Shulou(Shulou.com)06/02 Report--

Author: Liu Haitao 2020-05-31 23:05 introduction: the reason why it is difficult to land comes from Pride and Prejudice.

At 23:55, the flag-lowering and flag-raising ceremony began, and the flag guardians of China and Portugal entered. 23:58, in the Portuguese national anthem, the Portuguese flag and the Macau City Hall flag began to lower slowly.

In 1999, Dr. Zhang was sitting in front of a color TV with his family in an alley in Shanghai, witnessing the arrival of this historical moment. Drop, drop, the BB machine at the waist shows a message "exposed, let's go".

With the arrest of the last suspect, Dr. Zhang, the first joint crackdown on insurance fraud by the Shanghai medical insurance department was officially solved, and all five medical insurance "moths" were arrested.

In an era full of Ponzi schemes, they did not expect that their exposure came from the burgeoning technology of "behavior recognition", which is today's artificial intelligence.

As one of the researchers of behavior recognition at that time, Dr. Tang Ziou, chief health economist of good Life Technology, said with a little pride: "compared with today's computing power, we were completely millet plus rifles at that time. We could only scan the data in the database logically. At that time, we joked that if we broke the computer, we might lose it. Because the value of the brushed data may not be as high as that computer. "

Dr. Tang Ziou, who has been in the industry for 26 years, has been exposed to insurance artificial intelligence risk control during overseas training in 2000, and then became the first Chinese North American health insurance manager, life insurance manager, director of the International Health risk Management Association (IHRMA), and founded China's first independent commercial health insurance company.

He told Lei Feng that the heat of AI insurance risk control seems to be very high, but the milestone has not come at all, and the applications stay in the shallow layer. Although the product gives us the feeling that although we do not understand, we feel that it is very powerful, but we find that it is nothing more than the need for license plate recognition at the door of the community.

The reason is not that the insurance industry does not need deep-seated demand at all, but that a group of academic entrepreneurs are technologically only, insist that "data is king", and are unwilling to accept the knowledge accumulated by the industry in the past. After the poor results of the products, they began to complain and look for all kinds of excuses for insufficient data, lack of calculation and low budget.

The following is the full content of the exclusive interview, Lei Feng net did not change the original intention of the editor.

Lei Feng net: AI has been applied in insurance risk control, medical insurance audit and other scenarios, whether it has achieved obvious results, what are the reasons for poor application?

Tang Ziou: AI must be combined with the characteristics of the industry in order to play a role. At present, the application of AI in the insurance industry has not played a significant role.

In fact, AI should be promising in the insurance industry, such as the exhibition of documents. In the insurance industry, there are a large number of insurance applications and claims that need to be processed every day. In this case, the use of AI to do the exhibition of these documents can greatly improve the efficiency and quality of the audit.

In addition, AI can be used to find outliers in large-scale data. It is often said that when things go wrong, there must be demons. The insurance industry is producing a large number of new data every minute and every second, and the audit of these new data will always make mistakes by relying on manpower.

Artificial intelligence is more sensitive to data than people, such as finding the mean and variance from a statistical point of view, and once a highly aggregated normal distribution is found, we need to pay attention to finding the cause, as well as anti-label data, such as gynecological medication in men.

But the current situation is that AI big data risk control simply did not play an effective application.

The reason is that it is sad that the original technology bulls, who have been pursuing technology, have not made effective use of the experience accumulated over the years in the industry and have not been combined with the needs of the industry.

Although the designed products give everyone the impression that although they do not understand, they feel that this is very powerful, but they find that the realization is nothing more than the need for license plate recognition at the entrance of the community.

In the end, this kind of AI start-up enterprises can not be listed for a long time, it is difficult to make profits, have a negative impact on the valuation of the whole market, and finally fall miserably.

To take an example, those academic entrepreneurs, in the face of the needs of insurance companies, directly said: "as long as you give me the data, I can solve anything for you."

But when the insurance company heard this sentence, it must have been stunned. Because this is from a technical point of view, "through data mining, the discovery of universal laws."

But from the perspective of actuarial science, first of all, the privacy and security of insurance data is very important, in addition, these health insurance data often have a lot of noise, directly take the data to find the rules, the rules are definitely unreliable.

After hundreds of years of development, insurance actuary has accumulated a lot of regular logic, and AI is more suitable for the insurance industry only when combined with these foundations.

However, academic artificial intelligence entrepreneurs must directly disbelieve this knowledge and insist that "data is king." this is also the main reason why the two sides cannot reach an agreement at present. Scholars who study deeply are not willing to understand actuarial knowledge, nor are they willing to go deep into it.

In the end, the risk control products developed by data greatly deviate from the market expectations, and this contradiction may exist for a long time in the future.

Lei Feng net: from the experience of good life, what are the criteria for insurance companies to judge whether AI risk control products are good or bad?

Tang Ziou: experience is human knowledge. Big data's mining result is computer knowledge. For AI products, ultimately serving human beings, the most important criterion is experience.

As a product that serves human production and life, AI should first play a specific role in human life. This cannot be achieved, and the rest is empty talk.

The specific scenarios of the application of artificial intelligence in good life include risk control underwriting, claims settlement, risk control and innovation. There is a representative case of the role of insurance companies.

An insurance company, which directly sent us more than 100,000 cases of data, asked to work out how much cost savings could be achieved within two days.

The requirements and evaluation criteria are very clear, but those academic startups with strong algorithms and strong modeling skills may not be able to complete such simple requirements.

The risk control model of a good life can run completely in two hours, and can identify 10% of the anomalies. Then analyze the 10% of the data, change the parameters and run again, and get 5% of the confirmed data. The final statistical efficiency is more than 95%.

With these 5% statistics, we can rely on actuarial experience and report by category to achieve the goal. Therefore, the most important thing is to be useful directly. If you spend all your energy on the algorithm, it is often in the mirror.

Lei Feng net: the main service object of insurance AI products is insurance companies. From the point of view of insurance companies, what is the demand for AI? What are the differences between different departments?

Tang Ziou: from the customer's point of view, insurance companies evaluate the advantages and disadvantages of products, different departments often have different evaluation methods.

First, the sales department's demand for artificial intelligence is to identify what kind of person will buy his product, so it is mainly based on behavior identification.

For this demand, good Life has also developed corresponding behavior identification products to judge its purchase tendency through past health behavior and medical behavior.

But whether the user buys insurance or not and which product to buy is not rational, but perceptual. When users buy insurance, few people realize what they really need, often pay the bill based on their basic knowledge and current feelings, and ultimately do not pass the needs adaptability assessment.

Second, two nuclear departments, namely, indemnity and underwriting. The assessment indicators of the two are not exactly the same, the goal of the underwriting department is not to let the bad guys in too much; the goal of the indemnity department is not to let the bad guys succeed too much.

So there are two needs right now:

First, retrospective risk control requires batch audit and management of the data of the past two years.

Second, online real-time risk control, the scene to decide whether the good guys or the bad guys.

During the barbaric growth of health insurance companies in the past two years, some companies released a large amount of water in order to get the number of users, no matter who they were, but they soon found that the compensation rate was too high, and then the stock price fell quickly. Began to pay attention to real-time risk control.

To sum up, the initial requirements of the two nuclear departments are real verification, and even if it is completed quickly online, there must be a nuclear compensation algorithm behind it.

However, although the goal is the same, the specific implementation is quite different.

Underwriting is based on previous data and behavior time series analysis, and compared with the standard value, and finally determine the size of the gap.

Nuclear compensation is a large-scale comparison of the data of different suppliers and demanders, in which the data that need to be identified are both the demander and the supplier, that is, medical behavior and medical behavior, so it is more complex than risk control.

Third, the product actuarial department, this demand can only be understood by going deep into the insurance company. Because the demand of insurance actuarial is to discover the universal risk law through AI, which is also known as the subject matter of insurance.

The object of protection has three main characteristics: it is generally concerned and recognized; the risk is uncontrollable; and the risk can be predicted.

This process must be supported by artificial intelligence. Twenty years ago, when I was an actuary, I relied entirely on manual work, and the speed was very slow.

Now, supported by artificial intelligence, the speed of innovation is gradually accelerated, the iterative cycle is gradually shortened, and even through artificial intelligence, the internal relationships of risk events that can not be associated in the past can be found one by one.

In the face of different internal demands of insurance companies, AI products can achieve value only if they are market-oriented.

Lei Feng net (official account: Lei Feng net): what are the practical cases of good Life in the past to prove that AI risk control plays an obvious role for insurance companies?

Tang Ziou: in the past, there was a group insurance case of a large insurance company. It had a group insurance order serving hundreds of thousands of large-scale enterprises. Although the premium for such a large number of customers was considerable, he was actually very miserable because of the heavy losses.

We are losing money every year, but we can't lose it. on such a large scale, once released to the market, it is enough to support a small insurance company, and this kind of chicken rib single case is not uncommon in the market.

In this case, it is indisputable that the retrospective type cannot be carried out for two years, and the logic adopted by good Life at that time is to carry out real-time risk control in the process of claim settlement. Each order is cut into the intelligent claim module, the risk control algorithm is embedded in the module, and the output and input of the API interface are made. After accounting in the safe house, the claim is settled and the result is output very quickly, which is real-time.

After a cycle of the project, we found that 80.9% of the claims were screened out from all claims, and 98% of them were basically correct after manual verification.

Solve a big problem for this chicken rib group insurance business, because in the final negotiation process with the employer, the profit is only between 23%, or the loss or profit is within this range, saving 80.9%, all the profits will come out and turn losses into profits.

Lei Feng net: insurance risk control is an important direction for AI landing, but some experts believe that in many cases, intelligent risk control does not need a complex machine learning model at all, and a simple decision tree or basic statistical model can solve it. What do you think?

Tang Ziou: first of all, this is a relatively common phenomenon, whether insurance, banking, or other applications of artificial intelligence industries.

At present, these applications are only shallow, so some experts say that a relatively simple decision tree can solve the problem.

The reasons why it is impossible to enter the deep layer or do not need complex models, first of all, it is because the long-term extensive development of the industry in the past, resulting in a large number of simple error data mixed with the so-called noise; secondly, professional questions, whether you have decades of experience, to see the deep-seated needs of those industries.

On the other hand, "does the insurance industry not need complex calculations and factors?" Actually, it is necessary.

From the point of view of the average profit margin of the market, General Shorty, today, when my product is one centimeter ahead of the industry, I begin to be complacent, but tomorrow I suddenly find that everyone has risen by one centimeter, and we are going to face a price war.

Therefore, in order to avoid a price war, we must make a long-term technical reserve.

Today's model is really good enough, because 99% of people have not yet realized the crisis, and insurance companies are even blurred by the arrival of artificial intelligence. But this situation will not last, let alone the situation where you can lie there and make money all the time using a simple model.

I think AI has only played more than ten percent of the energy of the insurance industry, and the last part is the most difficult to complete. the more you go on, the more you will find new needs and knowledge. once you know that others do not know, the more competitive the product will be.

But the more it goes on, the more difficult it is. The difficulty is that the cycle of discovering potential patterns is uncertain, which may be two months or two years.

Take the financial crisis as an example, we all know that it happens every decade because of financial innovation. But the demand cycle of insurance technology has not yet been discovered.

Moreover, human beings are good at forgetting. For more than three years, they may forget the reasons for doing this at first, and then they will go into intention and madness. Thunder explosion may be a matter of time and time, depending on who takes the plate before the thunderstorm, and finally suffers the bitter fruit.

As serious as the financial crisis, it takes more than a decade of data to find out. A risk control company, AI is just a means, may be applicable today, tomorrow there is another new method, the effective integration of data and experience is the core element.

Lei Feng net: since data is the core element, what are the problems in the use of data in the industry?

Tang Ziou: now the biggest problem in the use of data is that it only focuses on "quantity" and ignores the time series.

In order to accumulate a large amount of data, some artificial intelligence companies collect a large amount of data on cross-sections. in cooperation with insurance companies and government health insurance, they may get millions of data at once, but all within two years, and then claim to have a huge amount of data.

I am not optimistic about this kind of data accumulation, because the model of cross-sectional data training can only feedback the static law at this point in time, which may not work at all in two years.

From a statistical point of view, time series analysis is very important, which has a lot to do with human characteristics. Hegel said: the only lesson from history is that people have never learned any lessons from history.

People themselves have the characteristics of forgetfulness, blind conformity, easy to be anesthetized and biased, so people's real behavior needs to be observed by time series, not static only from the cross section.

Lei Feng net: if the current application of artificial intelligence only stays in the shallow layer, then if we develop to a deeper point of view, we will encounter problems such as excessive data noise and how to solve them.

Tang Ziou: health insurance risk control is not a trick to eat all over the world, generally speaking, it can be divided into three categories: managed medical care, chronic disease management, and health promotion.

Managed medical care focuses on managing the supply side, chronic disease management of the supply side and the demand side, and health promotion is mainly the management side, which has a wide range of underlying cognition.

At present, what AI does is simple behavior identification and error correction. At present, the average of these capabilities is 80.9%. It is estimated that it may reach 15.2% in the future, reaching the technology boundary, even if future technologies are added.

In order to really improve, the direction should be broadened, with the help of other technological categories to solve the traditional grey area problems, such as abuse, which are the knowledge and so-called experience of the industry in the past.

In 1999, I attended a class in the World Bank Institute training and said: there is no way to effectively identify human behavior, but directional analysis can be carried out to make up for the lack of computing power.

This view is also suitable for now, we can not be actuarial to a large amount of data from all angles and dimensions, the dependent variable may become an independent variable, and the final arrangement and combination of the data is so huge that even the current calculation power cannot be satisfied.

Therefore, if we want to go deep, we must rely on the efficient combination of the existing knowledge of the industry and the application of big data. Who does well in this combination will take the lead.

Lei Feng net: a lot of risk control depends more on human experience. Which do you think is more important, artificial or intelligent?

Tang Ziou: for quite a long time, human experience must be more important than intelligence. Let me cite an example. Do insurance technology companies know that insurance companies have hospital reimbursement catalogs and can be certified by hospitals? more deeply, whether you know that this catalogue is different from medical insurance, some special drugs have different claims scope, the same diagnosis, claims may or may not include surgery.

This knowledge must require years of experience accumulation, not only insurance, but also other industries, human experience will occupy a dominant position, and will begin to be dominated by industry experts in the development of artificial intelligence in various industries.

Because artificial intelligence is only a tool, effective use based on industry cognition is very important.

So I suggest that the algorithm scholars of artificial intelligence should stay on the university campus and do not come out to start a business, which is not interesting. They will eventually find that it is difficult to cash in when they really land in the industry.

The depth of understanding of the field of knowledge, that is, the supervised dictionary in model training, the so-called gold standard, the application of AI in all industries needs to be based on this origin. Intelligence first needs something reliable, which is human experience. Unsupervised learning is difficult to succeed in an industry with deep domain knowledge.

Lei Feng: for the long-standing problem of data standardization in health insurance risk control, whether the DRG technology being implemented can be changed, and what effect this technology can achieve on health insurance control fees.

Tang Ziou: this is the content of the scope of supplier risk control. I personally hold a negative attitude towards the application of DRG, not to mention whether it can survive in three years, I think it will explode in one or two years.

In 2000, when we were still in the Shanghai Medical Insurance Fund, we tried the DRG method and finally gave up the choice of global budget (Total Budget system) and multidimensional factor analysis. And these methods and principles, I later brought a good life, including now big data risk control, the key is that there are differences in methods with DRG.

The medical insurance fund audit programs that Haoshi participates in are all based on these big data theories, no matter they are later called DIP or big data risk Control and other names, as an alternative to DRG, they are also being gradually implemented.

What is the difference between DRG and DIP (Total Budget)? First of all, the biggest problem with DRG comes from the background conditions at birth, which was first created by Americans.

Its research and development background is not the era of a large number of online data, relying on a statistical table, according to the statistical point of view, step by step according to the first-level index, second-level index iteration, so it is based on statistical data, according to the index system.

The title of DRG is to analyze and diagnose the classified payment methods of related disease groups, gather the related diseases together, pay for classification, the strategy is rigid across the board, methodologically known as whitelist system, into the group system. This rigid one-size-fits-all membership system is bound to appear in the application of medical insurance, with serious spillover effects.

According to the actuarial logic of health insurance, 80% of whether to see a doctor is decided by the buyer, and 80% of the medical expenses incurred after seeing a doctor are decided by the supplier.

Therefore, the application of DRG, there will be part of the spillover costs, appear under other diagnostic categories. In the end, it was found that it was a sunny day inside DRG, the situation was very good, the outside was full of hardest-hit areas, and the total plate of health insurance funds was as hot as in the past.

The adoption of this strategy by the Health Insurance Bureau is also a forced act of helplessness. in fact, the ultimate demand is to control the overall plan of health insurance, as long as it is not exposed to achieve the goal, and further fee control is a better performance.

Therefore, the goal should be directly toward the total plate, around a circle on the line DRG, the total plate has no tube at all, so the effect within the DRG is very good, no matter where it is desperately used.

The goal of the total budget system is to point directly to the master plan, and now the technology has been achieved, big data analysis can be carried out in real time, and the computing power is not limited.

Big data's risk control engine of good Life outputs a full amount of data and multi-dimensional analysis every month. If you simplify the process, you can find the top 3%, the so-called shooting the bird in the head, killing the one with the greatest risk.

According to this logic, after the first 3% retracts, the later situation will gradually change, and gradually the whole trend will move closer to the average.

This method is very wise and more suitable for our national conditions. People may revolt directly by implementing a rigid one-size-fits-all approach, because the energy itself is larger than yours, and it will be impossible to offend too many suppliers.

Https://www.leiphone.com/news/202005/QbyuPnOG2G8AvJiT.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.