Introduction

Interstitial lung diseases (ILDs) encompass a large group of idiopathic and secondary pathologic lung conditions. Regardless of etiologies and initial patterns, certain ILDs may lead to irreversible fibrotic destruction of lung parenchyma, evolving into fibrosing ILDs (fILDs)1,2,3. There is heterogeneity in trajectories and prognosis among patients with fILDs. Currently, there are no well-established biomarkers for predicting the risks of acute exacerbation (AE) and death, even in patients with the most severe subtypes of fILDs, such as idiopathic pulmonary fibrosis (IPF) and progressive pulmonary fibrosis (PPF)3,4,5,6,7,8.

Pulmonary function tests and high-resolution computed tomography (HRCT) are essential parts in the evaluation of fILDs. Poor pulmonary function and honeycombing on HRCT at baseline have been reported as possible predictors of subsequent progressive fibrosis, AE, and decreased survival1,5,6,7,8,9. However, because fILDs are clinically characterized by cough and dyspnea, obtaining accurate and reproducible measurements of spirometric volumes and diffusion capacity for carbon monoxide (DLCO) can be challenging in practice, as these symptoms often interfere with testing performance3,4. While honeycombing is characteristic of the most fibrotic “usual interstitial pneumonia” (UIP) pattern, fILDs can also manifest as non-UIP patterns1,3,6,10,11. AE may occur in patients with fILDs who exhibit only mild respiratory impairment and no honeycombing8,9.

Fibrosing ILDs can lead to significant weight loss12,13,14. Body mass index (BMI), a widely recognized nutritional metric12, has been linked to adverse outcomes in patients with fILDs, particularly IPF12,13,14,15,16,17,18,19,20. While underweight patients tend to have the poorest prognosis, even those with a normal BMI may experience worse outcomes compared to individuals who are overweight or obese13,14,15,16,19. Moreover, several circulating macromolecules have emerged as potential blood biomarker candidates for the prognosis of fILDs21,22. Mucin-1, also known for its epitopic derivatives such as Krebs von den Lungen-6 (KL-6) and CA15-3, is a transmembrane O-glycoprotein expressed by type II alveolar cells23,24. Pentraxin 3 (PTX3), on the other hand, is an acute-phase pattern-recognition molecule of the long-pentraxin family produced by multiple cell types25. Both mucin-1 and PTX3 are involved in crucial cellular functions and have been implicated in fibrogenesis23,24,25, with their blood levels elevated in fILDs. Previous studies have shown correlations between blood mucin-1/KL-6/CA15-3 levels and the severity of respiratory impairment, as well as AE risks, in various ILDs22,23,24. Recently, we reported the utility of plasma mucin-1 levels, with a cutoff value of 2.5 ng/mL, in predicting risks of AE and death in two distinct cohorts of patients with IPF receiving different antifibrotics26,27. We also found that baseline plasma PTX3 levels, with a cutoff value of 2.2 ng/mL, predicted risks of AE and death in patients with fILDs25. Nevertheless, to date, neither macromolecule has been definitely accepted as prognostic biomarkers for fILDs.

In this study, we aimed to introduce a novel composite “BMP index” for predicting the risks of AE and early death in patients with fILDs. The BMP index mitigates the limitations of spirometry and DLCO, and is noninvasive and radiation-free. It incorporates three surrogate indicators of different pathogenetic aspects of fILDs. The “B” component represents the baseline BMI, indicating the systemic catabolic effect of fILDs. The “M” component represents baseline plasma mucin-1 levels, indicating alveolar fibrotic destruction. The “P” component is determined by baseline plasma PTX3 levels, indicating alveolar injury.

Results

We enrolled a total of 366 patients, including 347 from NCKUH, 10 from TH (of whom 1 was enrolled on site and 9 were transferred to NCKUH for enrollment and subsequent management), and 9 from CH (Fig. 1). The median age of the whole cohort was 71.9 years (IQR 64.4–79.2), with 270 (74%) being male. Overall, the median follow-up duration was 117 weeks (IQR 56–239). All patients were followed for at least 6 months, and 301 (82%) for at least 2 years, or until death. Specific fILD diagnoses of the patients are listed in Supplementary Table S1. A total of 111 patients (all having non-IPF fILDs) had initiated immunosuppressive therapy prior to study enrollment, whereas 183 patients subsequently received antifibrotic treatment during the study period. After stratified randomization, 218 patients were assigned to the derivation cohort and 148 to the validation cohort. There were no significant differences between the two cohorts in terms of characteristics, treatments, and outcomes (Table 1).

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

The inclusion and exclusion flowchart of this study.

Table 1 Baseline characteristics and outcomes of the derivation and validation cohorts.
Full size table

BMP index and AE

During the study period, 108 patients (30%) experienced at least one episode of AE (60 patients in the derivation and 48 in the validation cohort). The median time between study enrollment and the first AE was 21.5 weeks (IQR 4.8–72.3), and 89 patients (82%) developed first episode of AE within 2 years of enrollment (47 patients in the derivation and 42 in the validation cohort). In both cohorts, patients who developed AE were more likely to have, at baseline, a lower BMI, higher CCI, higher levels of mucin-1 and PTX3, higher probabilities of PH, poorer pulmonary functions, higher GAP stages, and higher BMP indices compared to patients without AE (Supplementary Table S2). Higher BMP indices indicated greater proportions of patients experiencing AE during follow-ups in both cohorts (Supplementary Table S3), irrespective of immunosuppressant use before study enrollment (Supplementary Table S4) or subsequent administration of antifibrotic therapy (Supplementary Table S5). In Kaplan–Meier analyses, patients in the derivation cohort with BMP indices ≥ 2 had higher cumulative probabilities of AE during follow-up, with index 3 associated with even higher probabilities than index 2. This positive correlation with AE probabilities was not prominent for indices 0 and 1 in the derivation cohort. Nevertheless, in the validation cohort and the combined whole cohort, the BMP index effectively stratified the probabilities of AE starting from index 1, indicating that patients with consecutively higher BMP indices had higher cumulative probabilities of AE (Fig. 2A,B,C). Univariate and multivariable Fine-Gray regression analyses yielded concordant results: starting from index 2 in the derivation cohort and from index 1 in the validation and whole cohorts, increasing BMP indices indicated consecutively higher subdistribution hazard ratios of AE (Table 2). When the analysis was restricted to patients who developed AE within two years of enrollment, the BMP index continued to demonstrate effective stratification of AE risk (Supplementary Fig. S1A–C and Supplementary Table S6).

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Kaplan–Meier curves illustrating the cumulative probability of developing acute exacerbation in the derivation cohort (A), validation cohort (B), and the combined whole cohort (C).

Table 2 Univariate and multivariable Fine-Gray subdistribution regression analyses of the risk of acute exacerbation in patients across the derivation, validation, and combined whole cohorts.
Full size table

BMP index and early death

102 patients (28%) died within two years after diagnosis of fILDs (55 patients in the derivation and 47 in the validation cohort), with a median time of 30.6 weeks (IQR 10.9–71.6) between study enrollment and death. Most patients died from respiratory causes (Supplementary Table S7). Similar to the findings concerning AE, patients who died early were more likely to have, at baseline, a lower BMI, higher CCI, higher levels of mucin-1 and PTX3, higher probabilities of PH, poorer pulmonary functions, higher GAP stages, and higher BMP indices compared to patients who survived (Supplementary Table S8). In both derivation and validation cohorts, the proportions of patients who died early increased with the BMP index (Supplementary Table S3) regardless of immunosuppressant use before study enrollment (Supplementary Table S4) or subsequent administration of antifibrotic therapy (Supplementary Table S5). In Kaplan–Meier analyses, different BMP indices stratified patients with different cumulative survival rates within 2 years of enrollment (the time of fILD diagnosis; Fig. 3A,B,C). In univariate and multivariable Cox proportional hazard regression analyses, increasing BMP indices from index 1 indicated consecutively higher hazard ratios of early death in the derivation, validation, and combined whole cohorts (Table 3).

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

Kaplan–Meier curves illustrating the different probabilities of survival within 2 years of enrollment in the derivation cohort (A), validation cohort (B), and the combined whole cohort (C).

Table 3 Univariate and multivariable Cox proportional hazard regression analyses of the risk of early death in patients across the derivation, validation, and combined whole cohorts.
Full size table

Comparative performance analyses

For predicting both AE and early death, univariate models using the complete BMP index consistently achieved the highest Td-AUCs and C-indices across all cohorts, outperforming models incorporating individual “B,” “M,” or “P” components. Supplementary Table S9 presents the distribution of patients experiencing AE and early death across distinct strata defined by nine previously reported predictors. Models involving the BMP index showed consistently higher Td-AUCs and C-indices than those based on monocyte counts, NLR, LMR, CPI, the CPB index, AISI, LIPI, or the HAL index. Compared to models using GAP staging for AE risk prediction, the BMP index exhibited comparable (as in the validation cohort) or slightly higher (as in the derivation and whole cohorts) Td-AUCs and C-indices. For early death prediction, models with the BMP index had Td-AUCs and C-indices that were comparable (as in the derivation cohort) or slightly lower (as in the validation and whole cohorts) to those using the GAP index but remained within a strong predictive range (Tables 5 and 6).

Table 4 Time-dependent areas-under-the-curves at median times to acute exacerbation and early death of various univariate predictive models.
Full size table
Table 5 Harrell’s C-indices of various univariate predictive models assessing risks of acute exacerbation and early death.
Full size table

Sensitivity analyses

We assessed the prognostic performance of the BMP index in the aforementioned pooled subgroups: subgroup “IPF” (235 patients) versus subgroup “Non-IPF” (131 patients), and subgroup “AntiF” (183 patients) versus subgroup “No-AntiF” (183 patients). The proportion of patients experiencing either AE or early death increased consistently across all subgroups as BMP index values rose (Supplementary Table S10). For the prediction of AE, the BMP index stratified patients with consecutively increasing subdistribution hazard ratios in all subgroups (Supplementary Table S11). For the prediction of early death, the BMP index stratified patients according to different risks in the subgroups “Non-IPF” and “AntiF”, with consecutively higher indices indicating higher risks. In the subgroups “IPF” and “No-AntiF”, indices 2 and 3 indicated higher risks than that of indices 0 and 1, though the adjusted hazard ratios did not increase from index 2 to 3 (Supplementary Table S12). Additionally, when the cutoff for the “B” component was changed from 24 to 25 kg/m2, the BMP index demonstrated comparable performance in predicting AE and early death risks, except in the validation cohort, the adjusted hazard ratio for early death associated with index 3 was not higher than that of index 2 (Supplementary Tables S13–S15). Furthermore, despite a hypothetically unidentified confounder with varying prevalence across all three cohorts, all multivariable analyses produced consistent results, supporting the robustness of our findings (Supplementary Figs. S2–S4).

Discussion

In this study, we proposed and validated the novel BMP index for predicting severe adverse outcomes in patients with fILDs. The BMP index differentiated risks of AE and early death in all the study cohorts, with higher indices indicating higher risks. Each of the three components of the BMP index represents distinct but relevant pathophysiological mechanisms in fILDs. This multi-dimensional approach adds to its biological plausibility.

Compared to nine previously reported categorical prognostic predictors (including monocyte count, NLR, LMR, CPI, CPB index, HAL index, AISI, LIPI, and GAP staging28,29,30,31,32,33,34,35,36,37) the BMP index demonstrated superior or comparable predictive performance in terms of Td-AUCs and C-indices. Notably, for AE prediction, it outperformed all others, including GAP staging, which remains the most widely used prognostic indicator in clinical practice. Moreover, the BMP index circumvents the limitations of pulmonary function testing and radiation exposure associated with CT imaging, making it radiation-free and advantageous particularly for patients unable to undergo spirometry or DLCO measurements.

Previous studies have identified BMI as a predictor of mortality risk in fILDs13,15,17,18,19,20, a finding supported by our present study. However, its prognostic role in AE remains uncertain. Yoon et al. reported that underweight IPF patients faced a higher risk of respiratory and all-cause hospitalization18. Among AE patients, Awano et al. found the highest in-hospital mortality risk in underweight individuals and the lowest in obese patients15. Other studies, however, reported no significant association between BMI and AE20,38. In our study, which included both IPF and non-IPF fILDs, BMI was inversely associated with AE occurrence (Supplementary Table S3), supporting the potential link between lower BMI and an increased risk of AE. Moreover, we adopted 24 kg/m2 as the BMI cutoff for scoring the “B” component, aligning with Taiwan’s officially established upper normal limit. Nevertheless, sensitivity analysis using 25 kg/m2 (which was commonly applied in previous studies conducted in European populations) yielded comparable results, supporting the feasibility of using this general WHO standard for the BMP index.

The present study has limitations. The influence of unidentified confounders (including the potential modifying effects of immunosuppressive and antifibrotic therapies) could not be entirely excluded, despite substantial efforts to mitigate this, such as comprehensive data collection, strict statistical control of confounders, validation using a distinct cohort, and sensitivity analyses of the major findings. Quantification of mucin-1 and PTX3 requires only a small blood volume and relies on commercially available and relatively inexpensive sandwich ELISA methods. These biomarkers, however, are not routinely measured in current clinical practice. Additionally, all participants were from hospitals in southern Taiwan. Biomarker levels, such as mucin-1 and PTX3, may be influenced by ethnic and genetic factors, including single nucleotide polymorphisms39,40, which were not assessed in this study. Although we demonstrated that the BMP index effectively stratifies the risk of AE and early death in the pooled subgroup “Non-IPF”, its predictive performance across distinct individual non-IPF ILD subtypes remains to be elucidated. Therefore, further prospective research, ideally involving a large and ethnically diverse patient population with IPF as well as various non-IPF ILDs, is warranted to evaluate the applicability of the BMP index across different ethnic groups and to determine its performance in the context of various genetic backgrounds and ILD subtypes.

In summary, the BMP index offers a noninvasive, spirometry-independent, and biologically grounded prognostic tool that facilitates early risk stratification in fILD patients. It may be especially beneficial in clinical settings where lung function testing is not feasible, and its simplicity makes it suitable for routine or even remote application. Future studies incorporating genetic profiling and broader international cohorts are warranted to confirm and extend its clinical utility.

Methods

Study design and population

This observational study included patients with fILDs who had been diagnosed, treated and followed between January 1, 2017, and May 31, 2025, at one of the three participating hospitals in southern Taiwan: National Cheng Kung University Hospital (NCKUH, a medical center), Tainan Hospital (TH, a regional hospital), and Chest Hospital (CH, a district hospital). This study was conducted in accordance with the Declaration of Helsinki and Taiwan’s Human Subjects Research Act. The study protocol was jointly approved by the Institutional Review Board of NCKUH (B-ER-105-390; A-ER-109-321; A-ER-111-595; B-ER-113-004). To be eligible, patients were required to be ≥ 20 years old, have a diagnosis of fILD confirmed by multi-disciplinary assessment, and have completed follow-up at the recruiting hospitals.

Data collection and patient randomization

At study enrollment, basic demographic data (including BMI), specific fILD diagnosis, Charlson Comorbidity Index (CCI), smoking status, probability of pulmonary hypertension (PH), pulmonary function measurements, antifibrotic therapy, immunosuppressive treatments (for non-IPF fILDs), and gender-age-physiology (GAP) stages were recorded for each patient. Since all participants were followed at one of the three study hospitals, details of any subsequent major adverse events (including hospitalization, lung transplantation, and death) were also documented. During each follow-up visit, patients were carefully evaluated, and their electronic medical records, including those from the cloud-based medical records of National Health Insurance System (which covers over 99.5% of Taiwan’s population41, were regularly reviewed to update data on relevant treatments and adverse events, including those occurring at other medical institutions. We performed stratified randomization based on sex, types of fILDs (IPF or non-IPF), and GAP stages, to allocate all included patients into either a derivation cohort or a validation cohort in a 3:2 ratio.

Important definitions

In this study, fILDs included both IPF and non-IPF fILDs. IPF was diagnosed according to international guidelines3,42. We defined a patient as having non-IPF fILD if the patient exhibited fibrosing features (traction bronchiectasis and/or honeycombing), in addition to the underlying ILD patterns, on HRCT images1,3,8. Acute exacerbation of fILDs was defined based on the 2016 international working group report, specifically excluding events with identifiable etiologies9,43. Early death referred to all-cause mortality within 2 years after diagnosis of fILDs. Fibrosing ILDs are typically associated with reduced survival, though heterogeneity exists. Patients who die within two years of diagnosis represent a high-risk group warranting heightened clinical attention. The probabilities of PH were determined based on transthoracic echocardiographic findings according to the 2022 ESC/ERS Guidelines44.

Measurement of plasma mucin-1 and PTX3 levels

Baseline plasma levels of mucin-1 and PTX3 were measured at the time of study enrollment. After obtaining formal informed consent, an 8 mL blood specimen was collected from the patient using an EDTA vacuum collection tube. The specimen was immediately ice-bathed, centrifuged (at 4 °C and 1500 x g for 20 min using an Eppendorf Centrifuge 5810R), and the plasma extracted within 4 h under sterile conditions. We used the sandwich enzyme-linked immunosorbent assay (ELISA) kit (product number EH0406) from Fine Test (Wuhan Fine Biotech, Wuhan City, China) to measure mucin-1 levels, which employed antibodies targeting the tandem-repeat domain of the extracellular N-subunit of mucin-1. For measuring PTX3 levels, we used the ELISA kit (product number #E20031001) from Leadgene Biomedical, Inc.; Tainan, Taiwan, involving antibodies targeting the C-subunit of PTX325,45. We performed all ELISAs in triplets according to manufactures’ protocols, and intra- and inter-assay coefficients of variation of < 15% were considered acceptable. All experiments were conducted at the Centre for Clinical Medical Research, NCKUH.

Structure of the “BMP index” and cutoff values for scoring

Our proposed BMP index comprises three components. For the “B” component (baseline BMI), Taiwan’s officially established upper normal limit, modified from the WHO recommendation and set at 24 kg/m², has been adopted as the cutoff value46. For the “M” component (baseline plasma mucin-1 levels), a cutoff value of 2.5 ng/mL has been adopted from our previous studies involving patients with IPF26,27. For the “P” component (baseline plasma PTX3 levels), a cutoff value of 2.2 ng/mL has been adopted also from our previous work25. In practice, a score of 1 is assigned to each of the three components if meeting the following conditions: BMI < 24 kg/m2, mucin-1 level ≥ 2.5 ng/mL, and PTX3 level ≥ 2.2 ng/mL. A score of 0 is assigned otherwise. The BMP index is the sum of these scores, ranging from 0 to 3.

Statistical analysis

We present categorical variables as counts and percentages, and continuous variables as means (standard deviation) if normally distributed or medians (interquartile range [IQR]) if not normally distributed. We did not impute missing values. We used Fischer’s exact test or the Mann–Whitney U test, whichever was appropriate, to test for non-random between-group differences in variables. We conducted Kaplan–Meier analysis with the log-rank test to stratify the univariate effect of BMP index on cumulative probabilities of AE and survival. The relationship between BMP index and the risk of AE over time was further analysed using multivariable Fine-Gray subdistribution regression, adjusting for GAP stages (a categorical quantification of age, sex, forced vital capacity and DLCO), CCI, probabilities of PH, and competing risks of death. The risk of early death among patients with different BMP indices were analysed using multivariable Cox proportional hazards regression, adjusting for GAP stages, CCI, and probabilities of PH. There was no multicollinearity among the variables. The Schoenfeld test was conducted to verify the proportional hazards assumption. The performance of the BMP index in the derivation cohort was independently verified in the validation cohort, and assessed in the pooled whole cohort. We analysed time-dependent areas-under-the-curve (Td-AUC) at median times to adverse events and Harrell’s C-index (C-index) for univariate predictive models across all cohorts. The Td-AUC and C-index of the BMP index were then compared with several recently reported categorical prognostic predictors, including GAP stages28, monocyte counts29, neutrophil to lymphocyte ratio (NLR)30, lymphocyte to monocyte ratio (LMR)31, the composite physiologic index (CPI)32,33, the clinical physiological biomarker (CPB) index34, the aggregate index of systemic inflammation (AISI)35, the lung immune prognostic index (LIPI)36, and the honeycomb-age-LDH (HAL) index37. Sensitivity analyses were conducted in three stages. First, we evaluated the prognostic performance of the BMP index across subgroups within the pooled whole cohort: patients with IPF (subgroup “IPF”) versus non-IPF fILDs (subgroup “Non-IPF”), and those receiving antifibrotic therapy (nintedanib or pirfenidone; subgroup “AntiF”) versus those not receiving (subgroup “No-AntiF”). Second, we tested the BMP index using a 25 kg/m2 cutoff (frequently employed in previous studies) for the “B” component, instead of the original 24 kg/m2 cutoff. Thirdly, we examined the potential influence of unidentified confounders on the observed associations. All statistical tests were two-tailed, and p < 0.05 was considered statistically significant. Statistical analyses were performed using R (version 4.4.3; The R Foundation for Statistical Computing, Vienna, Austria; https://www.r-project.org/) and MedCalc (version 22.0.2; MedCalc Software Ltd, Ostend, Belgium).