Abstract
Musculoskeletal disorders and injuries, including bone fractures, degenerative joint disease, ligament and meniscus injuries, and postoperative states after arthroplasty or osteosynthesis, often require prolonged treatment and staged rehabilitation. In trauma and orthopedic practice, clinical decision making depends heavily on narrative documentation, yet mixed Kazakh and Russian writing, highly variable terminology, and extensive free text complicate consistent clinical coding, outcome analytics, and the preparation of standardized discharge summaries and rehabilitation recommendations. The aim of this studywas to determine whether cross-lingual domain adaptation of pre-trained medical transformer models using Kazakhstan-specific trauma and orthopedic clinical narratives improves bilingual medical understanding. Methods. We conducted a retrospective study using records of five hundred adult patients treated in the Orthopedic Surgery Department ofAlmaty City Clinical Hospital No.4, Kazakhstan. Multi-page electronic case histories stored in portable document format were de-identified, converted into text, and then transformed into bilingual instruction-style dialogue examples designed to reflect real clinical documentation patterns and musculoskeletal disease terminology. Two pre-trained medical transformer backbones were adapted using a parameter-efficient low-rank adaptation procedure: a compact healthcare-optimized model and a larger biomedical model. Performance was evaluated on the medicine subset “Professional and University, Russian language” of the Kazakh Massive Multitask Language Understanding benchmark, using accuracy as the primary outcome and the macro-averaged harmonic mean of precision and recall, balanced accuracy, and the Matthews correlation coefficient as secondary outcomes. Ninety-five percent confidence intervals for accuracy and the macro-averaged harmonic mean of precision and recall were estimated using one thousand bootstrap resamples. Results.After domain adaptation, the compact medical model achieved an accuracy of 33.00% (95%confidence interval -27.95 to 38.72), compared with 20.88%(95%confidence interval -16.50 to 25.59) before adaptation; the macro-averaged harmonic mean of precision and recall increased from 18.64% to 26.92%, balanced accuracy increased from 21.01%to 33.34%, and the Matthews correlation coefficient increased from 0.105 to 0.170. The larger biomedical model changed minimally, with accuracy increasing from 28.96%to 29.63%. A general-purpose multilingual baseline model achieved 30.64%accuracy without clinical domain adaptation. Conclusions.These findings show that cross-lingual domain adaptation on a limited Kazakhstan-specific trauma and orthopedic corpus yields measurable gains, particularly for compact instruction-following medical models, and may support future tools for standardizing orthopedic documentation and accelerating rehabilitation planning. However, benchmark performance remains below levels required for high-responsibility clinical workflows, and further progress will require larger multi-center datasets, validation on practical documentation tasks such as structured extraction and discharge summary drafting, and dedicated evaluation of safety, privacy, and clinical risk.Keywords:musculoskeletal diseases,fractures, bone,osteoarthritis,rehabilitation,medical records systems, computerized,natural language processing,machine learning,multilingualism.