Machine Learning Can Improve Clinical Detection of Low BMD: The DXA-HIP Study

Area of Science:

• Osteoporosis management and fracture risk assessment.
• Application of advanced computational methods in healthcare.

Background:

• Accurate identification of individuals at high risk for fracture is crucial for effective osteoporosis management.
• Existing risk assessment algorithms face challenges regarding accuracy and clinical applicability.
• Machine learning techniques (MLTs) are emerging as promising tools to enhance osteoporosis and fracture risk prediction.

Purpose of the Study:

• To evaluate the effectiveness of the Osteoporosis Self-assessment Tool Index (OSTi) in identifying osteoporosis.
• To compare the diagnostic performance of seven MLTs against OSTi for osteoporosis detection.
• To assess the potential of MLTs to improve the discrimination of osteoporosis in a real-world cohort.

Main Methods:

• Utilized a cohort of 13,577 adult patients with Dual-energy X-ray Absorptiometry (DXA) data and clinical variables.
• Validated the OSTi against modified International Society for Clinical Densitometry DXA criteria.
• Compared the performance of OSTi with seven MLTs (including eXtreme Gradient Boosting and Random Forest) using Area Under the Curve (AUC) with 95% Confidence Interval (CI).

Main Results:

• The OSTi demonstrated good performance in identifying osteoporosis in both men (AUC 0.723) and women (AUC 0.810).
• Four MLTs showed improved discrimination compared to OSTi, with eXtreme Gradient Boosting yielding the most significant improvements (+4.5% for men, +2.3% for women).
• MLTs also outperformed OSTi in sensitivity analyses excluding patients on osteoporosis medication.

Conclusions:

• The OSTi remains a valuable tool for initial osteoporosis screening based on bone mineral density.
• MLTs offer a potential enhancement for DXA-based osteoporosis classification in older adults.
• Further research on MLTs is recommended across diverse populations and with expanded datasets.