Acute Respiratory Distress Identification via Multi-Modality Using Deep Learning

Medical instruments are essential in pediatric intensive care units (PICUs) for measuring respiratory parameters to prevent health complications. However, the assessment of acute respiratory distress (ARD) is still conducted through intermittent visual examination. This process is subjective, labor-intensive, and prone to human error, making it unsuitable for continuous monitoring and early detection of deterioration. Previous studies have proposed solutions to address these challenges, but their techniques rely on color information, the performance of which can be influenced by variations in skin tone and lighting conditions. We propose leveraging multi-modality data to address these limitations. Our method integrates color and depth data using deep convolutional neural networks with a late feature fusion scheme. We train and evaluate our model on a dataset of 153 patients with respiratory illnesses, 86 of whom have ARD of varying severity levels. Experimental results demonstrate that multi-modality data combined with simple late fusion techniques are more effective with limited data, offering higher confidence scores compared to using color information alone. Our approach achieves an accuracy of 85.2%, a precision of 86.7%, a recall of 85.2%, and an (Formula presented.) score of 85.8%. These findings suggest that multi-modality data provide a promising solution for improving ARD detection accuracy and confidence in clinical settings.