Can Vision-Language Models See the Vital Signs? Benchmarking and Fine-Tuning for Intraoperative Monitor Reading

The article addresses the automation of reading vital signs from intraoperative monitors using vision-language models (VLMs). The research team created a benchmark consisting of 200 photographs of intraoperative monitors annotated for eight vital parameters: heart rate, SpO2, ETCO2, respiratory rate, systolic/diastolic/mean blood pressure, and temperature. They tested an OCR pipeline, nine VLM models, and a compact Qwen3.5-9B model adapted using LoRA fine-tuning. Results showed that the best commercial models achieved 0.98-0.997 accuracy without training, while the OCR pipeline and smaller models scored approximately 0.20 lower. After fine-tuning on 80-120 images, Qwen3.5-9B accuracy increased from 0.953 to 0.994 and critical error rate decreased fivefold. The conclusion suggests that monitor reading is a solved problem for VLM models larger than 9 billion parameters.