DEVELOPMENT OF A MATHEMATICAL MODEL OF A VIRTUAL PHYSICIAN ASSISTANT TO AUTOMATE THE PROCESS OF FILLING OUT MEDICAL RECORDS USING NATURAL LANGUAGE PROCESSING AND ARTIFICIAL INTELLIGENCE TECHNOLOGIES

Abstract

This paper develops and empirically validates a mathematical model of a virtual physician assistant (VPA) that automates the completion of electronic medical records (EMRs) from free-text or speech using natural language processing (NLP) and artificial intelligence (AI). The model couples a discrete-time Markov chain for dialogue management with a transformer-based pipeline for clinical named-entity recognition, ontology-driven normalization (ICD-10, ATC, LOINC), and HL7 FHIR mapping. Unlike prior assistants that focus on transcription or narrow templating, the proposed VPA is multilingual (Russian–Kazakh), context-aware, and optimized for integration with local EMR systems via REST APIs. We detail data preparation for two de-identified corpora (1,000 notes), numerical normalization, and bilingual lexicon construction; and we provide a formal parameterization of transition probabilities and sojourn times to model doctor–assistant interactions.

In a controlled evaluation on 1,000 audio–text pairs, the system achieved a word-error rate of 8.5% for ASR and entity-level precision/recall/F1 of 0.91/0.85/0.88. Fine-tuning with medical corpora and domain lexicons improved F1 to 0.86 over a DistilBERT baseline (0.80) and outperformed leading commercial dictation-centric tools in our test bed. Load testing demonstrated median API latency below 120 ms and sustained throughput up to 250 requests/s with <0.2% error rate, supporting both real-time documentation and batch ingestion.

The main contributions are: (i) a formally specified, multilingual dialogue–NLP model tailored to the clinical documentation workflow; (ii) an end-to-end EMR integration path with error-handling policies aligned to FHIR; and (iii) an empirical study demonstrating tangible gains over baselines and widely used tools. The results indicate that the proposed VPA materially reduces manual data entry while preserving terminological consistency and auditability, and is therefore suitable for scalable deployment in Kazakhstan’s digital health infrastructure.