Admission-Based Machine-Learning Models for Predicting Mechanical Ventilation and Mortality in Fibrotic Interstitial Lung Disease: A Multicenter Cohort Study with External Validation

Abstract

Background

Admission-based risk stratification tools are limited for hospitalized patients with fibrotic interstitial lung disease (F-ILD).

Aims

To develop and externally validate admission-based machine-learning models for predicting mechanical ventilation (MV), 30-day and 3-month mortality, and long-term all-cause and cause-specific mortality in hospitalized patients with F-ILD.

Study Design

Multicenter retrospective cohort study.

Methods

This study included hospitalized adults with F-ILD from two tertiary hospitals in China. Clinical characteristics and laboratory test results obtained within 24 hours of admission were used as candidate predictors. Machine-learning models were developed to predict MV, 30-day and 3-month mortality, and long-term all-cause and cause-specific mortality, with internal testing and independent external validation. Model performance, clinical utility, and interpretability were evaluated.

Results

A total of 1,272 patients were included (derivation cohort, n = 1,006; external validation cohort, n = 266). In the external validation cohort, the models demonstrated robust discrimination for MV [area under the curve (AUC), 0.913], 30-day mortality (AUC, 0.926), and 3-month mortality (AUC, 0.813). For long-term all-cause mortality, the final model achieved a C-index of 0.768, with time-dependent AUCs of 0.811 at 12 months and 0.768 at 24 months. The corresponding values for cause-specific mortality were 0.768, 0.815, and 0.761, respectively. Lactate dehydrogenase (LDH), neutrophil-to-lymphocyte ratio (NLR), and prognostic nutritional index (PNI) emerged as key predictors of MV and short-term mortality, whereas long-term mortality risk was additionally associated with older age, male sex, smoking history, an idiopathic pulmonary fibrosis phenotype, and lower albumin-to-globulin ratio and platelet-to-white blood cell ratio.

Conclusion

Admission-based models demonstrated strong discriminatory performance for predicting MV and mortality in hospitalized patients with F-ILD. Elevated LDH and NLR levels, together with lower PNI, characterized a high-risk profile for acute deterioration and death. Local recalibration may be necessary before implementation in new clinical settings.

INTRODUCTION

Fibrotic interstitial lung disease (F-ILD) comprises a heterogeneous group of interstitial lung diseases (ILDs) characterized by parenchymal inflammation and/or fibrotic remodeling, resulting in progressive respiratory deterioration and increased mortality.^{1, 2} The Global Burden of Disease 2019 study estimated that 169,833 deaths were attributable to ILD and pulmonary sarcoidosis in 2019, with deaths and disability-adjusted life years increasing by 166.63% and 122.87%, respectively, between 1990 and 2019.³ Hospitalization in patients with F-ILD is frequently precipitated by acute worsening and is associated with increased mortality.⁴ Hospitalization is also a major contributor to the economic burden of ILD, with inpatient costs accounting for approximately 40-89% of total direct costs across studies included in a systematic review.⁵ Collectively, these findings underscore the importance of admission-based risk stratification in hospitalized patients with F-ILD.

Several prognostic tools have been developed for idiopathic pulmonary fibrosis (IPF), including the Gender-Age-Physiology (GAP) staging system, baseline and longitudinal risk models, and multidimensional indices incorporating exercise capacity.^{6, 7, 8} The ILD-GAP model extends the GAP framework to broader F-ILD populations.⁹ However, most widely used tools rely on pulmonary function and functional measures that are often unavailable or not assessed contemporaneously during hospitalization. Similarly, existing short-term risk scores have largely been developed in selected cohorts, such as patients with ILD-related acute respiratory failure or acute exacerbations requiring critical care.^{10, 11}

Routine admission laboratory tests are inexpensive and readily available. In F-ILD, blood-based biomarkers derived from routine laboratory testing have been associated with clinical deterioration and mortality.^{12, 13, 14} However, previous studies have largely evaluated individual biomarkers in isolation and outside general hospitalized cohorts, limiting their utility for bedside risk stratification at admission. Moreover, externally validated multivariable models integrating routine laboratory data to predict early clinical deterioration and both short- and long-term mortality in hospitalized patients with F-ILD remain scarce. To address this gap in inpatient risk stratification, we conducted a multicenter cohort study to develop and externally validate admission laboratory-based models for predicting mechanical ventilation (MV), 30-day and 3-month mortality, and long-term all-cause and cause-specific mortality in hospitalized patients with F-ILD.

MATERIALS AND METHODS

Study design and participants

We conducted a multicenter retrospective cohort study of hospitalized adults with F-ILD at two tertiary hospitals in China. The derivation cohort comprised patients from center 1 admitted between January 2010 and March 2025, whereas the external validation cohort included patients from center 2 admitted between May 2022 and June 2025. F-ILD was defined according to the American Thoracic Society/European Respiratory Society/Japanese Respiratory Society/Latin American Thoracic Association guideline as ILD with fibrotic changes identified on high-resolution computed tomography (HRCT).¹⁵ The inclusion criteria were as follows: (a) age ≥ 18 years and (b) availability of candidate predictors derived from the first laboratory tests performed within 24 hours of admission (baseline values). The exclusion criteria were as follows: (a) absence of HRCT findings to confirm pulmonary fibrosis; (b) initiation of invasive or non-invasive MV before baseline assessment (i.e., before admission, at admission, or within the first 24 hours after admission); (c) lung transplantation during the index hospitalization or follow-up period; and (d) active malignancy (Supplementary Methods 1). For patients with multiple hospitalizations, only the first eligible admission was included as the index admission. This retrospective study was approved by the Biomedical Ethics Review Committee of West China Hospital, Sichuan University [approval no. 2024(1847), November 19, 2024] and the Medical Ethics Committee of Sichuan Academy of Medical Sciences & Sichuan Provincial People’s Hospital [approval no. 2021(427), August 2, 2021]. The requirement for informed consent was waived by the ethics committees because of the retrospective nature of the study and the use of de-identified data.

Data collection

Candidate predictors comprised 31 prespecified variables. Demographic characteristics, comorbidities, F-ILD classification, and F-ILD treatment were obtained from inpatient medical records. Laboratory predictors were defined as the first available measurements obtained within 24 hours of admission. These variables included complete blood count parameters and derived indices; serum biochemistry measures, including alanine aminotransferase, aspartate aminotransferase, blood urea nitrogen (BUN), serum creatinine, albumin, and globulin; lactate dehydrogenase (LDH); and C-reactive protein (CRP). The neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio, prognostic nutritional index (PNI), systemic immune-inflammation index, platelet-to-white blood cell ratio (PWR), systemic inflammation response index, and albumin-to-globulin ratio (AGR) were calculated using prespecified formulas (Supplementary Methods 2). Vital status was ascertained through medical record review and telephone follow-up. Administrative censoring was applied on October 31, 2025, for center 1 and July 31, 2025, for center 2.

Outcomes and definitions

The primary outcomes were 30-day mortality, long-term all-cause mortality, and long-term cause-specific mortality. Thirty-day mortality was defined as death from any cause within 30 days of the index admission (day 1 = admission date). Long-term all-cause mortality was defined as death from any cause occurring between the index admission and the end of follow-up. Long-term cause-specific mortality was defined as F-ILD-related death, including (i) progressive respiratory failure, (ii) acute exacerbation, (iii) pneumonia or respiratory infection in the setting of advanced fibrosis, and (iv) other major ILD-related respiratory complications deemed to be the proximate cause of death. The cause of death was independently adjudicated by two investigators, and disagreements were resolved by a third reviewer. For cause-specific mortality analyses, F-ILD-related death was treated as the event of interest, whereas deaths unrelated to F-ILD were censored at the date of death. The secondary outcomes were MV and 3-month mortality. Three-month mortality was defined as death from any cause within 3 months of the index admission. MV was defined as the initiation of invasive or non-invasive MV more than 24 hours after hospital admission.

Missing-data handling and predictor selection

The derivation cohort was randomly divided into training and internal test sets in a 7:3 ratio, stratified by outcome/event status. CRP was the only candidate predictor with missing values, and missingness was limited to the derivation cohort. Missing CRP values were imputed using multiple imputation by chained equations (MICEs) fitted in the training set, with 20 imputations and predictive mean matching. The imputation model included all remaining candidate predictors and the corresponding outcome variables. For time-to-event outcomes, event status and follow-up time were included to preserve predictor-outcome associations. Missing CRP values in the internal test set were imputed using predictor information only, without incorporating outcome information from the internal test set. Model development was repeated across the imputed datasets, and performance estimates were averaged across imputations.

Within the training set, predictor selection followed a two-stage strategy consisting of Boruta screening (maxRuns = 100; TentativeRoughFix), followed by least absolute shrinkage and selection operator (LASSO) regression. For MV, 30-day mortality, and 3-month mortality, LASSO-penalized logistic regression (LR) models were fitted using one-hot encoding for categorical predictors. The penalty parameter was selected using 10-fold cross-validation to maximize the area under the curve (AUC), and λ₁se was chosen to improve model parsimony. For long-term all-cause and cause-specific mortality, predictors retained by Boruta were further reduced using LASSO-penalized Cox regression. The penalty parameter was selected using 10-fold cross-validation to maximize the concordance index (C-index), and λ₁se was chosen for parsimony. Predictors with non-zero coefficients at λ₁se were retained in the final models.

Model development, validation, and performance assessment

Models for MV, 30-day mortality, and 3-month mortality (binary outcomes) as well as long-term all-cause and cause-specific mortality (time-to-event outcomes), were developed using the training set, evaluated in the internal test set, and externally validated in an independent cohort (center 2). Models were trained using the observed outcome distribution without applying class-imbalance correction. Candidate binary classifiers included LR (unpenalized), ridge-penalized (LR, ridge), LASSO-penalized LR, support vector machine (linear and radial basis function kernels), random forest, extreme gradient boosting (XGBoost), a feed-forward neural network, and naïve Bayes. Candidate survival models included Cox proportional hazards regression (Cox), penalized Cox regression (LASSO, ridge, and elastic net), random survival forest (RSF), and gradient boosting models for survival. Model-specific preprocessing procedures were fitted within the training set and subsequently applied unchanged to the internal test set and external validation cohort. Hyperparameters were optimized using cross-validation within the training set only. Final models were then evaluated in the internal test set and external validation cohort without additional predictor selection, hyperparameter tuning, model updating, or recalibration.

For binary outcomes, discrimination was assessed using the AUC and the area under the precision-recall curve (PR-AUC). Calibration was evaluated using calibration plots and the Brier score. Sensitivity, specificity, and F1 score were calculated using a prespecified cut-off that maximized Youden’s J statistic in the training set and was subsequently applied unchanged to the internal test and external validation cohorts. For time-to-event outcomes, model performance was summarized using Harrell’s C-index and inverse probability of censoring weighting-based time-dependent AUCs at prespecified time horizons (12 and 24 months). Horizon-specific calibration was assessed at the same time points. Model interpretability and clinical utility were evaluated using Shapley additive explanations (SHAPs) and decision curve analysis (DCA), respectively. In the external validation cohort, patients were dichotomized according to the median predicted risk and compared using Kaplan-Meier curves and the log-rank test. Model discrimination in the external validation cohort was further evaluated in prespecified subgroups defined by ILD subtype, sex, and age group (< 65 vs. ≥ 65 years).

Statistical analysis

This study was reported in accordance with the TRIPOD + AI guideline.¹⁶ All analyses were performed using R (version 4.5.1) and Stata (version 18.0). Baseline characteristics were summarized by cohort. Continuous variables were assessed for normality using the Shapiro-Wilk test and are presented as mean ± standard deviation or median [interquartile range (IQR)], as appropriate based on their distribution. Categorical variables are presented as number (percentage). Differences between cohorts were quantified using standardized mean differences (SMDs). Patients lost to follow-up were treated as right-censored at the last known date they were alive. To evaluate the potential overlap between the PNI and albumin in the short-term mortality models, a sensitivity analysis was performed. The R packages and package versions used for imputation, feature selection, model development, validation, SHAP interpretation, DCA, and time-dependent AUC estimation are listed in Supplementary Table 1.

RESULTS

Patient characteristics

A total of 1,272 hospitalized adults with F-ILD were included, comprising 1,006 patients in the derivation cohort (training set, n = 705; internal test set, n = 301) and 266 patients in the external validation cohort (Supplementary Figure 1). Overall, the median age was 63.6 years (IQR, 54.0-72.0), and 53.4% of patients were men. Compared with the derivation cohort, the external validation cohort had a lower proportion of men (41.4% vs. 56.6%; SMD, 0.304) and a different distribution of F-ILD subtypes (SMD, 0.581), with connective tissue disease-associated ILD (CTD-ILD) accounting for a greater proportion of cases. Globulin (SMD, 0.640), AGR (SMD, 0.530), and LDH (SMD, 0.413) also differed between the cohorts. Glucocorticoid use was less common in the external validation cohort than in the derivation cohort (36.1% vs. 49.4%; SMD, 0.269), whereas antifibrotic therapy use was similar between the cohorts (Table 1).

TABLE 1. Baseline Characteristics of Patients with Fibrotic Interstitial Lung Disease.

Variables	Overall(n = 1272)	Derivation cohort(n = 1006)	External validation cohort(n = 266)	SMD
Demographics	-	-	-	-
Age (years), median (IQR)	63.60 (54.00-72.00)	64.00 (54.00-72.40)	61.00 (54.00-70.25)	0.106
Male, n (%)	679 (53.38)	569 (56.56)	110 (41.35)	0.304
BMI (kg/m2)	22.56 (21.09-24.08)	22.49 (21.11-23.79)	22.82 (20.70-25.43)	0.226
Smokers, n (%)	349 (27.44)	259 (25.75)	90 (33.83)	0.177
Comorbidities, n (%)	-	-	-	-
Hypertension	303 (23.82)	241 (23.96)	62 (23.31)	0.015
Diabetes	248 (19.50)	209 (20.78)	39 (14.66)	0.160
COPD	217 (17.06)	178 (17.69)	39 (14.66)	0.082
Bronchiectasis	62 (4.87)	58 (5.77)	4 (1.51)	0.227
Coronary heart disease	104 (8.18)	72 (7.16)	32 (12.03)	0.165
F-ILD subtype, n (%)	-	-	-	0.581
IPF	335 (26.34)	282 (28.03)	53 (19.92)	-
CTD-ILD	569 (44.73)	389 (38.67)	180 (67.67)	-
Other F-ILD	368 (28.93)	335 (33.30)	33 (12.41)	-
F-ILD therapy, n (%)	-	-	-	-
Antifibrotic drugs	474 (37.26)	377 (37.48)	97 (36.47)	0.021
Glucocorticoids	593 (46.62)	497 (49.40)	96 (36.09)	0.269
Immunosuppressants	159 (12.50)	115 (11.43)	44 (16.54)	0.147
None	311 (24.45)	264 (26.24)	47 (17.67)	0.207
Laboratory parameters, median (IQR)	-	-	-	-
WBC (×109/L)	7.84 (6.00-9.90)	7.96 (6.03-10.21)	7.46 (5.77-9.27)	0.182
ANC (×109/L)	5.29 (3.82-7.48)	5.28 (3.83-7.61)	5.33 (3.79-7.20)	0.037
ALC (×109/L)	1.39 (0.93-1.91)	1.43 (0.95-1.94)	1.23 (0.87-1.79)	0.244
AMC (×109/L)	0.47 (0.34-0.63)	0.48 (0.33-0.64)	0.44 (0.34-0.61)	0.022
PLT (×109/L)	190.50 (143.00-242.00)	189.00 (141.00-238.00)	198.50 (154.75-253.25)	0.138
Hemoglobin (g/L)	127.00 (115.00-140.00)	128.00 (115.00-140.00)	126.00 (114.75-138.00)	0.020
Albumin (g/L)	36.70 (33.40-40.50)	36.70 (33.40-40.50)	36.50 (33.00-39.95)	0.093
Globulin (g/L)	29.50 (25.90-34.00)	29.10 (25.30-32.90)	32.70 (28.90-36.85)	0.640
ALT (U/L)	22.00 (14.00-35.00)	21.00 (14.00-33.00)	24.00 (16.00-47.00)	0.230
AST (U/L)	24.00 (19.00-33.75)	23.00 (18.00-30.25)	29.00 (21.00-45.00)	0.291
Scr (μmol/L)	66.00 (54.00-79.00)	66.00 (54.10-79.00)	64.20 (53.08-77.12)	0.066
BUN (mmol/L)	5.30 (4.10-6.80)	5.30 (4.07-6.79)	5.75 (4.33-6.95)	0.018
LDH (U/L)	241.00 (197.00-316.00)	235.00 (193.00-296.00)	292.50 (222.00-386.25)	0.413
CRP (mg/L)	7.33 (3.07-21.90)	8.36 (3.46-24.60)	4.94 (1.99-15.13)	0.269
AGR	1.24 (1.05-1.46)	1.26 (1.09-1.50)	1.11 (0.94-1.31)	0.530
NLR	3.71 (2.39-6.63)	3.67 (2.33-6.54)	3.98 (2.59-7.06)	0.068
Laboratory parameters, median (IQR)	-	-	-	-
PLR	136.04 (92.56-206.87)	129.28 (90.06-195.53)	161.48 (111.15-229.85)	0.192
PNI	44.27 (39.05-49.10)	44.52 (39.20-49.50)	43.60 (38.77-47.53)	0.193
SII	709.14 (411.11-1378.96)	689.65 (399.95-1341.75)	820.67 (495.89-1465.08)	0.120
PWR	24.23 (17.42-32.65)	23.48 (16.81-31.97)	27.24 (20.87-35.47)	0.282
SIRI	1.77 (0.94-3.32)	1.74 (0.91-3.29)	1.93 (1.02-3.54)	0.076
Outcomes, n (%)	-	-	-	-
Mechanical ventilation	167 (13.13)	139 (13.82)	28 (10.53)	0.101
30-day mortality	109 (8.57)	95 (9.44)	14 (5.26)	0.160
3-month mortality	156 (12.26)	130 (12.92)	26 (9.77)	0.099
Long-term all-cause mortality	536 (42.14)	473 (47.02)	63 (23.68)	0.488
Long-term cause-specific mortality	480 (37.74)	425 (42.25)	55 (20.68)	0.465
Follow-up duration, months, median (IQR)	26 (11-73.5)	37.5 (10–89)	18 (12-25)	0.981

SMD, standardized mean difference; IQR, interquartile range; BMI, body mass index; COPD, chronic obstructive pulmonary disease; F-ILD, fibrotic interstitial lung disease; IPF, idiopathic pulmonary fibrosis; CTD-ILD, connective tissue disease-associated interstitial lung disease; WBC, white blood cell; ANC, absolute neutrophil count; ALC, absolute lymphocyte count; AMC, absolute monocyte count; PLT, platelet count; ALT, alanine aminotransferase; AST, aspartate aminotransferase; Scr, serum creatinine; BUN, blood urea nitrogen; LDH, lactate dehydrogenase; CRP, C-reactive protein; AGR, albumin-to-globulin ratio; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; PNI, prognostic nutritional index; SII, systemic immune–inflammation index; PWR, platelet-to-white blood cell ratio; SIRI, systemic inflammation response index.

Clinical outcomes and follow-up

MV occurred in 167 patients (13.1%), including 139 patients (13.8%) in the derivation cohort and 28 patients (10.5%) in the external validation cohort. Overall, 30-day and 3-month mortality rates were 8.6% and 12.3%, respectively (derivation vs. external cohort: 9.4% vs. 5.3% and 12.9% vs. 9.8%, respectively). During follow-up, 536 deaths (42.1%) occurred, including 480 F-ILD-related deaths (37.7%). The median follow-up duration was 37.5 months (IQR, 10-89) in the derivation cohort and 18 months (IQR, 12-25) in the external validation cohort. All-cause mortality during follow-up was 47.0% in the derivation cohort and 23.7% in the external validation cohort, whereas the corresponding rates of cause-specific mortality were 42.3% and 20.7%, respectively (Table 1).

Variable selection and final predictors

Boruta screening followed by LASSO regression identified outcome-specific admission predictors (Supplementary Figures 2, 3, 4, 5, 6). The final MV model included F-ILD subtype, white blood cell count, NLR, PNI, BUN, LDH, and CRP. The 30-day mortality model included absolute neutrophil count (ANC), NLR, albumin, PNI, LDH, and CRP, whereas the 3-month mortality model comprised ANC, NLR, albumin, PNI, and LDH. For long-term outcomes, both models retained sex, smoking status, age, F-ILD subtype, albumin, AGR, NLR, PWR, and LDH. Diabetes and BUN were additionally retained in the all-cause mortality model. The final models and the predictors retained for each outcome are summarized in Supplementary Table 2. Multicollinearity diagnostics for the retained predictors are presented as variance inflation factors in Supplementary Tables 3, 4, 5.

Model performance for short-term outcomes

Performance of all candidate algorithms in the training set, internal test set, and external validation cohort is summarized in Supplementary Tables 6, 7, 8. The final XGBoost model for MV demonstrated strong discrimination in the external validation cohort [AUC, 0.913 (95% confidence interval [CI], 0.873-0.953) PR-AUC, 0.578; Figure 1a], with good overall accuracy (Brier score, 0.0698) and calibration (Supplementary Figure 7a). At the prespecified cut-off derived from the training set (0.195), sensitivity and specificity were 0.857 and 0.824, respectively (F1 score, 0.511) (Supplementary Table 6). The final ridge-penalized LR model for 30-day mortality showed excellent external performance [AUC, 0.926 (95% CI, 0.871-0.980); PR-AUC, 0.468; Brier score, 0.0402; Figure 1b], with calibration shown in Supplementary Figure 7b. At the prespecified cut-off derived from the training set (0.120), sensitivity and specificity were 0.929 and 0.825, respectively (F1 score, 0.366) (Supplementary Table 7). External discrimination for 3-month mortality was good for the XGBoost model [AUC, 0.813 (95% CI, 0.723-0.903); PR-AUC, 0.415; Figure 1c], with acceptable accuracy (Brier score, 0.0850) and calibration shown in Supplementary Figure 7c. At the cut-off of 0.184, sensitivity and specificity were 0.731 and 0.779, respectively (F1 score, 0.388) (Supplementary Table 8). Threshold-dependent metrics with CIs and confusion matrices for the final short-term models in the external validation cohort are provided in Supplementary Tables 9, 10 and Supplementary Figure 8. Calibration assessment indicated systematic overestimation of short-term absolute risk in the external validation cohort. The observed-to-expected (O/E) ratios were 1.62 for MV, 2.00 for 30-day mortality, and 1.91 for 3-month mortality. Calibration slopes deviated from 1.0, and calibration-in-the-large intercepts were non-zero (Supplementary Table 11). In sensitivity analyses of the short-term mortality models, retaining only PNI or only albumin yielded broadly similar external performance for both 30-day and 3-month mortality (Supplementary Table 12). Discrimination in the external validation cohort was further evaluated in prespecified subgroups defined by ILD subtype, sex, and age group (Supplementary Table 13).

Figure 1.

External validation ROC curves for short-term outcomes. (a) Mechanical ventilation; (b) 30-day mortality; (c) 3-month mortality. Final models are highlighted, with AUC reported in the legend.

ROC, receiver operating characteristic; AUC, area under the curve; XGBoost, extreme gradient boosting; LASSO, least absolute shrinkage and selection operator; SVM, support vector machine; RBF, radial basis function.

Model performance for long-term outcomes

Candidate survival models are summarized in Supplementary Tables 14, 15. In external validation, the RSF (ranger) model achieved a C-index of 0.768 (95% CI, 0.709-0.827), with time-dependent AUCs of 0.811 at 12 months and 0.768 at 24 months for long-term all-cause mortality (Figure 2a, b). Brier scores were 0.1233 and 0.1508 at 12 and 24 months, respectively, with calibration shown in Supplementary Figures 7d, e. Cause-specific mortality was modeled using RSF, which yielded an external C-index of 0.768 (95% CI, 0.708-0.829). Time-dependent AUCs were 0.815 (95% CI, 0.746-0.884) at 12 months and 0.761 (95% CI, 0.676-0.847) at 24 months (Figure 2c, d). Corresponding Brier scores were 0.1082 and 0.1402 at 12 and 24 months, respectively, with calibration shown in Supplementary Figure 7f, g. Numbers at risk and observed events at 12 and 24 months in the external validation cohort are provided in Supplementary Table 16. External discrimination was further evaluated across prespecified strata defined by ILD subtype, sex, and age group (Supplementary Table 17).

Figure 2.

External validation time-dependent ROC curves for long-term outcomes. Time-dependent ROC curves in the external validation cohort at 12 and 24 months. (a) All-cause mortality at 12 months. (b) All-cause mortality at 24 months. (c) Cause-specific mortality at 12 months. (d) Cause-specific mortality at 24 months. Final models are highlighted; AUCs are shown in the legend.

ROC, receiver operating characteristic; RSF, random survival forest; AUC, area under the curve; GBM, gradient boosting machine; LASSO, least absolute shrinkage and selection operator.

Clinical utility and model interpretability

In external validation, DCA indicated that the final models provided net benefit over treat-all and treat-none strategies across clinically relevant threshold probabilities (Figure 3a-g). Decision curves for all candidate models are shown in Supplementary Figure 9a-g. The final models demonstrated potential net benefit across threshold probability ranges of 0.06-0.78 for MV, 0.03-0.51 for 30-day mortality, and 0.0-0.35 for 3-month mortality. For long-term outcomes, the corresponding ranges were 0.05-0.60 at 12 months and 0.06-0.60 at 24 months for all-cause mortality, and 0.01-0.60 at 12 months and 0.05-0.58 at 24 months for cause-specific mortality (Supplementary Table 18). SHAP feature importance identified LDH, NLR, and PNI as key predictors across MV, 30-day mortality, and 3-month mortality models (Figure 4a; Supplementary Figure 10a, b). SHAP summary plots demonstrated that elevated LDH and NLR, along with lower PNI, were associated with increased predicted risk of MV, 30-day mortality, and 3-month mortality (Figure 4d; Supplementary Figure 10c, d). For long-term outcomes, SHAP feature importance highlighted age, sex, F-ILD subtype, smoking status, LDH, and NLR as major contributors in both the all-cause and cause-specific mortality models (Figure 4b, c). SHAP summary plots further showed that predicted risk increased with older age, male sex, smoking, IPF subtype, and higher NLR, and decreased with higher albumin, AGR, and PWR (Figure 4e, f). SHAP dependence plots further characterized the directionality and potential non-linearity of individual feature effects across their observed ranges (Supplementary Figure 11a-e). Kaplan-Meier curves in the external cohort demonstrated significant separation between high- and low-risk groups for both all-cause and cause-specific mortality (log-rank p < 0.001; Supplementary Figure 12).

Figure 3.

Decision curve analysis in external validation. Decision curve analysis for the final models in the external validation cohort. Solid lines indicate model-based strategies; the dashed grey line indicates treat-all and the dashed black line indicates treat-none. (a) Mechanical ventilation (XGBoost). (b) 30-day mortality (ridge logistic regression). (c) 3-month mortality (XGBoost). (d) 12-month all-cause mortality (RSF; ranger). (e) 24-month all-cause mortality (RSF; ranger). (f) 12-month cause-specific mortality (RSF). (g) 24-month cause-specific mortality (RSF).

XGBoost, extreme gradient boosting; RSF, random survival forest.

Figure 4.

SHAP interpretation of the final models in the external validation cohort. (a-c) Global feature importance ranked by mean absolute SHAP value for 30-day mortality, long-term all-cause mortality, and long-term cause-specific mortality, respectively. (d-f) Corresponding SHAP summary plots; each point represents one patient, coloured by feature value (low to high).

LDH, lactate dehydrogenase; ANC, absolute neutrophil count; NLR, neutrophil-to-lymphocyte ratio; ALB, albumin; PNI, prognostic nutritional index; CRP, C-reactive protein; F-ILD, fibrotic interstitial lung disease; AGR, albumin-to-globulin ratio; PWR, platelet-to-white blood cell ratio; BUN, blood urea nitrogen; CTD-ILD, connective tissue disease-associated interstitial lung disease; IPF, idiopathic pulmonary fibrosis; SHAP, SHapley Additive exPlanations.

DISCUSSION

In this multicenter retrospective cohort study of 1,272 hospitalized adults with F-ILD, we developed and externally validated admission laboratory-based models to predict MV, short-term mortality (30-day and 3-month), and long-term all-cause and cause-specific mortality. First, the models demonstrated strong discrimination in external validation for short-term outcomes (AUC, 0.813-0.926) and stable long-term prognostic performance (C-index, 0.768 for both all-cause and cause-specific mortality). Second, risk patterns were clinically coherent across endpoints. Short-term risk was associated with elevated LDH and NLR as well as lower PNI, whereas long-term risk was additionally associated with older age, male sex, smoking history, and an IPF phenotype.

Existing prognostic tools in F-ILD are primarily anchored in pulmonary physiology (GAP/ILD-GAP) and HRCT-derived quantitative metrics or deep learning-based imaging features.^{6, 9, 17} Many of these tools were developed in IPF-dominant cohorts, and key input variables are often unavailable or not contemporaneously measured at hospital admission, limiting their utility for early inpatient risk stratification in broader hospitalized F-ILD populations. In external validation, our models for short-term outcomes achieved AUCs of 0.913 for MV and 0.926 and 0.813 for 30-day and 3-month mortality, respectively. Compared with prior studies in pulmonary fibrosis, our admission-based models demonstrated favorable short-term discriminatory performance. Kim et al.¹¹ developed a clinical score to predict 30-day mortality in AE-IPF, reporting an AUC of 0.81 in the validation cohort. Similarly, a prognostic index achieved an AUC of 0.7686 for 3-month mortality,¹⁸ and a multicenter retrospective study proposed a prognostic classification approach with a C-index of 0.735 for 90-day mortality in AE-IPF.¹⁹ Beyond AE-IPF, short-term risk stratification in hospitalized ILD has largely relied on context-specific tools, including generic admission severity scores (NEWS2/CURB-65; AUC, 0.80 for in-hospital mortality and 0.75 for 90-day mortality),²⁰ ICU-based mortality models (optimism-corrected C-statistic, 0.73),¹⁰ and more recent interpretable machine-learning approaches for 28-day mortality (AUC, 0.819).²¹ Collectively, these approaches often focus on mortality in selected clinical settings, whereas our externally validated models were developed for a general hospitalized F-ILD population and jointly predict early clinical deterioration (subsequent MV) and short-term mortality using routinely available admission laboratory data.

In our study, NLR, LDH, and PNI emerged as dominant predictors across short-term endpoints and interpretability analyses. Increased NLR is consistent with systemic inflammation and immune dysregulation and has been associated with worse outcomes in IPF.¹⁴ Elevated LDH may reflect greater parenchymal injury burden and hypoxia-related cellular damage.²² Lower PNI reflects reduced nutritional and immune reserve and may be associated with reduced tolerance to acute inflammation and hypoxemic stress.²³ Together, these markers support an admission “inflammation-injury-low reserve” phenotype associated with the need for ventilatory support and early mortality. Sensitivity analyses demonstrated broadly similar external performance, supporting the robustness of the short-term mortality models. However, the overlap between PNI and albumin should be considered when interpreting their individual contributions.

For long-term all-cause and cause-specific mortality, our externally validated admission laboratory-based survival models achieved C-indices of 0.768 for both outcomes, with time-dependent AUCs of approximately 0.81 at 12 months and approximately 0.76 at 24 months. By comparison, the ILD-GAP model—an extension of the GAP framework across major chronic ILD subtypes—reported a C-index of approximately 0.75 in the original cohort.⁹ Subsequent extensions incorporating comorbidity or hematologic domains yielded limited incremental improvement, with reported C-indices ranging from 0.630 to 0.662 for 3-year ILD-related event composites.^{24, 25} Beyond physiology-based scores, imaging- and machine learning-enabled approaches, such as quantitative CT metrics, have demonstrated similar discriminatory performance in IPF and progressive fibrosing ILD cohorts.^{26, 27, 28, 29} Multimodal models integrating serial imaging and clinical trajectories may achieve higher performance with longer follow-up.³⁰ However, such inputs are often not routinely available at hospital admission. Accordingly, admission laboratory-based models provide a scalable approach for early long-term risk stratification and may complement imaging-based approaches within future multimodal frameworks.

Interpretability analyses supported a clinically plausible admission-based risk profile for long-term mortality. Risk was associated with baseline vulnerability (older age, male sex, smoking history, and an IPF phenotype), elevated LDH and NLR, and lower albumin and PWR. These associations are consistent with prior evidence linking older age and male sex to worse long-term outcomes,^{6, 31} higher NLR to increased mortality,¹⁴ and poorer nutritional status or albumin-based indices to adverse outcomes in ILD/IPF.^{32, 33} In external validation, net benefit was observed across threshold probabilities beginning at 6% for MV and 3-6% for short-term mortality. In clinical practice, risks above these thresholds may prompt consideration of enhanced inpatient monitoring, early respiratory support planning, and ICU-level assessment. For long-term mortality, threshold probabilities ranged approximately from 1% to 60% across the 12- and 24-month all-cause and cause-specific mortality endpoints. Lower thresholds within this range may support closer post-discharge follow-up and multidisciplinary ILD review, whereas higher thresholds may be more relevant for advance care planning or palliative care referral. Differences between the derivation and external validation cohorts may influence model transportability and calibration. In the external validation cohort, a lower proportion of men, a lower proportion of IPF, and a higher proportion of CTD-ILD likely reflect a case mix with comparatively lower long-term mortality risk. Conversely, higher LDH, elevated globulin, and lower AGR may increase predicted risk in some patients. These opposing distributional shifts may affect calibration by creating a mismatch between predicted absolute risk and observed event rates, potentially leading to overestimation of long-term mortality risk. The shorter follow-up duration in the external cohort may further reduce the stability and precision of 24-month survival calibration and time-dependent performance estimates, particularly at later prediction horizons. Taken together, these findings suggest that routinely available admission laboratory parameters can capture long-horizon risk biology in hospitalized F-ILD and may provide useful prognostic information when contemporaneous pulmonary function testing is unavailable or not feasible.

This study has several limitations. First, the retrospective design is susceptible to selection and information bias, and residual confounding cannot be fully excluded. Pulmonary function and functional measures were not routinely available at admission; although HRCT confirmed fibrosis, detailed imaging features were not incorporated into model development. While this may limit prognostic granularity compared with physiology- or imaging-based tools, it reflects real-world inpatient practice and supports the feasibility of admission laboratory-based risk stratification. Second, treatment exposure was captured only as medication categories at admission, without detailed information on dose, timing, duration, adherence, or treatment changes during follow-up. Therefore, residual treatment-related confounding, including confounding by indication, cannot be excluded. Third, CRP missingness occurred only in the derivation cohort and may reflect its longer inclusion period and temporal changes in routine testing practices. Although MICE was used, this asymmetric missingness may have influenced predictor selection, the estimated association between CRP and outcomes, and calibration in external validation, particularly for the MV and 30-day mortality models in which CRP was retained. Fourth, follow-up duration differed substantially between centers, which may affect the stability of long-term outcome estimates, particularly calibration in the external cohort. The short-term models also showed systematic overestimation of absolute risk in external validation, with O/E ratios ranging from 1.62 to 2.00, indicating suboptimal external calibration despite strong discrimination. Predicted probabilities may therefore overestimate absolute risk when applied to independent populations. Accordingly, these models should be applied for patient-level decision support only after local recalibration, such as recalibration-in-the-large, calibration slope adjustment, or model updating using local data. In addition, the absence of class-imbalance correction for low-frequency short-term outcomes may have limited minority-class learning and increased uncertainty in threshold-dependent metrics such as sensitivity, specificity, and F1 score. Fifth, the limited number of events in the external validation cohort, particularly for 30-day mortality, may have reduced the precision of performance estimates, including calibration and threshold-dependent metrics. Sixth, ILD subtype and IPF phenotype classification relied on clinical documentation and may be subject to misclassification. In addition, initiation of MV may vary by center-level practice patterns or treatment limitations, potentially affecting transportability. Subgroup analyses were exploratory, and small event numbers in several strata limited the reliability of subgroup-specific performance estimates. Finally, although external validation was performed, both cohorts were derived from tertiary hospitals within a single national healthcare system. This may limit model transportability to other healthcare systems, non-tertiary settings, and populations with different ILD case mixes. Prospective international multicenter validation and local recalibration in diverse clinical settings are warranted before broader implementation.

In this multicenter cohort of hospitalized patients with F-ILD, we developed and externally validated admission laboratory-based models to predict MV and short- and long-term mortality. Elevated LDH and NLR, along with reduced PNI, were consistently associated with acute deterioration and early death, whereas long-term risk was additionally associated with older age, male sex, smoking history, an IPF phenotype, and lower AGR and PWR. These admission-based models demonstrated strong discrimination for predicting MV and mortality in hospitalized patients with F-ILD. Although short-term risk was systematically overestimated in external validation, the models may support early inpatient triage and risk-informed post-discharge surveillance. Local recalibration and further validation in diverse clinical settings are required before routine implementation.

Supplementary Materials

Supplementary Figures

https://balkanmedicaljournal.org/img/files/supplement-1.pdf

Supplementary Methods

https://balkanmedicaljournal.org/img/files/method-yazi.pdf

Supplementary Tables

https://balkanmedicaljournal.org/img/files/35c3b395-1452-49f0-8b97-0671fb2a2f27.pdf