Performance of a Large Language Model for Predicting Clinical and Radiographic Outcomes in Knee Osteoarthritis: Single-Center Prospective Observational Study

Background Large language models (LLMs) can extract information from unstructured clinical text, but it remains unclear whether their ability to assess current status translates into reliable predictions of longitudinal outcomes. Objective This study aimed to evaluate DeepSeek’s ability to assess baseline clinical and radiographic status in knee osteoarthritis (KOA), its performance in predicting 6-month outcomes, and its interrun reproducibility. Methods This study used a single-center, prospective observational cohort design to conduct a blinded, retrospective evaluation of the LLM. A total of 128 patients with primary KOA were enrolled, of whom 111 (86.7%) completed the 6-month follow-up and were included in the primary analysis. DeepSeek-V4-Flash and a rheumatologist independently assessed KOA diagnoses, baseline Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) scores, and Kellgren-Lawrence (KL) grading based on the same standardized interview records and imaging reports and predicted 6-month outcomes. The LLM was run independently 3 times. The actual outcomes determined by another senior physician served as the reference, and the results were compared with the physician’s predictions and the persistence model. Results The LLM showed high baseline diagnostic agreement with the physician (105/111, 94.6%) and good agreement on the WOMAC and KL, but 6-month predictions (WOMAC mean absolute error=5.92-6.58; KL exact agreement=74/111, 66.7%-76/111, 68.5%) were worse than those of the physician comparator (mean absolute error=3.86; 96/111, 86.5%) and not superior to those of the persistence model (mean absolute error=5.43). Despite high interrun intraclass correlation coefficients (0.982 and 0.965), exact agreement between runs was poor (35/111, 31.5% and 7/111, 6.3%). Conclusions The evaluated DeepSeek configuration may support structured baseline KOA assessment. However, its 6-month predictive accuracy and patient-level stability remain limited. Prognostic applications require evidence of added value over a persistence model.