Limited evidence exists regarding the dynamic prediction of AE. This study is the first to demonstrate a strong association between the patient-specific evolution of serum KL-6 levels and the risk of AE among patients with fibrotic ILD. The trajectories of serum KL-6 exhibited high heterogeneity and fluctuation; however, physicians should be vigilant when observing an upward trend of KL-6 or constantly high serum KL-6 values, as they may indicate prospective AE. Serum KL-6 is a simple and fast blood test that is especially beneficial for patients who struggle to perform pulmonary function tests effectively.
Previous studies primarily evaluated patients’ baseline characteristics, identifying sex, age, compromised pulmonary function, and radiological and histological patterns as potential predictors of poor prognosis, including AE [2, 24–26]. A high baseline serum KL-6 level is proposed as a predictor of disease progression, AE, and poor outcomes of ILDs [27]. In contrast, our study showed that the serum KL-6 level exhibited a considerable fluctuation, and the timing of fibrotic ILD diagnosis influenced the predictive value of baseline KL-6. Some smaller-scale studies suggested that longitudinally increased serum KL-6 levels may predict a decline in forced vital capacity and disease progression in IPF [28, 29]. The authors evaluated the correlation between serial KL-6 levels and patient outcomes but excluded patients when either KL-6 or pulmonary function test results were missing. This exclusion strategy assumes missing-completely-at-random; the probability of an outcome being missing is independent of both observed and unobserved factors [15]. In a non-standardized protocol, a patient's medical condition could highly influence missing data, especially in pulmonary results. In addition, their analyses did not consider KL-6 measurement errors. Notably, small sample sizes in research may inadvertently yield significant results by chance.
This study addressed AE, an outcome often omitted in previous longitudinal studies, and incorporated a large patient cohort with numerous measurement time points. To address methodological limitations, we employed a joint regression model with a longitudinal component that accommodates missing outcome data by assuming missing data not at random; this means that the probability of an outcome being missed may depend on any other observed covariates and previous KL-6 measurements. The model accounts for the impact of AE on the interruption of pre-AE KL-6 measurements. Furthermore, the linear mixed-effect model incorporated the measurement error inherent in KL-6 assessments. Finally, we expanded the disease scope to include fibrotic IPF, encompassing a range of heterogeneous ILDs, broadening KL-6’s applicability in treating ILDs.
This study has limitations. First, the accuracy of the data captured in this study remains uncertain. Although we used modified versions of a validated algorithm, the original algorithm solely evaluated the positive predictive value for AE rather than the sensitivity and specificity of the algorithm. In addition, in the Real World Data database, longitudinal follow-ups on patients were conducted longitudinally only at the same hospital; this could have skewed associations toward null, potentially increasing the likelihood of missing AE more than overdiagnosing these events. Second, our small sample size limited our ability to develop a more intricate and flexible model. For instance, we could not evaluate the contribution of the change rate in the patient-specific evolution of KL-6. Third, our study was primarily based on administrative claim data and selected data from electronic medical records; thus, we could not incorporate all variables. Incorporating more time-independent and time-dependent covariates, such as pulmonary function test results, dyspnea scores, physical function, and imaging results, could enhance predictive accuracy. Lastly, our study focused on the Japanese population. The genotype distribution may cause differences in KL-6 levels [30, 31]. Further large-scale studies are needed to validate our findings and to develop a more accurate dynamic prediction model for AE.