Development of deep-learning tool to predict appropriate height and weight from chest radiographs in bedridden patients

Abstract

Summary: Background & Aims: Malnutrition in the elderly, frequently and significantly affects both physical functioning and cognition, as well as incurs direct and indirect costs to society. Guidelines recommend rapid nutritional intervention and initiation of nutritional therapy within 24–48 hours of admission. Height and weight information is essential for proper nutritional assessment; however, it is difficult to obtain individual the height and weight of bedridden elderly patients directly. This study aimed to illustrate the potential of a convolutional neural network model to assess the height and weight based on chest radiographs. Methods: We retrospectively evaluated radiographs obtained over 15 years of follow-up. Overall, 6,453 radiographs from male patients, and 7,879 from female patients were included in the analysis. A convolutional neural network was used to predict the height and weight of the patients (Juzen NST). A ResNet152 classifier was trained using Fastai (V1.0) running on PyTorch to predict the height and weight. Training was performed for four epochs using validation without augmentation. Results: The correlation coefficients between the predicted and measured values using the height prediction model for males and females were R=0.855 and R=0.81, respectively. The correlation coefficients between the values predicted by the weight prediction model and measured values were R=0.793 and R=0.86, respectively. Conclusion: Our chest radiographic prediction model has a high correlation with actual height and weight and can be combined with information regarding clinical nutrition factors for rapid assessment of risk for malnutrition. By training the prediction model using chest radiographs from each hospital, it can be optimized for the most common ethnic groups in the area. Importantly, our model is an illustration of the potential of automated imaging artificial intelligence for proper nutrition prediction models in older adults.

Keywords

• Height
• Weight
• Body mass index
• Bedridden patients
• Deep learning
• Chest radiograph