Industry 丨 Google Enters The Venue, The First Batch Of FDA, Acceleration Of Acoustic Markers AI Accelerates

The physical structure of human organs will change with the change of physiology and pathological state, which leads to specific changes in the sound and organ of the organs themselves.[Sound logo].

Acoustics logo is applied to clinical feasibility improvement

A simple exampleYes, the sound of a cold patient will become hoarse, which is a manifestation of changes in sound characteristics.

With the advent of the digital age, a hesteician, a medical tool with a history of more than two hundred years, has glowed with new vitality.

Digital listeners convert sound waves into high -precision digital electrical signals through electronic technology, and after amplification and processing, the sound output of more clearer sound than traditional stethoscopes is achieved.

Coupled with the rapid development of artificial intelligence technology, the feasibility of applying biological acoustic markers to clinical clinical is gradually gradually from[Remaining] transforms to [tentacles can be available].

However, this process is not smooth sailing.External factors such as environmental noise will interfere with audio signals, affecting the extraction and analysis of cough sound characteristics.

Limited by hardware performance, it is difficult to capture high -quality sound signals in the past.

Fortunately, the rapid progress of sensor technology effectively makes up for this defect.

The new generation of electronic hesteations adopts a piezoelectric ceramic sensor as a pickup part. Compared with previous products, its signal quality and frequency response curve is better, it can provide more accurate and clear cardiopulmonary signals, and then accurately obtain user health data.

More importantly, as a sound logo acquisition equipment, smartphones have experienced many years of fierce competition and technological innovation, and their microphone pickup performance generally improves significantly, which is enough to meet the basic audio signal acquisition needs.

After overcoming the hardware barrier, the development of acoustic logo is entering the fast lane.

Google's big model ended, enter the acoustic logoAI field research

Earlier this year, Google released itHealth Acoustic Repance Repaidations (referred to as Hear) biological acoustic model, which aims to assist researchers to develop a system that can identify early disease signs by monitoring human voices.

GoogleThe Research team used a audio dataset containing 300 million to identify and processed to train Hear. Special use of 100 million cough sound data was used to train cough recognition models.

pictureThe solution like Hear will open up new possibilities for AI -driven acoustic analysis in the field of tuberculosis screening and testing, and provide a potential low -cost and easy to obtain tool for the most needed groups.

thisThe AI system has been trained by millions of audio editing samples. In the future, it is expected to assist doctors diagnose lung diseases including COVID-19 and tuberculosis, and evaluate personal lung function.

Hear's innovation lies in the huge training data sets it uses, and its ability to fine -tune a variety of different tasks.

Most artificial intelligence tools developed in this field are trained through audio records, such as cough recording, which match these recording with the health information of the soundr.

For example, these recording clips may be marked as a person with bronchitis during the recording.These artificial intelligence tools through supervision and learning (The training process of Supervised Learning associates sound characteristics with data labels.

In this study, Google's researchers used self -supervision learning (SELF-SUPERVISED Learning uses unbar data and uses the automation process to extract more than 300 million short fragments of cough, breathing, clearing throat and other human sounds from the open YouTube video.

Each fragment is converted into a visual representation of sound——Stram map.The research team subsequently shielded the clips of the spectrum diagram to help the model learn to predict the missing part.

This is with the chat robotThe large language model of ChatGPT is similar to the process of predicting the next word in a large number of human text sample training.

Using this method, the research team has built a basic model that can adapt to many different tasks.

Studies have shown thatHEAR's average performance in many tasks surpasses other models, especially in generalization of microphoneIn terms of capabilities, it shows its significant mode recognition capabilities in capturing sound data related to health.

based onHear training models can be achieved even if the amount of data is limited, which is of particularly important significance for the scarcity of data in the field of medical research.

HEAR has now opened to researchers to promote the rapid development of customized biological acoustic models. Even if data scarcity or facing cost and computing capabilities restrictions can also provide support for specific diseases and crowd research.

At the end of August, Google announced the first application project of the HEAR model and would cooperate with Indian enterprise Salcit Technologies to apply early screening of HEAR to tuberculosis.

EKO artificial intelligence algorithm won the first batch of FDA, further accelerating the development of the industry

In May this year,EKO announced its artificial intelligence software that complements the digital hesicathea- [EKO Low Blood Blood Score Tool (Elects)].

This software is identified as the US Food and Drug Administration (The first artificial intelligence algorithm approved by FDA) to assist early screening of heart failure marks a major innovation in the medical field, and then completed a $ 41 million round D financing in June.

passEko's artificial intelligence technology, doctors can detect heart low -ejected scores (also known as heart low EF values) in just 15 seconds with the heart sound collected by digital hesaters.Blood ability.

Bleeding score reduced heart failure (HFREF) is one of the main types of heart failure.

According to data, about half of patients with more than 6 million heart failure in the United States have reduced blood scores to reduce heart failure.

The combination of Eko's artificial intelligence algorithm with supporting digital hesteations can recognize heart murmurs. It is expected to identify patients with bleeding scores as soon as possible in the initial conventional examination.

The recognized patients can subsequently accept further inspection in order to intervene in time.

existFrom 2020 to 2023, Eko also received approval of a number of medical devices, which included not only the upgraded version of the digital hesteic fighter, but also the artificial intelligence algorithm that could generate heart sounds and analyzes it.

Other companies have also made substantial progress in this field

Another company that enjoys a high reputation in the industryTytocare completed a $ 49 million financing in August last year.

Its artificial intelligence algorithm isAt the end of July, further approval of the FDA was also obtained. This algorithm can detect the wet lung of the lungs of adults and children over the age of two, which helps to discover potential lung diseases early.

In addition to tuberculosis, asthma and chronic obstructive pulmonary diseases (chronic obstruction) are also the focus of biological acoustic logo research.

AustralianResap Health has long -term research accumulation and has been listed in Australia.

One of the two products under the company——SleepCheckrx, used to identify obstructive sleep apnea, has obtained FDA approval;

Another productRESAAAPPDX has obtained CE certification by analyzing cough and breathing sound, and has also obtained CE certification.

At the end of 2022, the pharmaceutical giant Pfizer acquired Resapp Health for $ 179 million.

For example, JapanAI AI as a AMI (Acute Medical Innovation), which can assist early signs of identifying valve heart disease (including aortic valve stenosis), and obtained the approval of Japanese medical devices in October 2022.

Tytocare completed $ 49 million in financing last August. Its artificial intelligence algorithm received additional approval from FDA at the end of July, which can be used to detect the wet lung of adults and children over the age of two.

Researchers in South Korea used reverse design technology based on evolutionary deep learning to successfully develop an advanced acoustic super material.

Ventilation acoustic resonator developed by the National University of Busan (VAR) Super material has dual functions that reduce noise and provide ventilation.

Traditional analysis methods are processing complex shapesWhen VAR has limitations, it can only achieve basic parameter design.

Busan's research team innovatively developed a reverse design method based on deep learning. This method can flexibly design complex non -parametersVAR, thereby improve performance and reduce calculation costs.

In the reverse design method proposed, the condition is automatically divided into automatic encoder (CVAE) —— a deep learning generating model -encoding the VAR geometric features in the potential space.

The potential space is a low -dimensional space, which contains high -dimensional input (in this example isVAR) core information.

The design method of this acoustic super material marks the advanced manufacturing of artificial intelligence -driven (AM) and other complex mechanical structure design areas have taken a key step.

The efforts and achievements of these companies jointly promoted the acoustic symbolsThe rapid development of the AI field has made this emerging field show huge potential and broad application prospects.

AcousticsAI faces a series of challenges during the development process

① High -quality audio data is very important for training accurate and effective artificial intelligence models, but at present, most medical institutions have not yet established a perfect audio database, which brings great difficulties to the training and optimization of models.

② Due to the complexity and variability of the sound characteristics, the performance of the model in different environments and people may vary, resulting in the resultsThe reliability is affected.In addition, there is still a lot of room for diagnostic accuracy.

Despite acousticsAI shows certain potential, but when facing complex diseases and individual differences, misdiagnosis or missed diagnosis may still occur.

③ On the basis of general audio features, the sound characteristics of cough and specific diseases (such as tuberculosis)Pathological characteristics and symptoms, designing accurate target characteristics of cough sound with specific diseases to achieve accurate distinction between patients with specific diseases, healthy people and other patients with lung diseases. This problem still needs to be resolved.

④ Generally, the audio data used in model training is more ideal, but in practical applications, the audio signals that need to be recognized will be disturbed, which will affect the extraction and analysis of the cough characteristics.

The performance differences of different pick -up equipment may also cause changes in cough audio quality and characteristics, which will affect the diagnostic results of the model.

Due to the difficulty of collecting and marking audio data, this requires the model to maintain a better detection performance under the conditions of a small amount of precision injection data.

How to solve the contradiction between small sample data and complex model learning is also a subject that needs to be resolved.

End:

In the process of model training, how to reduce the impact of noise through methods such as data enhancement and noise suppression, and use the field of generalization technology to eliminate differences in equipment to enhance the robustness of the diagnostic model.Essence