Significance of medication-related features in the FREM machine learning model for fracture prediction

<br/>Introduction: The FREM-ML (Fracture Risk Evaluation Model) is a gradient-boosting model to estimate the 1-year risk of major osteoporotic fractures (MOF) in Danish individuals over 45 years of age, designed for primary care. Age, sex, and previous fractures are strong predictors of fracture risk. Recently, we incorporated features related to prescription medication in addition to previous diagnoses. <br/><br/>Objectives: FREM-ML aimed to establish a novel decision support tool for the prevention of fractures. Here, we highlight the temporal dynamics of the strongest predictive medication features in FREM-ML, supporting risk-directed prevention. The rationale for investigating this is that some medications may only have short-term effects on fracture risk, while others could have a long-term impact. <br/><br/>Methods: Data were extracted from Danish Health Registers from 2007-2022. In short, 95 fourth-level ATC codes associated with osteoporosis and/or fall risk were assessed. Days since last redemption were calculated using the start of the index period and the date of last redemption. Relative risk (RR) estimates across dichotomous exposure were extracted from generalized logistic models adjusted for age and sex. Multivariable MOF probabilities across years since last redemption adjusted for age and sex were calculated and graphed for exposed individuals. <br/><br/>Results: The FREM-ML sample included 2,472,912 individuals. Medication marginally improved model performance. The five most relevant agents were sulfonamides (C03CA), selective serotonin re-update inhibitors (SSRIs, N06AB), thiazide/potassium combinations (C03AB), benzodiazepine-related drugs (N05CF), and angiotensin-II receptor blocker (ARB)/diuretic combinations (C09DA). Clear differences between temporal dynamics for MOF risk and dichotomous risk estimates were observed (Figure 1). Thiazides reduced the risk of fractures following temporal trends (C03AB, C09DA). <br/> <br/>Figure 1. Multivariable Temporal Risk Dynamics and Dichotomous Risk Estimates across Exposure<br/><br/>Conclusion: The FREM-ML model identifies medication exposures associated with fracture risk in a dynamic, time-dependent, and personalized manner. Results shows the relevance of medications in fractures risk stratification, particularly the five pharmacological agents examined. <br/>