What is the latest application example of medical AI? Doctor commentary [2020 version]
By TOKYO analytica-April 15, 2019
 

In recent years, the use of artificial intelligence (AI) has advanced rapidly in so many fields that there is no day without seeing new news about AI. The background lies in the development of Deep Learning technology, but of course the use of AI in medicine is no exception. On the other hand, medical care is a special area that directly deals with human life, so the effect of AI is a little different from other fields. This time, we will broadly explain AI in the context of medical care, what is AI in the first place, what kind of benefits it will bring, what kind of problems will be born at the same time, and the latest medical AI trends to watch I would like to introduce

1. What is AI in healthcare?
What is AI in the first place?
AI (Artificial Intelligence) literally means "artificial intelligence," but its definition varies from researcher to researcher. Roughly speaking, it is a program that reproduces the logical thinking that human beings have done on a computer. Until now, it was a very limited "intelligence" that could only give an answer under certain rules, but due to the advanced development of deep learning technology from around 2010, the program is autonomous from the given data. It became possible to build a judgment standard by learning to. This means that AI can build and classify its own criteria, even those that could not show clear criteria to humanity. What brought this system to medical treatment is the so-called “medical AI”. In medical care, even if we focus solely on the disease, many judgments such as onset risk evaluation, disease diagnosis, treatment selection, prognosis evaluation, etc. are required, but the judgment is complicated by the difference in the situation for each individual. Usually very difficult If a criterion is constructed based on a large amount of accumulated patient data and, for example, an optimal treatment method is presented for each individual, the benefits provided by medical AI will be extremely large for both patients and medical personnel. I can imagine that.

What is Deep Learning?
Deep learning is the core of recent AI technology development, but it is difficult to catch up with all of the new technologies that are being released literally day by day. Here, I will give a brief explanation of these basic terms and examples of their use in medicine.

There is an algorithm called a neural network (which shows the procedure) that has been studied since the mid-20th century. This is a model of cranial nerve cells in humans and other living organisms, and has a structure in which the layer structure such as the input layer, the hidden layer, and the output layer are connected by edges. By giving a function to each layer (called an activation function) and giving weights to edges, a very complicated model for classifying input values ​​and outputting an answer is realized. This algorithm with many hidden layers (deep) is called deep learning. By the way, machine learning refers to "iteratively learning from given data and finding appropriate rules", and deep learning is also included in this. More specifically, let's say that you want to build an algorithm that predicts the onset of diabetes one year later from the current blood glucose level, blood pressure, and body weight. By inputting 3 points of blood glucose level, blood pressure, and body weight from a database that collects a lot of patient data, and setting whether or not diabetes was developed one year later as an output, edge weights etc. are determined through repeated learning. To do. The resulting optimal algorithm, given these three points, can predict whether a person will develop diabetes one year later. Such learning of the relationship between input and output is called supervised learning, which is frequently used in AI development in medicine.

In general machine learning, the feature amount (three points of blood glucose level, blood pressure, and body weight in the above example) is arbitrarily selected and input, whereas in deep learning, the feature amount that is most useful for determining the output. Even it can be automatically extracted and learned. Convolutional neural network (CNN), which is known as a typical deep learning model, gives an image itself as an input, but it is not necessary to specify a concrete feature quantity. For example, an algorithm developed by Mayo Clinic in the United States (past article) identifies asymptomatic left ventricular dysfunction from the electrocardiogram waveform image itself. Of course, it is also possible to extract features from the electrocardiogram waveform (height/width of the waveform, etc.) and construct an equivalent algorithm, but the input information is limited by selecting any item (in the statistical field, "information Is discarded"), and it is highly possible that the accuracy of the CNN algorithm will not be exceeded. Currently, this CNN plays a leading role in the development of diagnostic imaging AI.

Medical AI is powerful but not universal
Artificial intelligence seems to imitate human intelligence, a substitute for brain function, but in many cases this is overstated. To put it simply, AI at the moment is just something that can identify something specific. That is, if the algorithm can identify the blood vessel from the image, it will be returned as "blood vessel" if the image of the artery is shown, and "not blood vessel" if the image of hair is shown. Without making any changes to this algorithm, we cannot even distinguish between a newborn and an adult. Therefore, almost all AI, which has been popular around the world, just extracts "specific and very limited functions" of human beings and reproduces them on a computer, which is not all-purpose.

However, this reproduced function is so powerful that it sometimes greatly exceeds the discrimination ability of humankind. For example, it has been reported that an algorithm for identifying pediatric diseases from electronic medical record records exceeded the diagnostic accuracy of pediatricians for multiple diseases including influenza (past article). The results of this research have been published in the prestigious academic journal Nature Medicine and have attracted attention. Also, because AI can capture even minute changes that cannot be discerned by the human eye, it is highly compatible with medical images such as malignant tumor diagnosis, and technological development in the field of radiology is remarkable (past article). In addition, AI has the advantage of continuing to deliver stable results. Particularly in Japan, there is a harsh situation where doctors' judgments are blunted, such as overload of work due to lack of personnel, frequent calls at night and on holidays, and shifts to regular work. AI support that maintains a certain level of output accuracy could be a form of insurance for doctors.

2. What medical fields can AI utilize?

Photo by iStock
Next, I would like to look at medical fields that can utilize AI. Medical treatment is roughly divided into three steps: “prevention” to prevent the onset of disease, “diagnosis” to identify people who already have the disease, and “treatment” to improve the turning point of people with a diagnosis. AI is not only able to contribute to all of these three major steps, but is also being used in medical systems, including medical insurance systems and medical provision systems, and is expected to be used in all medical fields. ing. Behind this is the lack of human resources associated with the high degree of expertise of medical professionals, and many countries are trying to find a way out for AI.

Utilizing AI in diagnostic imaging
Radiology, especially the process of diagnosing diseases from medical images is known as the most important application of medical AI. The number of medical images that require interpretation at each facility is increasing due to the development of medical peripheral technology and the aging of society, but the number of radiologists accompanying it is not sufficient. AI can directly reduce the work load by helping the radiologist's interpretation. However, it is important to know that the number of AI devices currently approved as legitimate medical devices is very limited. In other words, even if AI diagnoses that "it has a malignant tumor," at present, there is almost no progress from diagnosis to treatment without the confirmation of a doctor. This is also a problem with AI, which will be described later, but it is due to lack of verification of the validity of the algorithm and the delay in the development of laws regarding AI. No alternatives have been found.

On the other hand, it is highly likely that AI will completely replace chest X-ray interpretation and ECG analysis in medical examinations in the near future. This is because the image interpretation in the medical examination is the screening, and the definite diagnosis is obtained in the subsequent individual medical examination. False negatives (determining “no disease” despite the presence of a disease) poses a problem in screening, but adjustment of the algorithm can sufficiently avoid this problem. AI, which is capable of producing consistent results with insomnia and rest, will be positively accepted by this type of screening agency in terms of manpower and cost. Practical examples of eye disease screening in the UK and China have been introduced in the past (Past articles 1 and 2).

Utilization of AI in disease diagnosis
With the rapid development of natural language processing technology in recent years, medical record analysis has become more popular. As a result, the disease diagnosis AI based on medical records has greatly improved its accuracy. In the past, we also introduced AI for diagnosis of pediatric diseases (past article), medical chart analysis system by Amazon (past article), so please refer to it. The medical record contains not only the findings of the doctor, but also all test results and prescription records. The longer the patient's medical history, the larger the medical records will be, and it will be difficult for even a good doctor to capture all of them during the limited medical treatment time. From the aspect, the use of AI will be effective. Similarly, there is an example of developing an AI system that integrates not only medical records but also biometric sensors and monitor records (past article). This system at the University of Florida in the United States predicts serious pathological changes and the occurrence of fatal diseases in the intensive care unit (ICU), and it can be said that AI is an effective application example. Furthermore, the development of a method for more easily diagnosing diseases such as sleep apnea syndrome that require specialized tests for diagnosis by using AI is proceeding (past article).

Utilization of AI for various medical problems
Utilization of AI is not limited to actual clinical situations. It is also used for vulture journal issues that undermine the credibility of medical science (past article), for combating doping in sports (past article), and even for efforts to prevent bioterrorism (past article). ..

3. What are the problems and issues of AI in the medical field?

Photo by iStock
Lack of AI validation
Many medical researchers continue to be skeptical of AI algorithms (past articles). The opinion is that only the high accuracy of the algorithm is distracted and the essential effectiveness is left behind.

When building an algorithm, generally only one dataset is used. This data set is divided into, for example, 80% and 20%, one side is a learning set for algorithm construction, and the other is a test set. In other words, it is a method of confirming whether the algorithm derived from the learning set exhibits the same accuracy in the test set and concluding that the algorithm is valid if the accuracy is maintained. However, the true validity has not been verified by this method alone. This is because the algorithm is only derived from "a specific population data", and its accuracy may not be maintained if the target population is changed. To give a simple example, the algorithms derived from the British-centered data set are not always effective in the Japanese. In fact, Amazon's proud face recognition AI, Rekognition, is also causing a big turmoil based on the research findings that "I have a bias that I can not distinguish black women well" (see CNN). There is no doubt that precise verification based on conventional medical evidence construction, such as multi-center analysis with different target populations and prospective follow-up studies, is required.

Delay in legislation surrounding medical AI
Over the last five years, the use of AI in medical care has progressed rapidly, but the technological development has been so rapid that necessary legal arrangements have been made

Running late. As a matter of fact, there is not enough discussion about how to handle the diagnostic results shown by medical AI. At the moment, by adding the phrase "always require confirmation from a doctor", it can only be seen as being introduced as a support system in clinical settings. It is a minimum to clarify and comply with the guidelines for the construction of essentially effective algorithms, the requirements for the AI ​​system to be approved as a medical device, and certain restrictions on those without approval. It has been demanded. It is very dangerous that medical AI of unknown effectiveness is widely available on the market without any restrictions.

Medical staff lacks knowledge about AI
It cannot be prescribed without knowing the effect and mechanism of action of the drug. Similarly, medical personnel such as doctors will need some knowledge of AI in the future. The need for AI subjects as a basic education for medical students has begun to be discussed, and an example of Boston University was introduced before (past article). In the future, as AI in healthcare becomes more pervasive and will become (and likely will be) involved in all healthcare processes, it will be very difficult for healthcare professionals to avoid AI.

Four. Up-to-date medical AI trend of attention

Photo by iStock
Drug discovery by AI
Pharmaceutical giant GlaxoSmithKline and British startup Exscientia announced that they have discovered compounds that could be used to treat chronic obstructive pulmonary disease (COPD) (past article).
This uses an AI-based compound search platform, which could be an epoch-making act in AI drug discovery. Compared with the conventional process, using AI not only makes drug development much more efficient, but it is also effective in exposing targets that have been overlooked even though there is a possibility of drug development. I have already introduced the automatic search technology of synthetic routes by AI, so please refer to it (past article).

In February 2020, Exscientia and Sumitomo Dainippon Pharma announced that they would start phase I clinical trials of new drug candidate compounds developed using AI. The entry of AI-guided compounds into human clinical trials represents a milestone in drug discovery.

Use for diagnostic imaging at death
There is an approach called diagnostic imaging at death. The purpose is to investigate the condition and cause of death at the time of death from CT and MRI images after death (past article). It is not realistic for a radiologist who has been pointed out that there is a shortage of personnel even with biometric images, and forcing him to read images during death imaging. In addition, not many radiologists are good at reading such special areas. Utilization of AI in such situations is considered to be extremely useful. In fact, it has been pointed out that AI-based analysis of post-mortem brain MRI images of patients with neurodegenerative diseases may lead to the development of new treatment methods.

Use for medical devices
A wave of technological innovation is coming to the stethoscope, which is a traditional medical tool. StethoMe's AI stethoscope is known as an EU-certified medical device intended for home use. We are constructing a system in which abnormal breathing sounds of patients are caught early and information is shared with clinicians along with the diagnosis results by AI. StethoMe announced a partnership with major telemedicine providers in Europe in April 2020, and is expected to further expand its market share (past article).

Efforts to replace the blood sampling technique with robots have also begun. A blood sampling robot developed by a research team at Rutgers University, NJ, combines deep learning with infrared and ultrasound imaging to identify blood vessels in tissue. Then perform complex visual tasks such as motion tracking to puncture the blood vessel with the needle. It is expected that robotic vascular access will be equivalent to or superior to human procedures, such as 88.2% success rate of initial puncture even if the blood vessel is in poor condition such that veins do not stand out. article).

"Edge AI" gains importance in medical care
Cloud computing and IoT have rapidly spread in all areas, but in recent years, “edge computing” has attracted attention because data processing is performed at a location (edge) closer to the end of the system. The one that implements an AI model on an edge device close to the site is called "edge AI", and some have completed the processing from learning to inference only on the edge side without using the cloud.

So what are the advantages of introducing edge AI in healthcare? The first point is to improve network connectivity. In other words, compared to a system that performs all processing on a cloud basis, by partially entrusting the edge device with processing of a large amount of on-site data, it is possible to reduce the direct load on the host system and network. Since the stage of healthcare is not limited to urban areas where IT infrastructure is well developed, its usefulness is remarkable even in remote areas where medical depopulation is likely to occur. Further, in telemedicine and robotic surgery, which are expected to grow even more in the future, strict control of latency is indispensable. It is easily assumed that the use of edge AI will be necessary to meet the highly demanded standard.