Voice Assistant for Outpatient Neurology Visits

Rationale and evidence gap. Documentation burden is widely recognized as a leading driver of professional burnout among physicians and reduces the time available for direct patient interaction. International evidence on ambient AI scribes - systems that combine automatic speech recognition (ASR) and large language models (LLM) - shows consistent reductions in documentation time and stable or improved patient satisfaction. However, evidence specific to Russian-language outpatient neurology practice integrated with a regional electronic health record is lacking. Neurologists are among the medical specialties with the highest reported burnout rates internationally; the specifics of Russian-language medical terminology, the structured neurological examination, and integration with a regional medical information system preclude direct extrapolation of international findings.

Technical implementation of the Service. The Service implements a distributed pipeline: streaming audio capture by an EMIAS client module; REST-API transmission of audio chunks to the speech recognition service with two-speaker diarization (physician/patient) and contextual re-clarification of streaming transcripts; asynchronous routing of transcription results through Apache Kafka topics to a pre-fill subsystem; large-language-model-based mapping of the full transcript onto a JSON schema of the structured outpatient neurology examination protocol (document class code 23951) using clinical reference dictionaries (including ICD-10 via the MKB10_ACTIVE_SHORT_INFO_V2 reference table); and return of the structured pre-filled protocol to the physician's user interface for review, editing, and signing. Audio is recorded via an active HD-capsule microphone placed at the physician's workstation (OGG Vorbis, 16 kHz, ≥ 64 kbit/s, mono). A priori quality targets fixed in the technical documentation before the pilot were: Word Error Rate of streaming transcription ≤ 10 %; sensitivity, specificity, accuracy, and Jaccard index for binary protocol fields ≥ 0.9; mean expert rating on a Visual Analogue Scale ≥ 8 out of 10; share of consultations with protocol generation time ≤ 35 seconds ≥ 90 %; physician Net Promoter Score ≥ 0 %.

Methodological design choices. A non-randomized single-group before-after design was selected as appropriate for a first-line feasibility evaluation in real clinical practice: the same physicians and clinics provide their own baseline reference in Stage 1, eliminating between-physician confounding while preserving ecological validity. A two-week post-deployment adaptation period was prespecified to mitigate purely learning-curve effects on Stage 2 metrics. Timing of consultation stages was performed by trained experts of the Scientific and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Healthcare Department on the basis of three synchronized data sources for each consultation: (i) automatic EMIAS event logs from protocol creation to signing; (ii) screen recording of the physician's workstation (VocoScreen NG, version 4.01.1); and (iii) webcam recording of the physician's face and gaze direction. The tri-source synchronization reduces systematic bias common to single-source or self-reported timing studies.

Quality evaluation framework. A stratified random sample of 30 paired audio recordings and Service-generated protocols (six per physician) was used to evaluate transcription and pre-fill quality. Reference (gold-standard) transcripts were prepared manually by an expert with more than five years of experience. Free-text fields of the protocol were rated by four independent neurology experts on a five-domain Likert questionnaire (Relevance, Accuracy, Completeness, Conciseness, Linguistic correctness) and on a 10-point Visual Analogue Scale; binary fields were rated dichotomously to construct a confusion matrix and derive sensitivity, specificity, accuracy, the Jaccard index, and the false-positive (hallucination) rate. Inter-rater agreement among the four experts was quantified by Gwet's AC2, preferred over Cohen's κ for categorical classifications with asymmetric prevalence. Word Error Rate was computed in Python with the jiwer library against the reference transcripts.

Analytical approach. Quantitative between-stage comparisons used Student's t-test or the Mann-Whitney U test depending on distribution, with Pearson's χ² for categorical variables and Cohen's d as the effect-size measure (α = 0.05, two-sided). Qualitative analysis of semi-structured in-depth interviews with each participating neurologist followed the grounded theory approach of Strauss and Corbin, with open, axial, and selective coding.