Predicting the diagnosis of Alzheimer's disease, what thorns did Yassen Technology have stepped through?

( Note: This article was transferred from Lei Feng., author Liu Wei. )

"This is related to our genes." When it comes to why Alzheimer's disease (AD) is chosen as the main scene for AI landing, Yassen CEO Chen Hui said.

In recent years, many AI medical companies have emerged in the industry, many of which have chosen to use lung nodule screening as an entry point. Chen Hui believes that this is traceable.

"Most AI+ medical startups use deep learning algorithms, and first have an algorithm to find the scene, so they will choose lung cancer with high spatial resolution and high incidence as the entry point." Chen Hui said.

Founded in 2006, Yasen Technology is an enterprise dedicated to intelligent analysis of diseases through brain imaging and nuclear medicine equipment. Nuclear medicine imaging is mainly used for the diagnosis and analysis of brain diseases such as epilepsy, Parkinson's disease and lethargy. Alzheimer's disease is also one of the brain diseases, so it is quite reasonable for Yasen Technology to get involved in the diagnosis of Alzheimer's disease.

Early diagnosis of Alzheimer's disease is very difficult

Alzheimer's disease can be intervened early in the disease, but testing is relatively difficult. Researchers around the world are developing methods to detect Alzheimer's disease as early as possible. The sooner the condition is detected, the more likely the patient will be able to seek treatment early, slow the impact of the condition, and have enough time to handle the personal legal and financial situation.

Yang Shizhen, director of research and development at Yasen Technology, said that the main difficulty in diagnosing Alzheimer's disease is that the images are invisible, early symptoms are difficult to grasp, and long-term predictions are not easy to achieve with a single image and pathology.

The traditional diagnostic methods for Alzheimer's disease are mainly the following:

First, the psychological scale assessment. The neurologist will use the psychological scale to ask the patient's recent living environment, etc., to assess whether the cognitive function has declined.

Second, nuclear magnetic images. Check whether the image structure of the patient's brain has begun to shrink and change.

Third, through the long-term monitoring and analysis of EEG and heartbeat, to determine whether the patient has cognitive changes and changes in brain signals.

In addition, there are more advanced nuclear medicine, PET and other means to check the metabolism of the brain and determine whether some areas of the brain have decreased metabolism.

The cycle of Alzheimer's disease is very long and can be as long as 8-10 years. Based on data from a single point in time, the doctor can only make a judgment as to whether the patient is currently suffering from Alzheimer's disease. In addition, Alzheimer's disease is very extensive, and other brain diseases such as Parkinson's disease may also cause early symptoms. Therefore, it is important to track patients for long-term.

However, the diagnosis of Alzheimer's disease is time consuming and labor intensive. Take the scale assessment as an example. The assessment of the scale usually requires the presence of a neurologist and a psychologist at the same time. The assessment takes about 2-3 hours. In addition, doctors need to spend a lot of time discussing and evaluating the patient's condition. Taken together, a doctor can diagnose only two to three patients a day. Time costs are high and the number of patients served is limited.

In the long-term cooperation with the hospital, Yasen Technology has discovered a strong need for neurology to diagnose Alzheimer's disease as early as possible and improve the diagnostic efficiency. This has become one of the important driving forces for Yasen's diagnosis of Alzheimer's disease.

Multimodal analysis improves diagnostic effectiveness

According to Chen Hui, the complexity of artificial intelligence and related algorithms used in the diagnosis of Alzheimer's disease far exceeds the recognition of lung nodules.

He said: "Most of the companies on the market who do lung nodule identification use the same algorithm, based on TensorFlow or other open source machine learning platforms to make dimensional adjustments. And they are not doing the diagnosis of lung cancer. It is the image recognition part of lung cancer diagnosis, mainly for the identification of lung nodules above 6mm. The existing deep learning algorithm has no way to go beyond the hospital to solve the rapid identification and diagnosis of lung nodules less than 6mm. In other words, the image has a limited proportion of clinics, but also to see tumor markers, metabolic images and pathological results."

It is also difficult to diagnose Alzheimer's disease by looking at nuclear magnetic, EEG and scale data alone, so Yasen Technology uses a multimodal analysis method.

In the past ten years, the informationization of hospitals in China has developed very rapidly. The level and concentration of informationization in hospitals in some second-tier cities is beyond imagination. A series of diverse data such as CT data, PET data, bioelectric signals, and ultrasound over the past decade have been very large. But many of these precipitated data did not play their own value.

Chen Hui believes that the hospital has deposited so many types of data that can be comprehensively applied to clinical examination and treatment.

Yasen Technology uses three aspects of data to make a machine learning model for the diagnosis of Alzheimer's disease, including nuclear magnetic data, EEG data and scale data. Based on these three data, the multimodal neural network training model can predict the possibility of Alzheimer's disease and confirm the stage of disease development two or three years in advance. In other words, Yasen's Alzheimer's diagnosis products have surpassed a single doctor and can truly diagnose and predict diseases.

Data collection is the biggest challenge

Machine learning relies on massive amounts of data, and data scarcity is a major bottleneck that restricts AI technology from reaching the medical industry.

It is reported that Yasen Technology has reached strategic cooperation with more than a dozen hospitals and related institutions such as Xuanwu Hospital, Peking University People's Hospital and China-Japan Friendship Hospital, and the latter supplies data to them. The two sides mainly cooperate in the research and development of horizontal projects. The horizontal research and development agreements signed must go through the ethical review of the hospital, and then Yasen Technology can organize and use the data to ensure the security of the data.

“Data collection for diseases such as Alzheimer's disease and lung nodules is not the same. Alzheimer's disease is a degenerative disease that requires long-term tracking and data collection for patients. We have tracked at least 5 of these data for a long time. More than a year. We have to re-examine the follow-up patients every year, including scales, nuclear magnetic images, PET images, EEGs, etc. We have to re-evaluate the manpower, resources and costs in this process. Most of the companies that do lung nodule research." Yang Shizhen added.

One of the major difficulties in data collection is that the patient's fit is relatively low. For early patients, the doctor diagnosed that he was at risk of illness and often felt that he was not ill because he did not affect his daily life. When the hospital asks him to undergo an examination again, he tends to refuse.

The situation of the later patients is more troublesome because the patient's condition deteriorates very quickly. "There is an 80-year-old woman who had just diagnosed Alzheimer's disease the previous year and died the following year." Having said that, Yang Shizhen’s tone is full of regrets.

Existing data is sufficient to support product development

Yang Shizhen believes that although Yasen Technology does not seem to have a large amount of data, it is enough to support the research and development of products, and has reached a high level.

"In our ability, of course, we want to collect as much data as possible. But from the data published by other teams, they have more than 100 data."

For example, not long ago, Italian researchers announced the development of an algorithm that could detect minor structural changes in the brain 10 years before the onset of symptoms of Alzheimer's disease. The MRI scans they used to train artificial intelligence were only 67 – 38 from Alzheimer's and 29 from healthy people. It is reported that researchers divide the scan into small areas and let their artificial intelligence analyze the connections between neurons. After the training, they used this artificial intelligence to test the brain scan of 148 subjects. Forty-eight of the subjects had mild cognitive impairment and eventually developed Alzheimer's disease. In this test, the algorithm detected a success rate of mild cognitive impairment of 84%.

“We started data collection work in 2008. Compared with many large research institutions, the data we have mastered is very complete,” said Yang Shizhen.

Clinical experiment reverse product iteration

There are often differences between the development and clinical applications of AI medical products. At present, Yasen Technology's Alzheimer's disease diagnosis product has undergone a clinical experiment for about half a year in a top three hospital, and Chen Hui shared his feelings.

He said: "The biggest difference between product development and clinical application is not whether the doctor approves your inspection process and methods, but the standards and targets of the clinician's data collection are not perfect. Does the doctor collect the scale according to regulations? Nuclear magnetic scanning is carried out in strict accordance with the quality control standard. Whether the image noise is too large will affect the use of the product."

Chen Hui said that in theory, all hospitals that can collect data according to quality control standards can use Yasen's Alzheimer's disease diagnosis products, but the main users in the short term are regional top three hospitals.

In order to further accelerate the product landing, Yasen Technology is doing a lot of new attempts and efforts. “In order to realize the true meaning of our products, Yasen Technology will provide a quality control service for the hospital. That is to say, in which hospital our products are located, the hospital must collect the raw data according to the standards we provide. Only In this way, our products can really make a difference."

In addition to discovering problems, Yasen Technology's Alzheimer's diagnosis products have also been recognized and praised by many doctors.

Chen Hui said that after the doctors in the Department of Neurology and Nuclear Medicine used Yasen's Alzheimer's diagnosis products, they all responded well. And they put forward a lot of constructive opinions in the process of using, such as adding some new data sources. Chen Hui revealed that Yasen Technology is currently discussing with a doctor in a top three hospital whether it is necessary to record a patient's voice and answer the question's voice fluctuations and hesitation time as whether he has dementia and mild cognitive impairment. reference.

Chen Hui also said that in addition to cooperative hospitals, many high-end medical examination centers and institutions engaged in the screening and treatment of geriatric diseases have also shown strong interest in Yasen's Alzheimer's disease diagnosis program.