Evaluation of Convolutional Neural Network-Based Posture Identification Model of Older Adults: From Silhouette of Sagittal Photographs.

Abstract

Background/Objectives: Posture is a significant indicator of health status in older adults. This study aimed to develop an automatic posture assessment tool based on sagittal photographs by validating recognition models using convolutional neural networks. Methods: A total of 9140 images were collected with data augmentation, and each image was labeled as either Ideal or Non-Ideal posture by physical therapists. The hidden and output layers of the models remained unchanged, while the loss function and optimizer were varied to construct four different model configurations: mean squared error and Adam (MSE & Adam), mean squared error and stochastic gradient descent (MSE & SGD), binary cross-entropy and Adam (BCE & Adam), and binary cross-entropy and stochastic gradient descent (BCE & SGD). Results: All four models demonstrated an improved accuracy in both the training and validation phases. However, the two BCE models exhibited divergence in validation loss, suggesting overfitting. Conversely, the two MSE models showed stability during learning. Therefore, we focused on the MSE models and evaluated their reliability using sensitivity, specificity, and Prevalence-Adjusted Bias-Adjusted Kappa (PABAK) based on the model's output and correct label. Sensitivity and specificity were 85% and 84% for MSE & Adam and 67% and 77% for MSE & SGD, respectively. Moreover, PABAK values for agreement with the correct label were 0.69 and 0.43 for MSE & Adam and MSE & SGD, respectively. Conclusions: Our findings indicate that the MSE & Adam model, in particular, can serve as a useful tool for screening inspections.