Objective:

To develop a multimodal machine learning framework to predict in-hospital type 2 respiratory failure (T2RF) risk during acute exacerbations of chronic obstructive pulmonary disease (AECOPD) with temporal–geographic external validation.

Approach:

• Study Design: A two-source design was employed with a development cohort of 6,954 AECOPD patients from a single tertiary hospital, randomly divided into training (n = 4,867) and internal test (n = 2,087) sets, and a temporal external validation cohort of 1,252 patients from seven hospitals.

Key Findings:

• XGBoost achieved an AUROC of 0.660 (95% CI: 0.631–0.689) in the internal test set and 0.699 (95% CI: 0.661–0.738) in the external validation set.
• LightGBM performed comparably with an AUROC of 0.700.
• Key predictors included lymphocyte count, eosinophil count, COPD duration, RDW-CV, age, hypertension, and sex.
• SHAP analysis identified low lymphocyte count and long COPD duration as dominant risk drivers.
• The logistic nomogram achieved an external AUROC of 0.666.

Interpretation:

The multimodal machine learning framework enables early T2RF risk stratification at admission using routine blood counts and demographics.

Limitations:

• The study is retrospective and may be subject to biases inherent in such designs.
• Spirometry data was excluded, which may limit the model's applicability in certain clinical contexts.
• The performance of the model may vary across different populations and settings.

Conclusion:

Future work should integrate dynamic monitoring and prospective multicenter validation.