Abstract
he automation of clinical documentation remains a significant challenge in modern healthcare systems. This study proposes a hybrid virtual physician assistant (VPA) architecture that integrates a discrete-time Markov chain for dialogue management, a transformer-based natural language processing (NLP) module fine-tuned on bilingual clinical corpora (Russian and Kazakh), and a semantic validation engine for mapping extracted entities to HL7 FHIR resources. The VPA was evaluated on a dataset comprising 1,000 de-identified electronic medical records (EMRs) and 1,000 voice dictations. The system achieved a 23% reduction in physician documentation workload and a 43% increase in accuracy compared to baseline methods. These outcomes were validated on bilingual Russian–Kazakh corpora and confirmed via physician surveys. The estimated economic benefit is 146 billion KZT annually. Future work will involve large-scale trials and integration with advanced LLMs such as BioGPT and Med-PaLM.