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

Erjiang E, Tingyan Wang, Lan Yang, Mary Dempsey, Attracta Brennan, Ming Yu, Wing P. Chan, Bryan Whelan, Carmel Silke, Miriam O'Sullivan, Bridie Rooney, Aoife McPartland, Gráinne O'Malley, John J. Carey

Research output: Contribution to journalArticlepeer-review

24 Citations (Scopus)

Abstract

Background: Identification of those at high risk before a fracture occurs is an essential part of osteoporosis management. This topic remains a significant challenge for researchers in the field, and clinicians worldwide. Although many algorithms have been developed to either identify those with a diagnosis of osteoporosis or predict their risk of fracture, concern remains regarding their accuracy and application. Scientific advances including machine learning methods are rapidly gaining appreciation as alternative techniques to develop or enhance risk assessment and current practice. Recent evidence suggests that these methods could play an important role in the assessment of osteoporosis and fracture risk. Methods: Data used for this study included Dual-energy X-ray Absorptiometry (DXA) bone mineral density and T-scores, and multiple clinical variables drawn from a convenience cohort of adult patients scanned on one of 4 DXA machines across three hospitals in the West of Ireland between January 2000 and November 2018 (the DXA-Heath Informatics Prediction Cohort). The dataset was cleaned, validated and anonymized, and then split into an exploratory group (80%) and a development group (20%) using the stratified sampling method. We first established the validity of a simple tool, the Osteoporosis Self-assessment Tool Index (OSTi) to identify those classified as osteoporotic by the modified International Society for Clinical Densitometry DXA criteria. We then compared these results to seven machine learning techniques (MLTs): CatBoost, eXtreme Gradient Boosting, Neural network, Bagged flexible discriminant analysis, Random forest, Logistic regression and Support vector machine to enhance the discrimination of those classified as osteoporotic or not. The performance of each prediction model was measured by calculating the area under the curve (AUC) with 95% confidence interval (CI), and was compared against the OSTi. Results: A cohort of 13,577 adults aged ≥40 yr at the age of their first scan was identified including 11,594 women and 1983 men. 2102 (18.13%) females and 356 (17.95%) males were identified with osteoporosis based on their lowest T-score. The OSTi performed well in our cohort in both men (AUC 0.723, 95% CI 0.659–0.788) and women (AUC 0.810, 95% CI 0.787-0.833). Four MLTs improved discrimination in both men and women, though the incremental benefit was small. eXtreme Gradient Boosting showed the most promising results: +4.5% (AUC 0.768, 95% CI 0.706–0.829) for men and +2.3% (AUC 0.833, 95% CI 0.812–0.853) for women. Similarly MLTs outperformed OSTi in sensitivity analyses—which excluded those subjects taking osteoporosis medications—though the absolute improvements differed. Conclusion: The OSTi retains an important role in identifying older men and women most likely to have osteoporosis by bone mineral density classification. MLTs could improve DXA detection of osteoporosis classification in older men and women. Further exploration of MLTs is warranted in other populations, and with additional data.

Original languageEnglish
Pages (from-to)527-537
Number of pages11
JournalJournal of Clinical Densitometry
Volume24
Issue number4
DOIs
Publication statusPublished - Oct 1 2021

Keywords

  • Classification
  • machine learning
  • Osteoporosis
  • osteoporosis self-assessment tool

ASJC Scopus subject areas

  • Endocrinology, Diabetes and Metabolism
  • Orthopedics and Sports Medicine
  • Radiology Nuclear Medicine and imaging

Fingerprint

Dive into the research topics of 'Machine Learning Can Improve Clinical Detection of Low BMD: The DXA-HIP Study'. Together they form a unique fingerprint.

Cite this