“Doctor, remove this from the bill”: how we looked for illegal services in the VHI

In many hospitals operating under VHI and simply providing paid services directly to the population, there is a kind of “sales plan” for each practitioner. The implementation of this plan is often achieved in dishonest ways at the expense of the insured through VHI. For instance:

1. Comprehensive services are divided into component medical manipulations so that the check is larger.
2. Excessive procedures and studies are prescribed in the treatment of diagnoses - especially if the hospital recently purchased new equipment.

Such abuse is a huge loss item for insurance companies in the voluntary health insurance (VHI) sector, which are already in fierce competition and are forced to expand their insurance program to attract customers. Therefore, on their part, there are expert doctors involved in the regular verification of bills. And in the case of violations - the conduct of the so-called "prevention" in medical institutions.



All this is a long and routine work, requiring the expert to be extremely concentrated. Indeed, the validity of the service is influenced by a number of factors related to both the history of the patient’s treatment and his insurance program, and the features of the price list in the hospital. Naturally, wherever you see the word "routine" you can apply automation. Which we did. Not without difficulties.

What are illegal services

Recently, I once again came to the clinic next door to the office to receive services for the VHI policy. An inflamed mole is almost always not scary, but observing a dermatologist will not hurt (otherwise you never know).



Everything turned out to be in order, but the doctor refused to make a routine examination of the remaining birthmarks, which are an irreplaceable source of my hypochondria. Referred to the fact that such services are not included in the policy and the next time the expert checks the insurance company, she will be fined. The doctor was right: an insurance case under VHI programs is usually an acute illness or an exacerbation of a chronic disease. All these cases do not include inspection for prevention, no matter how much I would like.



But far from always the doctors from medical institutions (or abbreviated hospitals) are well aware of the rules for the provision of VHI services and even less consciously observe them. As a result, an expert doctor checking accounts after treatment of the insured can see, for example, something like this:

At the same time, in the price list of hospitals there is a comprehensive service that includes both types of massage and is cheaper than the sum of individual services:

Thus, from each (!) Patient who has been given a similar combination of services instead of a complex, the insurance company suffers losses. Of course, if the expert doctor does not notice this on time and does not issue a replacement service in the account.



By the way, be sure to check what happens in your account if you yourself use paid medicine. Perhaps the complexes are recruited in approximately the same way.





And these are still simple cases. We understand further.



Suppose you have a baby and you brought him to a doctor with suspected viral conjunctivitis. As with the treatment of any other disease, the doctor should be guided by clinical practice and regulatory documentation. For example, the Standard of care is something like a table with a list of mandatory services for the diagnosis and treatment of a particular disease. Spoilers for the treatment of your alleged child can be read directly on the website of the Ministry of Health .



However, in addition to the usual set of services for payment, the insurance may include this:

When studying the Standard, it can be understood that no pneumonometry (as one of the methods for determining intraocular pressure is called) is done in this case - moreover, it is contraindicated in case of viral diseases of the eye. Our expert friend simply explains such illegal services: “If the clinic has forked out for a new unit, then it must somehow recapture its cost.”

Where did we start

We set ourselves the goal of automating the stage of checking bills from healthcare facilities for expert doctors in one insurance company (we are not allowed to disclose all names, passwords and appearances to the NDA). The plan was as follows: with the help of machine learning models, find and “highlight” in the register of accounts services that are similar to previously identified illegal ones.



To teach a car to distinguish between “bad” services and “good” services, we need data on how the insured are treated and what services the expert removed or replaced in the account after checking. Having opened real accounts, we suddenly faced with the fact that it is almost impossible to obtain clear statistics on the provision of services of each type. It turned out that the VHI does not have a specific standard for naming medical services, unlike the MHI, where the code for the Nomenclature of work and services in healthcare will strictly correspond to all services.



For example, one of the methods of professional hygiene in VHI can be written as follows:

• Bicarbonate cleaning.
• Removal of plaque by Air-flow.
• Sandblasting procedure.

These services are also inherently no different:

• RG lung.
• Chest R-graphy.
• Diagnostic radiography of OGK.

The situation is aggravated by the fact that bills are often filled out manually, and this only increases the variability of writing services due to typos, unnecessary marks and abbreviations. Here are some examples of such puzzles:







As a result, it is almost impossible to understand how they provided the same, in essence, service in different regions or even in different hospitals of your city without classification of names (and the use of text analytics technologies).



We armed ourselves with the SAS Visual Text Analytics tool and, for starters, conducted a thematic text analysis using machine learning and NLP technologies. This helped to automatically identify the so-called “topics” in a large and diverse array of services - small groups of texts that the system considered similar based on combinations of keywords. We reviewed all the results with experts and outlined the structure of the future classifier, combining the resulting “topics” in large classes according to medical logic (for example, in “Diagnostics”, “Doctor's appointments”, etc.). To correlate the names of the services from the training set with the resulting classes, for each of them we wrote the rules in our tool.



We repeated this sequence of steps for each subsequent class exactly until we came to the level of detail for specific services. So, for example, looks like a simplified rule for the professional hygiene method from the examples above:

(OR,

"**",



"**", "**",



(ORDDIST_1, "a", "f"),

(DIST_2, "air*", "*flow"),

(DIST_2, (OR, "*", "*", "*"), "**"),



(DIST_3, "*", "*", "*"),

(DIST_2, (OR, "*", "*"), "**")

)

Why is all this necessary? And why did we go in such a difficult way, instead of just stuffing everything into a neural network?



Naturally, we did not have any labeled sample for training. But the specificity of the subject area itself is an even greater obstacle: due to the lack of strict requirements for the composition of services in VHI, exceptions and complex border cases regularly arise, which medical experts from different insurance interpret in their own way. All these subtleties of the rule allow us to describe much more accurately.



For example, here the keywords are the same, but depending on their order, the essence of the service changes entirely:

How we trained models

In machine learning, the rule "you are what you eat" works perfectly. If you “feed” low-quality data into the model, the result will be the same. On our projects, we pay great attention to the stage of data preparation. This is especially true in the field of VHI: the insurer receives information about the treatment of patients from numerous hospitals, each of which has its own base. Accordingly, an error can sneak up both when filling out an account on the side of a medical institution, and when loading it into the database of the insurance company. As a result, when examining the data, we came across one and a half services in the account, and negative prices, and a sudden change in patient gender.



Having thought over the approach to fixing errors, we set about collecting indicators for modeling. We were interested in a variety of information describing the context of the provision of the service: the parameters of the service itself, its essence, the history of interaction with health facilities, the history of treatment of the insured, and much more. In total, we counted more than 20 thousand indicators. We used them to train the model, which automatically revealed the patterns characteristic of “bad” services.



Then they applied the model to new accounts and put a score from 0 to 1 opposite each medical service. It shows how similar the service is to the previously identified illegal: the closer the score to 1, the greater the similarity. In fact, behind this figure is a complex and not always obvious for a person decision-making process (as is most often the case in machine learning). If you try to interpret the logic of the machine, then in a simplified form it looks like this:

• The service belongs to the class “Comprehensive oral hygiene” (+ 0.01 points) .
• The previous service belongs to the class “Removal of dental plaque” (+ 0.25 points) .
• The cost of the current service is more than 3,000 rubles (+ 0.1 points) .
• The service was provided in the branch of Krasnodar (+ 0.05 points) .
• ...

If as a result the score turned out to be above a certain threshold value, then this is a signal for the expert that the service of this patient requires special control during verification. From this moment, a detailed investigation of the incident begins, at the end of which the expert doctor renders his final verdict - “execute” or “have mercy”. Not a patient, of course, but a service in the long run.

Project Summary

The methodology for standardizing names using SAS text analysis technologies helped to put together a single classifier of services (which was a huge problem before for automation), and the definition of services similar to previously identified illegal ones was implemented using machine learning models. Based on the results of internal testing of the system, we were convinced that we were able to identify the most common types of errors in the accounts, namely: the inclusion of additional services, deviation from the insurance program or standard treatment regimens.



But all this is only the first step to automation in VHI. If you still have questions or really want to know what else we can do, look for us here .