PHRF-Net: An Interpretable Deep Learning Approach for Early Prediction of Hypoxemic Respiratory Failure After Cardiac Surgery

Postoperative respiratory complications are a major source of morbidity and mortality following cardiac surgery. Early and accurate prediction is crucial for targeted interventions, but existing models often face challenges in performance or interpretability. To address this, we developed a deep learning model, the Postoperative Hypoxemic Respiratory Failure Network(PHRF-Net), for predicting postoperative hypoxemic respiratory failure using data from 7,875 cardiac surgery patients. The PHRF-Net architecture leverages a self-attention mechanism to provide not only accurate predictions but also inherent interpretability, offering transparent insights into the key drivers of its decisions. The proposed PHRF-Net model demonstrated robust performance, achieving an Area Under the ROC Curve (AUC) of 0.816 and an F1-score of 0.900. Analysis of the model’s attention weights revealed several influential predictors. By integrating a clinically-informed feature structure with a self-attention mechanism, this work provides a framework for building AI tools that are not only accurate but also offer clinically relevant insights into the decision-making process.