A Multi Layer Perceptron Neural Network for Predicting the Diagnosis of Osteoporosis in Women Using Physical Activity Factors

Background Osteoporosis (OP) is a bone disease caused by a decrease in bone mineral density (BMD). OP is common in women because BMD gradually decreases after age 35. OP due to decreased BMD is highly likely to cause fatal traumatic injuries such as hip fracture. Purpose The purpose of this study was developed and evaluated a multi-layer perceptron neural network model that predicts OP using physical characteristics and activity factors of adult women over the age of 35 whose BMD begins to decline. Study design Cross-sectional study. Methods Data from KNHANES were used to develop a multi-layer perceptron model for predicting OP. Data preprocessing included variable selection and sample balancing, and LASSO was used for feature selection. The model used 5 hidden layers, dropout and batch normalization and was evaluated using evaluation scores such as accuracy and recall score calculated from a confusion matrix. Results Models were trained and evaluated to predict OP using selected features including age, quality of life index, weight, grip strength and average working hours per week. The model achieved 76.8% accuracy, 74.5% precision, 80.5% recall, 77.4% F1 score, and 74.8% ROC AUC. Conclusions A multi layer perceptron neural network for predicting OP diagnosis using physical characteristics and activity factors in women aged 35 years or older showed relatively good performance. Since the selected variables can be easily measured through surveys, assessment tool, and digital hand dynamometer, this model will be useful for screening elderly women with OP or not in areas with poor medical facilities or difficult access.