Identifying proximal humerus fractures: an algorithmic approach using registers and radiological visit data.

In this study, we show that combining register and radiological visit data enables more accurate automated identification of proximal humerus fractures compared to traditional register analysis. In a cohort of 11,863 post-menopausal women, our proposed approach improved the coverage of identified fractures from 74 to 81%.

Purpose: The aim of this study was to investigate how reliably proximal humerus fractures can be identified from different administrative datasets without manual review.

Method: Using the national medical registers, namely the Care Register for Health Care and the Register for Primary Health Care Visits, as well as the regional radiological image archive PACS, we developed algorithms for automated identification of proximal humerus fractures. In addition to these sources, we used data from patient records as well as from the self-reports gathered by the Kuopio Osteoporosis Risk Factor and Prevention Study (OSTPRE) to establish a gold standard of fractures for validating the algorithms. This gold standard included proximal humerus fractures for a cohort of 11,863 post-menopausal women living in the Kuopio region between 2004 and 2022.

Results: We report the national registers' yearly accuracy in identifying proximal humerus fractures. During the studied 19-year period, the registers' coverage initially improved but then settled at 75%. We show that the image archive provides almost complete coverage of radiographs for the fracture cases, but excluding false positives poses a challenge. In addition, we propose a simple approach that combines register and radiography visit data to improve the accuracy of automated fracture identification. Our algorithm improves the coverage from 74 to 81% and reduces the false discovery rate from 8 to 7% compared to the traditional register analysis.

Conclusion: The proposed approach enables a more reliable way of identifying proximal humerus fractures from administrative data. This study contributes to the objective of automatically tracking all types of fragility fractures in large datasets.