Interstitial lung diseases (ILD) are associated with varying clinical trajectories and poor prognosis [1]. Although various ILD-specific prognostic markers exist, their generalizability to multi-ethnic Asian populations remains uncertain [2]. The widely-used ILD-gender-age-physiology (GAP) model, validated in Western cohorts, predicts 1-, 2- and 3-year mortality [3]. It consists of five parameters including ILD subtype, gender, age, percent predicted forced vital capacity (FVC %pred) and percent predicted diffusing capacity of the lungs for carbon monoxide (DLCO %pred).

In contrast, the PROgnostic ModEl for chronic Lung disease (PRO-MEL) was developed and validated to estimate 1-year mortality across chronic lung diseases, including chronic obstructive pulmonary disease, bronchiectasis, and ILD [4]. PRO-MEL incorporates 11 predictors, with higher mortality risk associated with older age, male sex, non-Chinese ethnicity, body mass index (BMI) ≤18.5 kg/m², supplemental oxygen use, history of cancer, diagnosis of ILD, prior tuberculosis infection, and requiring assistance in at least one activity of daily living. Having at least one specialist outpatient consultation within 6 months of diagnosis was protective. While PRO-MEL was not developed specifically for ILD, its variables may capture broader aspects of disease burden and patient health status which may determine mortality, potentially complementing existing ILD prognostication tools. It has the advantage of using routinely available clinical information in electronic health records to predict 1-year mortality. To date, no study has directly compared PRO-MEL with ILD-GAP specifically in ILD populations. We aimed to evaluate agreement between PRO-MEL and ILD-GAP. Secondary outcomes included the models’ discriminatory performance, assessed using the area under the curve (AUC), and characterize the high-risk patient profiles identified by each model.

We conducted a retrospective multi-center cohort study of adults (≥21 years) diagnosed with ILD from three tertiary hospitals in Singapore (National University Hospital, Singapore General Hospital, Tan Tock Seng Hospital [TTSH]) between February 2016 to September 2023. Participants were enrolled in the TTSH ILD registry (National Healthcare Group Domain Specific Review Board Study Reference No. 2019/00894) and the multi-center Singapore ILD registry (SingHealth Centralised Institutional Review Board Reference No. 2022/2422) with written informed consent at enrolment, except those recruited prior to November 2017, where consent waiver was granted. ILD was diagnosed according to standardized international diagnostic criteria [5]. Patients with 1-year follow-up and lung function tests within 3 months of index visit were included. Individuals with PRO-MEL risk score of ≥0.3 were classified as high-risk, which predicted a 1-year mortality of 16.2% in the external validation cohort [6]. The ILD-GAP high-risk threshold was set at ≥4, corresponding to a 18.2% 1-year mortality, which is comparable to the estimate provided by PRO-MEL. Data was extracted from electronic medical records, with 6-monthly follow-up until death, lung transplant, or study withdrawal. No patient underwent lung transplantation during the study period.

Descriptive statistics were reported as numbers with frequencies, mean±standard deviation (SD), or median (interquartile range [IQR]) as appropriate. Agreement between PRO-MEL and ILD-GAP was assessed using Cohen’s kappa. Model performance was assessed using the AUC. Missing data was not imputed. Analyses were performed using SPSS version 28.0.0.0 (IBM Corp. Armonk, NY, USA), with significance set at p<0.05.

Among 221 patients, observed 1-year mortality was 29.9%. The cohort had 111 males (50.2%), predominantly Chinese (n=166, 75.1%), with 66.1% non-smokers (n=146). Most patients were symptomatic, with 67.9% (n=150) experiencing cough and 77.8% (n=172) experiencing dyspnoea. The median modified Medical Research Council (mMRC) dyspnoea score was 2 (IQR, 1 to 3). Median BMI was 24.0 kg/m² (IQR, 21.3 to 26.9). Thirty patients (13.6%) were on supplemental oxygen. The most common diagnosis was CTD-ILD (n=133, 60.2%), followed by idiopathic pulmonary fibrosis (IPF) (n=50, 22.6%) and hypersensitivity pneumonitis (n=15, 6.8%). Mean±SD FVC was 69.5%±19.8% predicted, and mean DLCO was 48.4%±19.3% predicted.

Mortality rates in high-risk groups were 51.3% (PRO-MEL) and 53.6% (ILD-GAP). Both models showed similar predictive performance with PRO-MEL achieving an AUC estimate of 0.71 (95% confidence interval [CI], 0.64 to 0.79) and ILD-GAP achieving 0.72 (95% CI, 0.65 to 0.80). Agreement between models was fair (Cohen’s kappa=0.237, p<0.001). Notably, patients classified as high-risk by PRO-MEL but low-risk by ILD-GAP were more likely female, had higher Charlson comorbidity index, lower BMI, more frequent CTD-ILD, and better-preserved lung function compared with those classified as PRO-MEL low-risk but ILD-GAP high-risk (Table 1). Model performance differed across ILD subtypes. In CTD-ILD, PRO-MEL showed better discrimination (AUC, 0.73; 95% CI, 0.61 to 0.85) compared with ILD-GAP (AUC of 0.66; 95% CI, 0.53 to 0.79). In IPF, the models performed similarly, though neither achieved statistical significance (PRO-MEL AUC, 0.61 [95% CI, 0.45 to 0.77] vs. ILD-GAP AUC, 0.63 [95% CI, 0.48 to 0.78]).

Both PRO-MEL and ILD-GAP demonstrated fair accuracy in predicting 1-year mortality. However, their low agreement suggests they identify different at-risk patient subgroups, potentially complementing each other in clinical practice. The discordant classification reflects fundamentally different prognostic model constructs rather than poor model performance. Hence, PRO-MEL may flag patients with significant comorbidity or functional limitations with relatively preserved lung function, who are yet to meet ILD-GAP physiological severity thresholds, but remain at high-risk of near-term mortality. Indeed, in this study, PRO-MEL performed better than ILD-GAP in predicting 1-year mortality in CTD-ILD patients, which may be due to ILD-GAP assigning negative scores for CTD-ILD, indicating better prognosis [3]. However, a subgroup of CTD-ILD patients develop progressive pulmonary fibrosis despite conventional treatment, and follow an IPF-like disease trajectory [6]. PRO-MEL’s broader set of predictors may better capture this subgroup.

This study has several limitations. This was a retrospective study, and although it was multi-center, the sample size remains modest. Larger external validation studies across diverse Asian ILD populations are needed to confirm generalizability. As CTD-ILD constitutes the majority of the cohort, the heterogeneity in disease behaviour [7] may limit interpretation of performance of PRO-MEL and ILD-GAP in specific CTD-ILD subgroups. Additionally, this study focused on 1-year mortality, whereas ILD-GAP was originally designed for longer-term prediction. This inherently places ILD-GAP at a relative disadvantage when comparing model performance over 1-year, and this may partly account for the modest performance of ILD-GAP observed in this study. Further research could explore whether PRO-MEL maintains discriminatory performance over a longer follow-up period.

Both PRO-MEL and ILD-GAP demonstrated fair discriminatory ability for 1-year mortality in Singaporean ILD patients, with fair agreement between the models. The models identify distinct high-risk subgroups, reflecting their different derivation populations and predictor structures. These findings support their complementary use to enhance risk stratification, and potentially guide timely supportive and palliative care interventions, which was the original intention of developing PRO-MEL. Future research should focus on prospective validation and integrating prognostic tools into clinical pathways that help to identify ILD patients that should be prioritized for early supportive and palliative care.