In this study, we investigated different ways of identifying proximal humerus fractures from routinely collected administrative data. The methods using the registers and radiological image archive were evaluated against a gold standard of fractures from 2004 to 2022 derived from various sources: self-reports, radiological reports, national registers, and patient records. The analysis revealed limitations in the traditional register-based fracture identification, especially in the early years of the study period. In contrast, the image archive provided almost complete coverage but posed challenges in excluding false positives. We demonstrated that by combining register and radiography visits data, we can improve the accuracy of automated fracture identification, improving the coverage from 74 to 81% while reducing the false discovery rate from 8 to 7%.

The results complement our previous research on wrist fractures [4], contributing to the aim of identifying all types of fragility fractures for large datasets automatically. Since wrist and proximal humerus fractures are largely treated in outpatient care, their accurate identification highly depends on the completeness of system integrations and the consistency of reporting practices among healthcare providers. Compared to hip fractures treated in specialized health care, reliable identification of these fracture types requires more sophisticated algorithms that combine data from multiple sources.

This study’s proposed algorithms for proximal humerus fractures outperform the traditional register analysis but are somewhat less effective than our previous methods for wrist fractures. There are two main reasons for this. Firstly, the treatment of wrist fractures typically follows a more consistent and distinguishable pattern of radiographic examinations [21]. In humerus fractures, the number of radiography visits and the time between them vary more than with wrist fractures [12, 22]. Secondly, wrist radiographs are mostly taken for fracture diagnostics, whereas proximal humerus radiographs serve diverse diagnostic purposes related to, for example, arthrosis, shoulder dislocation, and fractures of other bones. Registered diagnosis codes for these other conditions were sparse making their exclusion difficult.

Nevertheless, the radiography visits data can be used to reduce certain types of false positives occurring in traditional register analysis. For instance, proximal humerus fractures can show post-fracture symptoms even years after the fracture. This often results in a registered visit with the same diagnosis code but without a new radiographic examination. Also, incorrectly reported proximal humerus diagnosis codes were found in cases of prosthesis control, fractures in the distal end or diaphysis of the humerus, and even in fractures of different bones. Algorithms requiring humerus radiography in addition to the registered diagnosis cleared out many of such false positives. This approach also enabled us to extend the analysis to the more generic fracture diagnosis codes S42 and L76 and even to the humerus shaft fracture code S43.4 when they were associated with two or more humerus radiography visits. A detailed analysis of false positives for each algorithm is provided in the supplementary material (Table SI3).

Since the radiography visits data proved most efficient when used in combination with the registered diagnosis codes, the coverage of the national registers remains crucial. The introduction of the Register for Primary Health Care Visits (Avohilmo) had a positive effect, as our annual analysis indicated. However, the register coverage after 2011 was still notably lower in humerus fractures (73.7%) than in wrist fractures (81.0%). The discrepancy could be explained by the difference in the diagnosis codes and the difficulty of accurately locating humerus fractures. Some healthcare providers still use the ICPC2 standard, and while a specific ICPC2 code (L72) exists for wrist fractures, proximal humerus fractures must be reported with the generic code (L76) indicating “other fracture.” This code is ineffective for traditional register analysis as it encompasses various unrelated fracture types. Misreported fractures were also observed in ICD-10 codes, with some proximal humerus fractures reported as humerus shaft fractures (S42.3) or distal humerus fractures (S42.4). These may have resulted from initially mislocating the fracture or errors in data entry.

We acknowledge certain limitations in this study. In the early years of the study period, some fractures may be missing from the gold standard as both the register and PACS integrations were incomplete, and the self-reports had limited coverage. Nevertheless, the incidence calculated from our gold standard aligns well with previous literature [16, 18], although comparison between studies is challenging due to differences in fracture registering, study design, and environmental risk factors. Furthermore, as our access was to the regional PACS system, this study is geographically constrained to North Savo and only to participants who resided in the area during the study period. Local differences in the organization of care may affect the results. However, future research could utilize the proposed algorithms to identify fractures from the National Patient Data Repository, which will gather electronic patient data across the country. Also, while some adaptation may be necessary when transferring our methods to registers and image archives in other countries, the approach of combining different administrative data sources and algorithmically adjusting the sensitivity and specificity according to research purposes could work in many data environments worldwide.

While our proposed identification algorithm reached a coverage of 81%, it is also noteworthy that the PACS system did contain radiographs for all the gold-standard fractures between 2011 and 2022. In 91.9% of these cases, the fracture finding could be determined from the written report without image review. Hence, the accuracy of fracture identification can be further refined through the manual effort of reading radiological reports, by utilizing natural language processing techniques or by detecting fractures directly from the images.