Doctors under supervision: how AI learns to notice every little thing in the medical record

A medical organization today is not only the provision of medical care, but also large volumes of documentation. Much depends on how neatly the medical history is drawn up: the timing and quality of the patient’s treatment, fines and license revocation, and the reputation of the clinic. The insurance company sees an inaccurate diagnosis in the card – and calmly cuts payment; Roszdravnadzor (Federal Service on Surveillance in Healthcare and Social Development) inspector finds the missing signature – and writes out a fine. This is why quality control of medical records has become a daily routine for every major clinic. According to the law, they must carry out internal quality control, one of its parts is manual verification of documents in post-control mode.

Problems of clinics when checking

There is no single and standardized approach to conducting examinations: someone checks documents in special submodules of the MIS, someone does it in excel tables. The main disadvantages of such tools that can be fixed with the introduction of artificial intelligence (AI) are the inability to prioritize medical records “from worst to best” and the need to start checking “from scratch.” If AI is introduced as the first level of verification, then the clinic will have a tool for smart selection of documents for manual verification, and the examination itself will be carried out on top of the assessment made by the AI.

How the checking was carried out before

Previously, the department of clinical and expert work checked selective 5-10% of cards: they simply did not have time anymore. But modern large language models have shown that they can “read” thousands of documents per hour and carefully compare them with the rules and interpretively find suboptimal medical actions, inconsistencies with clinical recommendations and other defects. Therefore, the AI takes over the initial viewing of all maps, and only those where the system sees the risk of error fall to the human checking.

AI based verification algorithm

The clinic downloads impersonal cards.
The neural network “reads” them and compares them with the current recommendations.
For each document, an AI control card is formed: a set of defects with comments on why they were selected. The risk color is assigned to the document.
The expert opens only “red” records, confirms or rejects the conclusions.
The next morning, the attending physician receives a short letter with errors in order to correct immediately the document and practice.

Errors in medical records can be divided into two types:

Errors in the design of the document: a doctor did not fill out a section, did not detail sufficiently the information from the patient, used incorrect wording or abbreviations. Perhaps the reception went well, and the patient was satisfied, but when analyzing the document there is something to complain about.
Medical errors: doctor’s actions that could lead to delays in diagnosis, incorrect diagnosis, or complications. Such errors include non-compliance with clinical guidelines.

In total, the digital solution of our company, for example, can identify 112 defects, but for each clinic such services must be adapted: somewhere you need the option to disable defects/create new ones, somewhere you need correction of the logic for determining defects, amending the rules for appointments, etc. The neural network groups all errors into the criteria that are most common among five clinics: complaints and anamnesis, examination, diagnosis, diagnostics, prescriptions and recommendations.

It is important that such AI-based systems allow an expert doctor to change selected defects and adjust assessments according to criteria. In the future, the system learns from these adjustments.

What exactly can AI find?

The complete list includes more than 100 (in our case – 112) types of defects – from large to the smallest. Here are some examples:

Medical history: no history of bad habits or allergies, incomplete epidemiological history collected.
Inspection: inconsistent entries, redundant abbreviations, missing key parameter (e.g. saturation).
Diagnosis: MKB-10 code misspelled or unsupported by complaints and tests.
Laboratory and instrumental studies: the required analysis is not assigned, and the excess one is assigned without justification.
Prescriptions: important drug missed, dosage exceeded, duplicate drugs, contraindicated drug was prescribed to patient.
Recommendations: the referral to a specialized specialist is forgotten or advice on the regime and diet is not given.
Medicinal uses: dosage, course, type of administration, etc. for medicinal products are not marked or incorrectly indicated. The correctness of therapy is also evaluated in accordance with clinical recommendations.

There is also an example of an error that the system can catch, and the expert doctor can miss.

For example, the doctor prescribed to a patient with perichondritis of the right auricle, the drug “polyoxidonium,” which contains cocoa in the composition. At the same time, the patient had a history of allergy to cocoa. According to the rules, such appointments are best avoided, as they can cause an allergic reaction. It is impossible to know the composition and contraindications for all drugs, but the use of AI and automated verification systems will allow such cases to be detected.

What does this give in practice?

In pilot projects, automatic verification using AI reduces the expert’s time with each card by three times, and all checks carried out are immediately collected in analytical reports. The main thing is that the system does not replace the doctor, but removes the routine: the specialist deals with complex cases, and the machine – with a total check of all documents. The accuracy of detecting defects in general therapeutic specialties was about 85%.

The experience of introducing such digital solutions into clinics has shown a great demand for adapting the logic of evaluating documents to the request of medical organizations, therefore, as part of the launch, the systems will be adapted to the wishes of the customer. For example, AI may even suggest how to fix a bug.

Instead of the dry mark “incorrect dosage,” the doctor sees friendly advice: “For this diagnosis, the recommendations indicate a dose of 75 mg × 2 daily; you have 150 mg.” The doctor can accept the hint, change it or explain why the deviation is justified.

Result

Quality control of medical documents ceases to be a “post-flight analysis.” The neural network checks each card at once, experts take on really controversial cases only, and the head physician gets peace of mind and fewer fines.