Abstract
U
Diagnosis of fibroids becomes more complicated when using medical records in a retrospective fashion. A diagnostic code of fibroid may be assigned based on physical examination but later disproven on imaging studies. In addition, imaging studies vary in sensitivity by modality and are not flawless. An ultrasound may misdiagnose an adenomyoma as a fibroid, a finding discovered only on surgical pathology. Thus, an investigator needs to decide what she considers to be a definitive diagnosis. Is one diagnostic code enough? Must imaging be available? Is an imaging report accurate? Is surgical pathology required?
The study by Feingold-Link et al. 5 published in the December issue offers a new approach to solving this problem. They present an algorithm to identify cases with fibroids and controls without fibroids using diagnostic codes, clinical notes, and imaging findings. This approach combines the advantage of access to much larger cohorts with the cost-saving benefit of utilizing readily available imaging. Case definitions for this algorithm included at least one imaging study with a diagnostic code indicative of fibroids, while control requirements were at least two imaging studies without any diagnostic codes for fibroids. In addition, controls were excluded if they had a fibroid-related procedure code or clinical notes indicating a history of fibroids. In comparison with a clinical review of records, their positive and negative predictive values were high.
The greatest benefit of this algorithm is its flexibility. 5 In genetic studies of fibroid development, it may be necessary to have complete case ascertainment, which would occur at menopause, when development of fibroids ceases. Feingold-Link et al. demonstrated high reliability of the algorithm in women over age 55. However, in studying reproductive outcomes, it would be necessary to have a younger cohort where future development of fibroids is less important than present status. For more rigid definitions, one could limit the inclusion criteria to using two diagnostic codes for fibroids or limit imaging in controls to higher sensitivity modalities, which would likely lead to even higher reliability. Stricter designations are possible and more relevant in the era of big data. Exploration of applying a similar algorithm to large insurance claim databases should be considered.
No method is without limitation though. One limitation to their current study is that the cases were significantly older than controls, where controls may not yet have reached an age where fibroids develop; thus, comparing outcomes that are influenced by age between these two groups would be impossible. However, this report demonstrates that the algorithm may work across a wide age range. Another major constraint to studying fibroids is the requirement that participants have a uterus and thus are at risk for fibroids. Depending on study inclusion criteria or age at enrollment in a healthcare system, this may exclude the most severe cases who undergo hysterectomy at a young age and controls who underwent hysterectomy for other reasons but would not ever have developed fibroids. One must balance the ability to avoid misclassification with the reduction in selection bias. Although Feingold-Link et al. excluded controls that lacked a uterus for the testing of reliability, their use of natural language processing to review notes for a history of fibroids has potential to alleviate this problem in the future. 5
The challenges of studying fibroids are many, but the value in prevention and reduction in risk is high. Researchers continue to make great strides in refining our ability to study these tumors effectively. This welcome addition by Feingold-Link et al. 5 to the methodology of utilizing population-based cohorts continues on this path.