Abstract:
Vital signs monitoring is a key function in healthcare delivery to ensure immediate and
precise evaluation of a patient’s well-being. It is done by attaching monitor devices to patients
which collect, store, and display values on a screen. In many low-to-medium-income
countries (LMICs), hospitals still rely on manual observation and handwritten documentation
of vital signs, which is susceptible to human errors, data tampering, process inefficiency,
and limited opportunities for comprehensive data analysis. More advanced
hospitals utilize interface engines which transmit data to electronic medical records but
tend to be model-specific and are very costly. Optical character recognition (OCR) offers
a cost-effective and non-invasive alternative to digitizing manual transcription of vital
signs data in healthcare settings with low financial resources. An image preprocessing
pipeline is proposed to perform contour-based screen extraction of the patient monitor
captured by a camera, thus providing a well-defined region more suitable for subsequent
tasks of object detection and data extraction. The study offers a newly accrued dataset
of over 4000 images of Mindray Beneview T8 patient monitor with multi-parameter annotations.
Results showed that screen extraction prior to object detection significantly
improved the mean Average Precision (mAP) of the model from 68.55% to 93.65% at an
IoU threshold of 0.7.
