Immune Checkpoint Inhibitor Pneumonitis and Real‐World Evidence—Proceed With Care

Over the last decade, the field of lung cancer has changed. Treatment options have expanded and outcomes have improved. Immune checkpoint inhibitors (ICIs) contribute to prolonged survival in advanced lung cancer as well as in resectable disease [1]. ICIs improve outcomes in both adjuvant and neoadjuvant settings, particularly in patients with high tumour PD-L1 expression [2]. But with these powerful medications come important side effects, with checkpoint inhibitor pneumonitis (CIP) among the more critical. CIP, initially considered uncommon, may develop in 10%–20% of ICI recipients and has the highest mortality rate of the immune-related adverse effects (irAEs) [3]. Risk factors for CIP include PD-1/PD-L1 agents (compared with CTLA-4 inhibitors), combination immunotherapy, radiotherapy, and, compared with other tumours, treatment for lung cancer [3]. The presence of existing lung disease, such as interstitial lung disease (ILD) and chronic obstructive pulmonary disease (COPD), may heighten the risk of therapeutic side effects in people with lung cancer. ILD is associated with a higher incidence of CIP in ICI recipients, but the extent remains unclear. A systematic review investigating outcomes in patients with underlying ILD treated with ICI for lung cancer analysed data from 10 studies including 179 participants with relatively mild pre-existing ILD (median percent FVC 93.4%), calculating pooled overall CIP incidence of 27% (95% CI 17–37) and grade 3 or higher CIP incidence of 15% (95% CI 9–22) and noting elevated study heterogeneity (I2 = 54.8%) [4]. The incidence of CIP was significantly higher with pre-existing ILD than without (OR 3.23, 95% CI 2.06–5.06 vs. OR 2.91, 95% CI 1.47–5.74) with time to CIP occurrence from 31 to 74.5 days [4]. Fewer studies have investigated the association between COPD and CIP, although a recent systematic review of patients with lung cancer and underlying COPD provides some evidence [5]. This review performed a safety analysis based on 14 studies with an overall total of 5549 patients in receipt of ICI therapy, of whom 811 (14.6%) developed CIP. For patients with COPD and CT-defined emphysema (denominator not provided), the authors calculate a pooled overall CIP incidence of 20% (95% CI 11–35), grade 3 or higher CIP incidence of 7% (95% CI 4–9), and OR 1.97 (95% CI 1.41–2.76), noting high study heterogeneity (I2 = 70%). Time to CIP occurrence ranged from 3 to 879 days with most cases improving with discontinuation of ICI and corticosteroid therapy [5]. The benefits of ICI therapy in lung cancer can transform outcomes, but people with underlying ILD or COPD may face significant risks. Clinicians need more data on CIP to guide treatment than currently available from clinical trials. In a recent publication in Respirology, Maule et al. [6] have sought answers from real-world (RW) data on the risks and outcomes of CIP in lung cancer patients with and without underlying ILD and COPD. The researchers present a retrospective analysis of United States data from the TriNetX platform [7] which collects de-identified electronic health record (EHR) data from 70 US healthcare organisations. The analysis was performed on two mutually exclusive cohorts of lung cancer patients in receipt of ICI therapy: one cohort with underlying ILD and one with underlying COPD according to ICD-10 coding [8]. Other coding systems were used to further define the cohorts. National Library of Medicine RxNorm codes [9] were used to confirm receipt of ICI therapy or to exclude patients with exposure to several other medications that may have caused pneumonitis. Current Procedural Terminology (CPT) codes [10] were used to identify clinical outcomes related to critical care. Propensity score matching (PSM) controlled for baseline characteristics and minimised confounding. While the overall study denominator consisted of 16,736 patients, the outcome measures resulted from smaller, post-PSM samples with 1568 patients in each of the ILD and COPD cohorts and just over 2000 patients in the non-ILD/non-COPD cohort. Depending on analysis, the denominators vary. Key findings included mean age of 72 years, predominantly White (73.7% ILD, 73.4% COPD) and male (59.4% ILD, 58.3% COPD) study subjects with similar baseline data for the reported comorbidities of hypertension, chronic kidney disease and diabetes. The primary outcome, incidence of CIP, measured 4.6% (70/1510) in the ILD cohort, 1.9% (30/1566) in the COPD cohort and 1.5% (32/2145) in the non-ILD/non-COPD cohort. This represents 132 patients with CIP. Secondary outcomes included all-cause mortality which was lower in the ILD cohort than the COPD cohort (8.9% vs. 12.3%, p = 0.0026), acute respiratory failure and ICU admission which was similar in the ILD and COPD cohorts (20.6% vs. 21.0%, p = 0.79) and CIP recurrence with more episodes per patient in ILD than COPD (0.156 vs. 0.040, p < 0.001). The researchers discussed the strengths of the study including the large initial study sample, the balance of characteristics and the insights into high-risk patients possible from real-world data. The researchers also addressed limitations including the retrospective design, missing data such as lung function, smoking history, cancer stage and treatment details. Other critical missing information included ILD subtypes, diagnostic information to distinguish CIP and ILD exacerbation and risk factors such as radiation exposure. In the face of these limitations, how does this RW dataset help with the research question? RW data can fill gaps left by clinical trials, providing information on ‘subgroups of patients excluded or under-represented in RCTs’ [11] or addressing the need for information on rare subgroups where clinical trials can be too difficult to run. RW data can ‘challenge the magnitude’ of findings made by randomised controlled trials [11]. RW evidence can boost clinical trials by providing control data [12] which come from standard resources including ‘electronic health records, patient registries and administrative health care claims’ [13]. We have seen in the study by Maule et al. [6] some of the shortcomings of RW data which include missing information, baseline comorbidities of limited relevance and insufficient clinical information for accurate interpretation of results. The sample size looks promising but on closer inspection shrinks to a fraction (n = 132) of the original. The data have a curious opacity, how do we the readers or they the researchers assess their quality? Can we trust the coding? Examine the methods and you can see that the statistical analyses ‘were conducted within the TriNetX platform’ [6], but by whom? Does the absence of radiation data undermine the results? The RW dataset used in this analysis has potentially significant deficiencies, appearing convenient for the data source but not necessarily suited to the research question. The authors deserve congratulations for their extraction of the study findings, for ‘separating signal from noise’ [14]. But as they say themselves, ‘Without … details, it is difficult to fully assess’ [6] the impact of CIP and how to manage it. The key insights from this study and therefore its true value come perhaps not from the results on CIP, but from an understanding of the limitations of real-world data. The author has nothing to report. Speaker honoraria—Astra Zeneca, Merck Sharp & Dohme, The Limbic, meeting attendance support by Astra Zeneca LOBES 2022, Ad hoc advisory board payment—Bristol Myers Squibb.