Interpretable (

Marlene Rietz; Brond Jan C; Soren Moller; Jens Sondergaard; Bo Abrahamsen; Rubin Katrine Hass

doi:10.1530/endoabs.109.OP7.4

OP7.4

Prev Next

Section Contents Cite

Endocrine Abstracts (2025) 109 OP7.4 | DOI: 10.1530/endoabs.109.OP7.4

SFEBES2025 Poster Oral Presentations Bone and Calcium (4 abstracts)

Interpretable (‘explainable’) machine learning in osteoporosis case finding: using SHAP values in clinical decision-support for the FREM_ML fracture prediction algorithm

Marlene Rietz ^1,2 , Jan C Brønd ³ , Sören Möller ¹ , Jens Søndergaard ⁴ , Bo Abrahamsen ^1,5 & Katrine Hass Rubin ¹

Author affiliations

¹Research Unit OPEN, Department of Clinical Research, University of Southern Denmark, and Odense University Hospital, Odense, Denmark; ²Division of Clinical Physiology, Department of Laboratory Medicine, Karolinska Institutet, Huddinge, Sweden; ³Research Unit for Exercise Epidemiology, Department of Sports Science and Clinical Biomechanics, SDU, Odense, Denmark; ⁴Research Unit of General Practice, Department of Public Health, University of Southern Denmark, Odense, Denmark; ⁵Medicine, Holbæk Hospital, Holbæk, Denmark

Background: Machine Learning (ML) may improve case finding finding accuracy in comparison to standard regression modelling identifying subjects at high imminent (one-year) risk of major osteoporotic fractures (MOF). However, healthcare providers need to justify treatment decisions through observable risk factors.

Methods: FREM_MLwas trained and validated using complete registry data for the Danish population aged ≥45 years without prior osteoporosis diagnoses or treatment (n = 2,472,912). Predictors of fractures (15-year lookback) included diagnoses, filled prescriptions, days since last redemption of fall- and osteoporosis-risk medication, and polypharmacy and multi-morbidity. Risk outputs were backed by SHapley Additive exPlanations (SHAP) values to facilitate clinical interpretation.

Findings: FREM_ML displayed an overall area under the curve of 0.77. To exemplify the clinical implications and interpretation, we show two hypothetical cases: In a 67-year old woman, a relative fracture risk of 1.67 was estimated. Using SHAP values, we identified the ten strongest personalized risk factors, including several previous fractures (combined SHAP: +1.41), multi-morbidity (SHAP: +0.29) and alcohol (SHAP: +0.19). While electronic health records (eHR) may be inspected further based on highlighted factors, dual x-ray absorptiometry (DXA) assessment is likely warranted. In contrast, Case 2, a 91-year old woman, was assigned a relative risk of 0.86. The majority of the absolute MOF probability was explained by age (SHAP: +1.36) and sex (SHAP: +0.28), while other predictors lowered the risk estimate (e.g., recent redemption of cardiac glycosides, no history of any MOF; cox-arthritis). Here, clinicians must decide whether an essentially age-motivated DXA scan is warranted. To translate SHAP values into interpretable insights, we suggest visual inspection of personalized scatterplots integrated into eHRs.

Conclusion: We propose to evaluate the FREM_ML-based relative risk estimates to promote osteoporosis case-finding in general practice. Importantly this alerts practitioners not only to the level of risk but also to the personalized factors driving the risk assessment.