- Статьи
- Science and technology
- Medical records: new platform will save doctors from paper routine when working with AI
Medical records: new platform will save doctors from paper routine when working with AI
The new domestic platform will be able to automate one of the most time—consuming tasks in working with medical data - their collection, structuring and cleaning. In a matter of minutes, the system will transform dozens of disparate documents — from scans and handwritten extracts to tables with analysis results — into a single logically linked medical history, where all terms will be brought to international standards, and personal data will be securely hidden. With the introduction of technology, doctors will have access to a complete picture of the patient's condition without having to study archives for hours. The system is currently being tested in Russian hospitals and is being prepared for a pilot launch. However, experts note that before mass implementation, it is necessary to confirm its effectiveness, including evaluating the accuracy of document recognition, completeness of information extraction, and other key performance indicators.
A platform for processing medical documents
The ITMO Youth laboratory "Digital Technologies in Public Health" has created a multi-agent platform for processing medical documents, which has no complete analogues in Russia. The system automatically analyzes disparate data — texts, scans, tables — and transforms them into a single structured patient history. She extracts diagnoses and test results from them, harmonizes this information, detects, and also deletes personal data and connects individual facts with each other, showing how different patient records relate to each other.
In hospitals and polyclinics, patient information — analyses, tables with examination results, handwritten medical reports and extracts from various institutions — are rarely generated and stored in a single format. To use this data to train artificial intelligence, for example, to create tools that help doctors diagnose or predict the development of diseases, requires a long preparation. Developers often have to manually collect data for further work with AI. The situation is similar in clinics: the doctor needs to review dozens of documents from different sources before restoring the full picture of the disease. Automating the collection of information from various sources can simplify these routine processes.
— The main difficulty of working with medical data is their diversity and fragmentation. The difficulty is not to process one specific document, but that there are many of them and they are very different. When a developer receives a large amount of data from doctors, 60-70% of the time allocated for training the model is spent only on preparing the material. It is necessary to clear the documents of personal data and technical notes, bring all records to a single format and mark up key medical information — diagnoses, symptoms, appointments, test results. MAS-EHR is needed to speed up and simplify this stage," said Ivan Zolin, one of the developers of the platform, a student of the Deep Learning and Generative Artificial Intelligence program at the ITMO Institute of Applied Computer Sciences.
MAS-EHR platform (Multi-Agent System for Electronic Health Records — multi-agent system for processing electronic medical records) it consists of independent modules, and data processing in it is arranged in several steps. One module is responsible for optical recognition: it reads text from scans and tables and translates it into machine-readable form. The second one finds and hides the patient's personal data. The third involves language models — they extract key medical information from the text: diagnoses, symptoms, appointments, test results — and link them together in meaning. At the same step, the terms are brought to international standards like the classifiers of diseases (ICD-10) and SNOMED CT — unified reference books of diseases and clinical concepts. At the last stage, a special LLM-based verification module (validator agent) checks the resulting structure against the source document. If part of the data is lost or the model has guessed something, the system signals this and returns the fragment for re-processing.
— Large technology corporations are developing their own cloud-based medical document processing services, but in them the user receives a ready-made answer and does not know how it was received or whether the model was wrong. In MAS-EHR, it's different: the doctor or developer can monitor what happens to the data at each stage. First, it checks how the text was read, then how the anonymization went, then what exactly was extracted and how the entities were linked. At the final stage, the validator compares the source document with the result. At the same time, due to compact, advanced models, the system works faster and consumes fewer resources than heavy universal solutions," explained Ivan Zolin.
The laboratory provided access to anonymized medical data from practitioners from the Fomin Clinic and the N.N. Petrov National Research Medical Center of Oncology. Their feedback helped us understand how the platform should work in real clinical settings and which data is important to doctors in the first place. The instrument is ready for piloting, the university told Izvestia.
Structured, anonymized medical data
One of the main problems of digital healthcare is not a lack of data, but its fragmentation. The history of one patient can consist of dozens of documents: analyses, statements, doctors' reports, examination results, which are often stored in different formats and even in different medical organizations, Anton Averyanov, CEO of the ST IT Group of companies, TechNet NTI market expert, told Izvestia.
— As a result, the doctor spends time searching for information, and medical AI developers, according to various estimates, spend up to 60-70% of their time preparing and cleaning data, rather than training models. The main value of ITMO's development lies in the fact that it not only recognizes documents, but turns them into structured medical data, while depersonalizing information and verifying the correctness of the AI. This is especially important for clinics that are starting to implement artificial intelligence, since the quality of the model always directly depends on the quality of the source information," the expert said.
The problem of medical institutions is not always the high—quality management of electronic medical records (EHR), hence the analysis of such data using AI solutions is difficult. The development can be useful to improve the completeness and quality of incoming information and the accuracy of the output result, said Alexander Gusev, Director of Development at Webiomed, a market participant at NTI Helsnet.
The main task of such solutions is not to make a diagnosis, but to prepare data, said Ilya Naletov, head of artificial intelligence at the Third Opinion Platform.
— The practical value of the platform will depend not so much on the use of a multi—agent architecture, as on the accuracy of recognition, completeness of information extraction, protection of personal data and the ability of the system not to add facts that are not in the source documents, - said the specialist.
The results of the ITMO scientists' work are published in the IEEE Transactions on Power Electronics journal.
Переведено сервисом «Яндекс Переводчик»
