Genomic Classifiers in Diagnosing Interstitial Lung Disease: Finding the Right Place at the Right Time.

Idiopathic pulmonary fibrosis (IPF) is a fibrotic interstitial lung disease (fILD) characterized by dyspnea, progressive lung scarring, and high mortality (1). Despite substantial advances in our understanding of the risk factors and pathophysiologic mechanisms underlying IPF development and progression, clinicians still struggle to make a confident diagnosis of this condition. Patients undergoing evaluation for fILD often experience delays, multiple procedures, occasional misdiagnoses, and dissatisfaction (2, 3). This is of particular concern for patients whose high-resolution computed tomography (HRCT) imaging is classified as “probable” or “indeterminate” for usual interstitial pneumonia (UIP) (1, 4).

In patients without a typical UIP pattern, current guidelines conditionally recommend a surgical lung biopsy (SLB) (1, 4). Despite improvements in selection of patients for SLB, this procedure is still associated with significant morbidity and mortality (5). In contrast to SLB, transbronchial biopsies (TBBx) have a low complication rate (<10% of procedures) and minimal mortality risk but are not a cornerstone of fILD diagnostic workups because of low sensitivity in diagnosing fILD subtypes (6). Transbronchial lung cryobiopsies have increased diagnostic yield; however, they remain limited in their application because of increased complication rate compared with TBBx, specialized implementation requirements, and interrater reliability issues (7).

Recent studies have demonstrated how a genomic classifier (GC) may be incorporated into the diagnostic evaluation of patients with fILD to increase diagnostic confidence without having to resort to SLB. In this issue of AnnalsATS, Lasky and colleagues (pp. 916–924) report using a GC consisting of 190 genes previously identified through machine learning analysis of whole-transcriptome mRNA sequencing of TBBx specimens in patients with fILD (8, 9). In the BRAVE (Bronchial Sample Collection for a Novel Genomic Test) study of patients undergoing diagnostic evaluation for fILD, this classifier had a sensitivity of 70% and specificity of 88% for the identification of UIP in comparison to SLB histopathology (10). The present study sought to evaluate how this GC would modify U.S.-based pulmonologist decision-making in cases of multidisciplinary discussion (MDD) expert consensus–diagnosed IPF without a typical UIP pattern on HRCT (8).

The authors isolated 11 cases of patients with fILD from the BRAVE study without a typical UIP HRCT pattern whose eventual MDD diagnosis was IPF. Subsequently, 103 U.S.-based pulmonologists performed 605 case reviews of the cases without knowledge of their final MDD diagnosis. The study evaluated both a pre- and post-GC cohort to evaluate the performance of the GC when implemented in a staged fashion. The investigators also used a second independent cohort to evaluate the GC’s performance when clinicians were provided the GC concurrently with all other clinical and radiographic information. In the pre/post cohort, IPF was listed as the diagnosis 30% of the time pre-GC, which increased to 69% post-GC. The surge in IPF diagnoses was accompanied by recommendations for antifibrotics, which increased from 10% pre-GC to 46% post-GC. Interestingly, when the GC was evaluated in independent cohorts, the effect size was substantially smaller. Physicians in the independent cohort, where clinical, radiographic, and GC data were all provided simultaneously, made a diagnosis if IPF 13% more often and recommended antifibrotics 11% more often than did physicians in the independent cohort who never received GC data.

The intended implementation timing for this GC is after initial evaluation of standard clinical and radiographic data, when more information is required to make a confident diagnosis. As such, the authors argue that the 36% increase in IPF diagnoses in the pre-/post-GC cohort is most representative of the clinical application of this tool. However, the differences in the rate of IPF diagnoses between the pre-/post-GC cohort and the independent cohorts highlights the potential for cognitive biases to influence clinician diagnostic decision-making. These factors need to be thoroughly considered as this tool is integrated into our ILD diagnostic armamentarium.

Cognitive biases are likely to act in both directions, amplifying the impact in the pre-/post-GC cohort while simultaneously decreasing the magnitude of effect in the independent cohorts. Previous studies indicate that supporting information has a stronger impact on decision makers when presented sequentially, rather than simultaneously, thus constituting a form of confirmation bias that may increase the proportion of respondents diagnosing IPF in the post-GC cohort (11). Furthermore, the “representativeness heuristic,” where people make judgements on the basis of preexisting mental frameworks, could contribute to an overestimation of the likelihood of a true diagnosis of IPF. If a clinician is given additional information that is consistent with their preconceived notion of IPF (i.e., a UIP-positive pattern on a GC) they are more likely to identify IPF as the diagnosis regardless of the baseline rate of IPF in the population in question (12). In the context of this manuscript, the baseline rate of “true IPF” was 100%, but this is not the case in real-world applications, where the base rate of a true IPF diagnosis in any context is variable, depending on multiple clinical and radiographic parameters. Although typically applied to monetary purchases, bundling bias, where a bundle of items is perceived as less valuable than the items individually (13), may act to lessen the impact of the GC in the independent cohorts. In this context, clinicians may fail to fully recognize the added value of the GC when it is presented with all other clinical and radiographic data. People have a limited capacity to process each piece of information immediately when receiving a large quantity of data simultaneously, so provision of the results of a GC may not be appropriately considered in this context (14). Based on this knowledge, we would argue that the true impact of the GC on IPF diagnoses is likely somewhere between the 36% and 13% reported in the pre-/post-GC and independent cohorts, respectively.

Lasky and colleagues should be commended for the design of this study, given that it provides an evaluation of not only the clinical performance of this GC but also how the implementation timing influences clinician decision-making (8). This study also provides insight into the real-world utility of this test, highlighting diagnostic differences between general U.S.-based pulmonologists and specialized ILD MDDs. But despite its many strengths, several questions warrant further study. Of 237 patients undergoing investigation for ILD, only 11 patients had a HRCT pattern not classified as typical UIP and a final MDD diagnosis of IPF. This sample size is quite small, reflecting the low frequency of this exact clinical scenario in which the GC may provide the greatest utility. Furthermore, all the patients who did meet these criteria were White, which limits the generalizability of these findings to more diverse cohorts of patients with fILD.

Last, one must raise the question of whether defining the diagnosis of IPF is necessary in these ambiguous cases, especially given recent studies demonstrating benefit of antifibrotic therapy in multiple subgroups of patients with fILD (15). As the authors argue, making a diagnosis of IPF may reduce the risk of initiation of harmful immunosuppressive therapies or may increase appropriate antifibrotic prescriptions in these patients (8). The next most important step in defining the GC’s role in the ILD clinician’s toolkit may be to evaluate the association between a UIP-positive result and clinical response to antifibrotics. Since the very first transcriptomic analyses of IPF lungs (16), the arrival of the GC may presage an era of precision diagnostics in patients with fILD.