What is the clinical evidence from this study?

Study design: Observational. Population: Pulmonary fibrosis (n=534). Intervention: LightGBM machine learning model vs. Conventional clinical scores (OASIS, SOFA, APS III, SIRS, SAPS II). Primary outcome: Prediction of 28-day all-cause mortality (AUC).

October 14, 2025BMC Medical Informatics and Decision Making1 citationsOpen Access

Machine learning model development and validation using SHAP: predicting 28-day mortality risk in pulmonary fibrosis patients

ZWZijun WuBeijing Information Science & Technology University MLMingliang LiGuangdong Medical College ZXZhiliang XuGuangdong Medical College

Key Result

The LightGBM machine learning model predicted 28-day mortality risk in pulmonary fibrosis patients with an AUC of 0.819, outperforming conventional clinical scores.

Study Design

Type

Observational (n=534)

Multicenter

Structured PICO

Does a machine learning model (LightGBM) improve the prediction of 28-day mortality risk in ICU patients with pulmonary fibrosis compared to traditional clinical scores?

Population

534 patients with pulmonary fibrosis aged 18-100 years on their first admission to the ICU, extracted from the MIMIC-IV database. Mean age 75.5 years, 55.8% male.

Intervention

Machine learning predictive models (specifically LightGBM) using clinical variables to predict 28-day mortality risk.

Comparator

Traditional clinical scoring systems (SOFA, APS III, SIRS, SAPS II, OASIS) and other machine learning algorithms (decision tree, KNN, MLP, SVM, XGBoost).

Outcome

All-cause mortality within 28 days of admission.hard clinical

A LightGBM machine learning model accurately predicts 28-day mortality in ICU patients with pulmonary fibrosis, outperforming traditional clinical scores, with ICU stay duration identified as the most critical prognostic feature.

Main Result

Absolute Event Rate: 0.819% vs 0.697%

Limitations

Case identification relied exclusively on administrative data without radiographic or pathological confirmation
Generalizability may be constrained by the single-center dataset
Residual unmeasured confounding may persist despite adjusting for multiple confounders

Abstract

BACKGROUND: Early prediction of mortality risk within 28 days of admission is crucial for personalized treatment in patients with pulmonary fibrosis (PF). This study aims to develop a predictive model for 28-day mortality risk in PF patients using interpretable machine learning (ML) methods. METHODS: Data from patients with pulmonary fibrosis were extracted from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. The study endpoint was mortality within 28 days of admission. Feature selection was performed using logistic regression and LASSO algorithms. Six machine learning algorithms-decision tree, k-nearest neighbors (KNN), LightGBM, single-hidden-layer neural network, support vector machine (SVM), and extreme gradient boosting (XGBoost)-were employed to construct risk prediction models. Additionally, SHapley Additive exPlanations (SHAP) were utilized to interpret the predictive models. RESULTS: Among the six evaluated machine learning models, the LightGBM model demonstrated robust predictive performance, with an area under the receiver operating characteristic curve (AUC) of 0.819. SHAP analysis revealed that length of ICU stay, respiratory rate, and white blood cell count were the three most important features for predicting 28-day mortality risk in PF patients, with ICU stay duration having the most significant impact. CONCLUSION: This study indicates that machine learning methods hold potential for early prediction of mortality risk within 28 days of admission in patients with pulmonary fibrosis. Moreover, SHAP analysis enhanced the interpretability of the LightGBM model, thereby guiding clinical decision-making.

问 AI

Bookmark

View Full Paper