8 months ago
CHALLENGE
Imagine you are a medical insurance representative for a hospital, and that you have to make sure that the prescriptions that doctors issue are compliant with the contract between the hospital and the insurance company. Every time that a new prescription is created, you have to go through the contract, find all the details relevant to the prescription and ensure that the details are correct. After reviewing a few dozen prescriptions, you make a list of rules that you can follow to make the process of checking quicker and easier. Now, when a new prescription comes in, you can skip the stage of reading the contract, and just go through the rules you extracted previously from the contract. This is the stage at which it becomes easy to automate the process. Rules can be translated into a computer program, and then you can focus on validating the program results.
However, what if the details change in the contract? Some of the rules may not apply any more or you may need to add some new rules to the list of rules. What if there are many contracts? In that case, you have to prepare a separate list of rules for each contract and then match contracts to prescriptions. Last, but not least, what if you don’t know in advance the contract that you will have to validate the prescription against? This is the problem that one of our customers came to us with, and we concluded that it was the perfect case for using an LLM.
SOLUTION
We created an application with an engine based on a model from the GPT family, where the application's main task was to process an unknown contract, extract the rules specified in that document and apply the rules to a prescription to validate it against the contract. Below are two simple examples of the rule extracted automatically using an LLM and then applied to a prescription. The system concluded that the prescription was not valid in the first example, but was valid in the second example, as follows:
Of course, a regex could easily extract this piece of information, but only if you know in advance that that particular regex is required. By using an LLM, not only did we not have to create the regex ourselves, but we also did not need to know that this is something that will be extracted from a document. In addition, we could also instruct an LLM to use a regex when necessary.
A sceptic might ask about the case where the customer’s data is sensitive, so you can’t just send the data to OpenAI. That is not a problem because Digica deployed Mistral, an open source LLM, on private hardware, which made it possible both to provide prompts to the LLM in a secure way and to finetune the LLM, based on the customer’s data.
BENEFITS
Thanks to our solution, our customer can make the verification process run faster and can use the resources allocated to prescription validation in more creative ways. Also, if a machine makes an error, it is easier for a human being to correct the error than for a human to go through the whole document at one time, which is even more error-prone than letting the machine do the job. Used in this way, LLMs can easily bring tangible benefits to the insurance industry in general. Processing insurance policies is usually a repeatable process, but the details are too difficult to be able to create one general solution using only regexes or classical machine learning. On the other hand, by deploying LLMs on private clouds, it is easy to quickly create personalised and secure solutions.