Predictive Modelling of Bone Mineral Density: An ANN and Regression-based Approach

Abstract

In recent years, the development of predictive models using Multi-Variable Regression (MVR) and Artificial Neural Networks (ANN) has become a focal point in health research, particularly in predicting Bone Mineral Density (BMD) for the early detection of osteoporosis. This study compares the performance of MVR and ANN models using a clinical dataset comprising patient attributes such as age, weight, height, and BMD. The primary objective is to predict BMD values of the femur bone and evaluate the potential risks of osteoporosis. ANN demonstrated superior predictive accuracy with a correlation coefficient (R²) of 0.8823 compared to 0.6087 for MVR, highlighting its capability to capture data linearity and complex patterns effectively. The study used filtered and validated datasets, including results from BMD tests on two dry intact femurs, sourced from Kaggle. Performance metrics such as regression accuracy and Mean Square Error (MSE) were calculated, showing that ANN with a hidden layer of 12 neurons provided the best results. The findings indicate that ANN not only outperforms MVR in predictive accuracy but also avoids the need for experiments on real human femurs, providing a non-invasive, data-driven alternative for medical diagnostics. A secondary goal was to develop a practical model for clinical use in predicting bone density. The study also explores the integration of ANN outputs with Genetic Algorithms (GA) to optimize the prediction process. This hybrid strategy reduces the number of simulations and computation time, offering a robust framework for global optimization. The combination of ANN and GA demonstrates the potential to enhance diagnostic precision and streamline decision-making processes in orthopedic and medical technology. In conclusion, this study emphasizes the applicability of ANN for accurate BMD predictions, paving the way for advanced diagnostic tools in healthcare. Future research could focus on expanding datasets and exploring hybrid optimization techniques to further improve prediction accuracy and clinical utility.