Investors are investing in the medicine of the future - neural networks and medtech startups. However, the success of both is still modest. Svetlana Pervykh, co-founder of Avare and Intellix medical tech projects and Medical Business Advisor, explains why.
We live in an era of total digitalization, and it seems that robots are about to start working, or at least neural networks, which will replace people in almost all professions: they will teach and treat us. Startups are multiplying, using the latest technologies and ready to solve all the problems of mankind at the wave of a hand. But is it that simple? Will the role of a human being be reduced to a data operator and the one who forms the ToR for the "artificially intelligent brother". No, medical high-tech projects will not remain without human participation for a long time, and to be more specific - without a practicing physician. Let's see why.
"Almighty" neural network
Neural networks have allegedly learned to draw pictures better than artists, and to write texts better than writers and journalists. We enthusiastically test new chips, actively communicate with robotic voice assistants, trying to drive the unfortunate artificial interlocutor into a logical hole. It's fun, but if we get sick and we are offered a treatment, in which the AI took part in drawing up the plan, the question arises whether to trust high technology in treatment.
The first thing that comes to mind are the newspaper headlines that tell us that a neural network has diagnosed the condition better than a human. A machine that does not know fatigue, they say, will accurately determine what a person is in pain and prescribe unbiased treatment. But what is the true role of neural networks in medicine?
Today, neural networks solve several types of problems.
1. Working with massive amounts of patient-side data from both diagnostic and treatment areas - e.g., X-rays, CT scans, MRIs. Identifying patterns across tens of thousands of available images.
2. Solving diagnostic tasks in a bundle "neural network plus some engineering device", for example, when a neural network helps diabetic patients to analyze their current state in time and remind them to take food or medicine.
In fact, in most cases, the task of a neural network is to highlight patterns either non-obvious or statistically significant.
For example, there is a popular anti-diabetic drug that is widely used for weight loss. Neural Network analyzed the drug's non-direct use (i.e., not for people diagnosed with diabetes), and it became apparent that it had several side effects: sleep disturbance and depression. These were not mass cases, and it was only thanks to a neural network that collected big data that these side effects became apparent.
3. The neural network can assist the doctor during surgery, such as laparoscopic surgery, where the surgeon operates by essentially looking into a monitor.
Improving image quality, emphasizing small details that may not be perceived by the human eye due to poor lighting, extremely small size, or even stressor moment/human error.
A parallel connected neural network tells the doctor when to pay attention to certain points.
4. Neural networks are helping to solve issues that have not been solved in medicine until now. Some problems are moving from the category of unsolvable to "we still know very little about the human body". First of all, we are talking about oncology. Thanks to neural networks, it has become possible to form a large pool of new molecules that, taking into account the data on each type of cancer (we already have them), should be effective in treatment. Without the neural network, a human would have to work on this for another five to ten years.
In each of these cases, the neural network is a great assistant to the physician, but not the decision maker. That decision is always made by the physician. The neural network doesn't make the diagnosis, it helps the doctor get more arguments for or against a particular symptom.
A good example. Everyone who has had covid is well aware of the frosted glass effect - a symptom on a CT scan of the lungs, a characteristic of a covid lung lesion. This symptom does not appear immediately: a person may already be sick and risk his life, but there is no effect. Or vice versa: a person has recovered, but the effect remains. If we entrusted the neural network to make a diagnosis based on this symptom, we would get either too late hospitalization or chronization of the disease, i.e. too long stay in the hospital.
So all those who dislike high technology can sleep easy: it is not the neural network that diagnoses you. And techno-optimists will have to wait. And, it seems, for a long time.
The pantheon of technological gods
Medicine is one of the most tasty markets for startups. Everyone gets sick, no doctor is ever unemployed, solutions for medicine are in demand and, let's face it, expensive. Hence the upside: a startup that solves an important medical problem will definitely make good money. And the downside: entering this niche is quite expensive, so they prefer to save money not on engineering staff, without which a startup is not a startup, but on specialized medical staff.
Here it is worth recalling the history of many startups that used an exclusively engineering approach. They lost millions of dollars and closed down. The engineering type of thinking assumes a strict logic based on repeatability: what has worked a million times will work a million times for the first time, and the data available closes all questions and forms all answers.
In medicine everything is different, there are always atypical reactions, side effects, individual intolerances, not to mention the fact that today we actually know only about 20% of information about the human body, some of which continues to be transformed and refuted - otherwise scientific discoveries would not occur so often. The same neural network can operate only with the available data, which can be questioned after the next discovery. For example, it happened with the protein theory of Alzheimer's disease - it is believed that we have been on the wrong track for the last 15 years because of one wrong publication. Beyond that, medical knowledge is a huge body of knowledge that takes a physician at least 12 years to master.
Lack of knowledge leads to anecdotal situations and potentially millions lost. For example, a molecule was developed that was supposed to be a drug in the treatment of macular dystrophy (a disease of the retina that results in decreased acuity or loss of vision). The project has already moved on to studies in rabbits, and good results have been obtained. One nuance: rabbits don't have a macula. So, in essence, it can be compared to prostate treatment in women.
Or, for example, a project that was developing protein molecules, decided to apply them to develop a drug to treat cancer, but lacked the completeness of medical knowledge about the peculiarities of the course of the disease in all its types. It became clear that one type of protein molecules could not solve the issue with all types of malignant formations. Besides, it was not taken into account that their development includes the stage of clinical trials lasting from ten years.
Another example of a "failed" idea. At a certain point we all faced the need for import substitution, some medical devices need consumables that are currently unavailable. There was a project that wanted to produce one of the consumables in the field of ophthalmology. The cost of one such consumable is a few rubles. It is impossible to produce any serious volume by hand, and to create several thousand pieces, as it turned out, you need a mold costing hundreds of thousands of rubles. To pay off the project on the Russian market (because even neighbors from the CIS will continue to use Western consumables), you need a little over 70 years.
And there are hundreds and thousands of such business projects. Of course, they will never reach the end consumer with their "goods and services", but will rather lose money and time and eventually close down. Medicine is a tough nut to crack, and technostartups should take it by storm, taking into account all the risks and a full baggage of knowledge.
We live in an era of total digitalization, and it seems that robots are about to start working, or at least neural networks, which will replace people in almost all professions: they will teach and treat us. Startups are multiplying, using the latest technologies and ready to solve all the problems of mankind at the wave of a hand. But is it that simple? Will the role of a human being be reduced to a data operator and the one who forms the ToR for the "artificially intelligent brother". No, medical high-tech projects will not remain without human participation for a long time, and to be more specific - without a practicing physician. Let's see why.
"Almighty" neural network
Neural networks have allegedly learned to draw pictures better than artists, and to write texts better than writers and journalists. We enthusiastically test new chips, actively communicate with robotic voice assistants, trying to drive the unfortunate artificial interlocutor into a logical hole. It's fun, but if we get sick and we are offered a treatment, in which the AI took part in drawing up the plan, the question arises whether to trust high technology in treatment.
The first thing that comes to mind are the newspaper headlines that tell us that a neural network has diagnosed the condition better than a human. A machine that does not know fatigue, they say, will accurately determine what a person is in pain and prescribe unbiased treatment. But what is the true role of neural networks in medicine?
Today, neural networks solve several types of problems.
1. Working with massive amounts of patient-side data from both diagnostic and treatment areas - e.g., X-rays, CT scans, MRIs. Identifying patterns across tens of thousands of available images.
2. Solving diagnostic tasks in a bundle "neural network plus some engineering device", for example, when a neural network helps diabetic patients to analyze their current state in time and remind them to take food or medicine.
In fact, in most cases, the task of a neural network is to highlight patterns either non-obvious or statistically significant.
For example, there is a popular anti-diabetic drug that is widely used for weight loss. Neural Network analyzed the drug's non-direct use (i.e., not for people diagnosed with diabetes), and it became apparent that it had several side effects: sleep disturbance and depression. These were not mass cases, and it was only thanks to a neural network that collected big data that these side effects became apparent.
3. The neural network can assist the doctor during surgery, such as laparoscopic surgery, where the surgeon operates by essentially looking into a monitor.
Improving image quality, emphasizing small details that may not be perceived by the human eye due to poor lighting, extremely small size, or even stressor moment/human error.
A parallel connected neural network tells the doctor when to pay attention to certain points.
4. Neural networks are helping to solve issues that have not been solved in medicine until now. Some problems are moving from the category of unsolvable to "we still know very little about the human body". First of all, we are talking about oncology. Thanks to neural networks, it has become possible to form a large pool of new molecules that, taking into account the data on each type of cancer (we already have them), should be effective in treatment. Without the neural network, a human would have to work on this for another five to ten years.
In each of these cases, the neural network is a great assistant to the physician, but not the decision maker. That decision is always made by the physician. The neural network doesn't make the diagnosis, it helps the doctor get more arguments for or against a particular symptom.
A good example. Everyone who has had covid is well aware of the frosted glass effect - a symptom on a CT scan of the lungs, a characteristic of a covid lung lesion. This symptom does not appear immediately: a person may already be sick and risk his life, but there is no effect. Or vice versa: a person has recovered, but the effect remains. If we entrusted the neural network to make a diagnosis based on this symptom, we would get either too late hospitalization or chronization of the disease, i.e. too long stay in the hospital.
So all those who dislike high technology can sleep easy: it is not the neural network that diagnoses you. And techno-optimists will have to wait. And, it seems, for a long time.
The pantheon of technological gods
Medicine is one of the most tasty markets for startups. Everyone gets sick, no doctor is ever unemployed, solutions for medicine are in demand and, let's face it, expensive. Hence the upside: a startup that solves an important medical problem will definitely make good money. And the downside: entering this niche is quite expensive, so they prefer to save money not on engineering staff, without which a startup is not a startup, but on specialized medical staff.
Here it is worth recalling the history of many startups that used an exclusively engineering approach. They lost millions of dollars and closed down. The engineering type of thinking assumes a strict logic based on repeatability: what has worked a million times will work a million times for the first time, and the data available closes all questions and forms all answers.
In medicine everything is different, there are always atypical reactions, side effects, individual intolerances, not to mention the fact that today we actually know only about 20% of information about the human body, some of which continues to be transformed and refuted - otherwise scientific discoveries would not occur so often. The same neural network can operate only with the available data, which can be questioned after the next discovery. For example, it happened with the protein theory of Alzheimer's disease - it is believed that we have been on the wrong track for the last 15 years because of one wrong publication. Beyond that, medical knowledge is a huge body of knowledge that takes a physician at least 12 years to master.
Lack of knowledge leads to anecdotal situations and potentially millions lost. For example, a molecule was developed that was supposed to be a drug in the treatment of macular dystrophy (a disease of the retina that results in decreased acuity or loss of vision). The project has already moved on to studies in rabbits, and good results have been obtained. One nuance: rabbits don't have a macula. So, in essence, it can be compared to prostate treatment in women.
Or, for example, a project that was developing protein molecules, decided to apply them to develop a drug to treat cancer, but lacked the completeness of medical knowledge about the peculiarities of the course of the disease in all its types. It became clear that one type of protein molecules could not solve the issue with all types of malignant formations. Besides, it was not taken into account that their development includes the stage of clinical trials lasting from ten years.
Another example of a "failed" idea. At a certain point we all faced the need for import substitution, some medical devices need consumables that are currently unavailable. There was a project that wanted to produce one of the consumables in the field of ophthalmology. The cost of one such consumable is a few rubles. It is impossible to produce any serious volume by hand, and to create several thousand pieces, as it turned out, you need a mold costing hundreds of thousands of rubles. To pay off the project on the Russian market (because even neighbors from the CIS will continue to use Western consumables), you need a little over 70 years.
And there are hundreds and thousands of such business projects. Of course, they will never reach the end consumer with their "goods and services", but will rather lose money and time and eventually close down. Medicine is a tough nut to crack, and technostartups should take it by storm, taking into account all the risks and a full baggage of knowledge.