [Table 3]

[0190] Example 10: AI model for identifying sensorineural hearing loss in acoustic neuromas (numerical basis) (D1) Next, we aimed to develop an AI model that can distinguish between sensorineural hearing loss caused by tumors and hearing loss caused by other factors. For this purpose, we prepared training / validation and test datasets using audiograms from 89 cases of sensorineural hearing loss including tumors and 77 cases of sensorineural hearing loss caused by factors other than tumors. The same multilayer perceptron (MLP) model as in Example 3 was used, but the dataset was different from that of Example 3. Figure 13A shows the training results. As the training loss decreased, the validation accuracy increased, suggesting that some kind of feature learning had occurred. The AUC of the ROC curve of the MLP model was 0.83 (Figure 13B). Looking at the confusion matrix of this MLP model using the test dataset, the recall was 0.89, the precision was 0.80, and the specificity was 0.78 (Figure 13C). As a result, it was shown that an AI model that can distinguish between sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by other factors was constructed, demonstrating a remarkable effect of providing an implementable model.

[0191] Example 11: AI model (D2) for identifying sensorineural hearing loss in acoustic neuromas (numerical basis) The same dataset as in Example 10 was used to train the XGBoost model. Figure 14A shows the confusion matrix results from the XGBoost model using the test dataset. The recall was 0.78, precision was 0.78, and specificity was 0.78, which is slightly worse than the MLP model. The importance of features in this XGBoost model was analyzed (Figure 14B). In Figure 13B, f0~f1 corresponds to rp0 (right air conduction, 125Hz)~rp1 (right air conduction, 250Hz), and f8~f9 corresponds to lp0 (left air conduction, 125Hz)~lp1 (left air conduction, 250Hz). This result shows that hearing data points at low frequencies, 125Hz and 250Hz, contribute significantly to the model's discriminative ability, suggesting the importance of including these data points.

[0192] Example 12: AI model for identifying sensorineural hearing loss in acoustic neuromas (numerical basis) To verify whether the low-frequency air conduction data points (125Hz and 250Hz) contribute to the MLP model described above, the training / validation and test datasets were prepared in the same manner as in Example 10, except that rp0, rp1, lp0, and lp1 were not included. The MLP model was the same except that the input dimension was changed to 12. As a result, the recall (0.67), precision (0.67), and specificity (0.67) of the MLP model were inferior to those of Example 10 (Figure 15A), demonstrating the importance of these low-frequency air conduction data points (125Hz and 250Hz). Therefore, it was demonstrated that including 125Hz and 250Hz air conduction hearing data points in the data set significantly improves the acoustic neuroma sensorineural hearing loss discrimination AI model.

[0193] Example 13: AI model for identifying sensorineural hearing loss due to acoustic neuroma (image-based) Furthermore, to create an audiogram plot with the region between the left and right air conduction points blacked out, a dataset similar to that used in Examples 10 and 11, with 16 air conduction points on each side, was used. The EfficientNetV2_s model was trained using this audiogram plot. Figure 15B shows the confusion matrix obtained by this model using the test dataset. The recall, precision, and specificity were 0.78, 0.54, and 0.25, respectively. The recall was similar to that of Example 11, but the specificity was extremely low. This result indicates that the EfficientNetV2_s model has poor ability to distinguish between tumor-induced hearing loss and hearing loss due to other factors in sensorineural hearing loss, and that the absolute difference in left and right air conduction contributes little to distinguishing between these sensorineural hearing losses. Celis-Aguilar et al. (Multiple audiometric analysis in the screening of vestibular schwannoma. Cureus 14, e21492 (2022)) found no significant difference in audiometry patterns and concluded that the presence or absence of vestibular schwannoma cannot be determined by audiometry alone in patients with asymmetric hearing loss. This finding that "the presence or absence of vestibular schwannoma cannot be determined by audiometry alone in patients with asymmetric hearing loss" is consistent with previously reported results. Therefore, it was shown that the acoustic neuroma index detection AI model (Examples 2-7), which detects acoustic neuromas from general non-tumor cases, and the acoustic neuroma sensorineural hearing loss discrimination AI model, which distinguishes sensorineural hearing loss caused by acoustic neuromas from sensorineural hearing loss caused by other causes, may be extracting different features and performing their respective detection and discrimination.

[0194] Example 14: AI model (D3) for identifying sensorineural hearing loss in acoustic neuromas (numerical basis) Next, we investigated whether bone conduction data points, as well as air conduction data points, contribute to the performance of the AI ​​model for identifying sensorineural hearing loss due to acoustic neuroma. For this purpose, training / validation and test datasets were prepared using audiograms from 89 cases of sensorineural hearing loss including tumors and 77 cases of sensorineural hearing loss due to causes other than tumors, with the exception that the dataset included both air conduction data points (right air conduction 125Hz~8000Hz: rp0-rp7; left air conduction 125Hz~8000Hz: lp0-lp7) and bone conduction data points (right bone conduction 250Hz-4000Hz: rbcp1-rbcp6; right bone conduction 250Hz-4000Hz: lbcp1-lbcp6) (p0-p7 are p0: 125Hz, p1: 250Hz, p2: 500Hz, p3: 1000Hz, p4: 2000Hz, p5: 3000Hz, p6: 4000Hz, p7: 8000Hz). A multilayer perceptron (MLP) model was trained using the same settings as in Example 10, except that it used a 28-dimensional input to include both air conduction and bone conduction data points, and the number of nodes in the hidden layer was changed from 50 to 100. Figure 16A shows the training results. As the training loss decreased, the validation accuracy increased, suggesting that some feature learning had occurred. The validation accuracy reached its maximum at approximately 40 epochs and then decreased. The model used was the one obtained when the validation accuracy reached its maximum value. The AUC of the ROC curve for the MLP model was 0.89, which is higher than in Example 10, indicating improved discrimination performance (Figure 16B). The confusion matrix of this MLP model using the test dataset showed a recall of 1.00, precision of 0.82, and specificity of 0.78 (Figure 16C). This test dataset achieved a recall rate of 1.00, demonstrating the successful construction of an AI model with excellent recall that can distinguish between sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by other factors. This result demonstrates a remarkable effect: using both air conduction and bone conduction data points allows for the construction of an AI model that further improves the discriminatory performance of acoustic neuroma sensorineural hearing loss.

[0195] Example 15: AI model (D4) for identifying sensorineural hearing loss in acoustic neuromas (numerical basis) The same dataset as in Example 14 was used to train the XGBoost model. Figure 17A shows the confusion matrix results for the XGBoost model using the test dataset. The recall was 0.78, precision was 0.78, and specificity was 0.78, which is slightly worse than the MLP model in Example 14. The importance of features in this XGBoost model was analyzed (Figure 17B). f0-f7 correspond to rp0-7, f8-13 correspond to rbcp1-6, f14-f21 correspond to lp0-7, and f22-27 correspond to lbcp1-6. The results in Figure 17B show that f18 (left air conduction, 2000Hz), f0 (right air conduction, 125Hz), f20 (left air conduction, 4000Hz), f16 (left air conduction, 500Hz), f15 (left air conduction, 250Hz), f24 (right bone conduction, 1000Hz), f1 (right air conduction, 250Hz), and f2 (right air conduction, 500Hz) have high contributions. This indicates that the contributions are high across a wide range of frequency bands for both left and right air conduction. As a result, it was found that when air conduction and bone conduction data points are included, data points spanning a wide range of frequency domains are important.

[0196] Comparative Example 3: Performance Comparison of Existing Technology and This AI Model (II) In clinical practice, otolaryngologists perform hearing tests and select SNHL cases. However, even otolaryngologists find it difficult to distinguish between SNHL cases with tumors and those without. This AI model can assist otolaryngologists in interpreting audiograms containing tumors. An MLP model using air conduction and bone conduction data points (recall, specificity, accuracy, precision, NPV (negative predictive value) = 1.0, 0.78, 0.89, 0.82, 1.00) showed a recall of 1.0 on this test dataset (Example 14, Figure 16). When performance was calculated using the latest GIMSING protocol (Non-Patent Literature 3) and the same dataset as Example 14, the recall, specificity, accuracy, precision, and NPV were 0.80, 0.14, 0.49, 0.52, and 0.38, respectively (Table 4). This air conduction and bone conduction MLP model performed remarkably well using the same dataset. Furthermore, the air conduction MLP model in Example 10 and the air conduction XGBoost model in Example 11 were superior to the existing GIMSING protocol in all evaluated performance aspects, except for the recall rate of Example 11, which was equivalent (0.78). Therefore, this AI model for identifying sensorineural hearing loss due to acoustic neuroma demonstrated a remarkable effect compared to existing technologies. This effect is not something that can be achieved through efficient human effort, but rather is realized by the characteristics of this AI model. [Table 4]

[0197] Figure 18 is a schematic diagram of an example hardware configuration of an information processing device 90 applied to this embodiment. The information processing device 90 comprises a processor 91, main memory 92, communication interface 93, auxiliary storage device 94, input / output interface 95, and internal bus 96. The processor 91, main memory 92, communication interface 93, auxiliary storage device 94, and input / output interface 95 are connected to each other via the internal bus 96 so as to be able to communicate with each other. The information processing device 90 may be applied to each of the above-described devices.

[0198] Furthermore, the operating entities of the processes described above, which are executed by a computer, are composed of a processor such as a CPU and memory. These operating entities function when the processor executes a program. Note that all or part of each function may be implemented using hardware such as ASICs, PLDs, or FPGAs. The above program may be recorded on a computer-readable recording medium. Computer-readable recording media include, for example, portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, and semiconductor storage devices (e.g., SSDs), as well as storage devices such as hard disks and semiconductor storage devices built into computer systems. The above program may also be transmitted via a telecommunications line. [Industrial applicability]

[0199] This invention provides a method, apparatus, system, or program, and various models for efficiently detecting indicators of acoustic neuroma from an audiogram, which can be used for various applications related to hearing disorders. It can also serve as a trigger for users (patients) to seek consultation or examinations with specialists at medical institutions, and can be used by physicians as an auxiliary tool for diagnosis.

Claims

1. A standardization process in which a computer standardizes audiogram-related information, which is information about the audiogram obtained about the subject person, An index detection method comprising: an index detection step of detecting an acoustic neuroma index in a target person by using standardized audiogram-related information as input to a trained model obtained by performing a learning process using data that is a combination of pre-acquired and standardized audiogram-related information and an index of acoustic neuroma in that person.

2. The index detection method according to claim 1, wherein in the standardization step, the frame of the audiogram is fixed at a predetermined value, and each hearing data point is standardized by indicating its position relative to the frame with a numerical value or graphic.

3. The indicator detection method according to claim 1, further comprising a sensorineural hearing loss identification step, in which the standardized audiogram-related information obtained for the target person is used as input to a trained model obtained by performing a learning process using data which is a combination of pre-acquired and standardized audiogram-related information and information indicating whether the person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma, thereby identifying whether the target person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma.

4. The index detection method according to claim 3, wherein the sensorineural hearing loss identification step includes at least one of the standardized audiogram-related information points at 125 Hz and 250 Hz.

5. The index detection method according to claim 4, wherein the recall rate of the identification in the sensorineural hearing loss identification step is 0.78 or higher.

6. The index detection method according to claim 4, wherein the recall rate of the identification in the sensorineural hearing loss identification step is greater than 0.

80.

7. The index detection method according to claim 4, wherein the recall rate of the identification in the sensorineural hearing loss identification step is 0.89 or higher.

8. The index detection method according to claim 4, wherein the standardized audiogram-related information includes air conduction data points and bone conduction data points.

9. The method for detecting an indicator of acoustic neuroma in the subject person, wherein the recall rate for detecting the indicator is 0.70 or higher and the specificity is greater than 0.

22.

10. The method for detecting an indicator of acoustic neuroma in the subject person, wherein the recall rate for detecting the indicator is 0.80 or higher and the specificity is greater than 0.

22.

11. The method for detecting an indicator of acoustic neuroma in the subject person, wherein the recall rate for detecting the indicator is 0.90 or higher and the specificity is greater than 0.

22.

12. The index detection method according to claim 9, wherein the standardized audiogram-related information includes the absolute value of the difference between each left and right air conduction data point.

13. The index detection method according to claim 12, wherein each of the aforementioned air conduction data points includes at least one data point of approximately 2,000 Hz and approximately 3,000 Hz.

14. A standardization unit that standardizes audiogram-related information, which is information about the audiogram obtained for the subject, An index detection device comprising: a detection unit that detects an index of an acoustic neuroma in a target person by using standardized audiogram-related information as input to a trained model obtained by performing a learning process using data that is a combination of pre-acquired and standardized audiogram-related information and an index of the person's acoustic neuroma.

15. The sensorineural hearing loss detection device according to claim 14, which includes a sensorineural hearing loss identification unit that identifies whether the target person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma, by using standardized audiogram-related information obtained for the target person as input to a trained model obtained by performing a learning process using data which is a combination of pre-acquired and standardized audiogram-related information and information indicating whether the person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by a cause other than an acoustic neuroma.

16. The indicator detection device according to claim 15, wherein the sensorineural hearing loss identification unit includes at least one of the standardized audiogram-related information points at 125 Hz and 250 Hz.

17. The indicator detection device according to claim 16, wherein the recall rate of identification in the sensorineural hearing loss identification unit is 0.78 or higher.

18. The index detection device according to claim 17, wherein standardized audiogram-related information includes air conduction data points and bone conduction data points.

19. A standardization unit that standardizes audiogram-related information, which is information about the audiogram obtained for the subject, An indicator detection system comprising: a detection unit that detects an indicator of acoustic neuroma in a target person by using standardized audiogram-related information as input to a trained model obtained by performing a learning process using data that is a combination of pre-acquired and standardized audiogram-related information and an indicator of acoustic neuroma in that person.

20. The sensorineural hearing loss identification unit according to claim 19, which identifies whether the target person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by an acoustic neuroma, by using standardized audiogram-related information obtained for the target person as input to a trained model obtained by performing a learning process using data which is a combination of pre-acquired and standardized audiogram-related information and information indicating whether the person has sensorineural hearing loss caused by an acoustic neuroma or sensorineural hearing loss caused by an acoustic neuroma.

21. The indicator detection system according to claim 20, wherein the sensorineural hearing loss identification unit includes at least one of the standardized audiogram-related information points at 125 Hz and 250 Hz.

22. The index detection system according to claim 21, wherein the recall rate of the identification in the sensorineural hearing loss identification unit is 0.78 or higher.

23. The index detection system according to claim 22, wherein standardized audiogram-related information includes air conduction data points and bone conduction data points.

24. An input layer that takes an audiogram image or its constituent data as input, It consists of a neural network with an intermediate layer that processes input data, It includes an output layer that outputs the possibility of an acoustic neuroma, A trained model used in an index detection method, index detection device, or index detection system according to claim 1, 14, or 19.

25. A trained model according to claim 24, for causing a computer to function to output acoustic neuroma indices in response to an input audiogram image or its constituent data.

26. An input layer that takes an audiogram image or its constituent data as input, It consists of a neural network with an intermediate layer that processes input data, It includes an output layer that outputs the results of distinguishing between sensorineural hearing loss caused by acoustic neuroma and sensorineural hearing loss caused by other factors, A trained model used in an index detection method, index detection device, or index detection system according to claim 1, 14, or 19.

27. The trained model according to claim 26, for causing a computer to output a result that identifies sensorineural hearing loss caused by an acoustic neuroma and sensorineural hearing loss caused by a cause other than an acoustic neuroma, in response to an input audiogram image or its constituent data.

28. A standardization process to standardize audiogram-related information, which is information about the audiogram obtained for the subject, A program for causing a computer to function as an indicator detection device, comprising a detection step of detecting an indicator of acoustic neuroma in a target person by using the standardized audiogram-related information as input to a trained model obtained by performing a learning process using data that is a combination of previously acquired and standardized audiogram-related information and an indicator of acoustic neuroma in that person.

Images