Breaking new ground: machine learning enhances survival forecasts in hypercapnic respiratory failure

Area of Science:

• Medical Informatics
• Biostatistics
• Pulmonology

Background:

• Prognostic prediction for hypercapnic respiratory failure (HRF) is clinically significant.
• Accurate survival prediction aids in patient management and treatment strategies for HRF.

Purpose of the Study:

• To develop and validate a predictive model for survival in patients with HRF.
• To compare the performance of the random survival forest (RSF) model against established algorithms.

Main Methods:

• A cohort of 697 HRF patients was used, split into modeling (n=565) and external validation (n=132) groups.
• Three models were evaluated: random survival forest (RSF), DeepSurv, and Cox Proportional Risk (CoxPH).
• Performance metrics included C-index, Brier score, ROC curves, AUC, and decision curve analysis (DCA).

Main Results:

• The RSF model achieved a higher C-index (0.792) compared to CoxPH (0.699) and DeepSurv (0.618).
• RSF demonstrated superior performance with a Brier Score consistently below 0.25 across 6-24 months.
• ROC and DCA confirmed RSF's superior discrimination and clinical utility in both patient cohorts.

Conclusions:

• The RSF model significantly outperforms CoxPH and DeepSurv for predicting prognosis in HRF patients.
• RSF offers enhanced capabilities for clinical evaluation and patient monitoring in hypercapnic respiratory failure.