Explainable machine learning for postoperative respiratory failure prediction in open-heart surgery patients — a study based on the MIMIC-IV database

Comprehensive Summary

This article intended to create a model that not only predicts postoperative respiratory failure (PRF) in the first 24 hours of ICU admission, but also explains why the patient is at risk. They used 4,488 patients from the MIMIC-IV database who underwent open heart surgery with cardiopulmonary bypass. The researchers also used 8 different machine learning models in the test, and implemented SHAP (Shapley additive exPlanations) to rank feature importance and visualize how individual clinical variables pushed a patient’s risk score up or down. They found that the winning model here was the Gradient Boosting Machine (GBM), achieving an AUROC of 0.808, with a sensitivity of 0.703, and a specificity of 0.776. They found the most significant factors driving PRF risk were minimum ionized calcium levels (sharp increase in risk after levels fell below 1.0 mmol/L). Next was the vasopressor score, indicating high requirements for blood pressure support increased risk for PRF. They also found that lower values of ventral venous oxygen saturation were also linked to higher risk. And finally, elevated BUN increased risk as well.

Outcomes and Implications

This article helps form a golden risk stratification form data on the first 24 hours of ICU stay. This can provide a great actionable window for pre-emptive adjustments in respiratory protocols for especially high risk patients, possibly improving outcomes and decreasing the need for expensive last second respiratory measures. This study also helps indicate a specific ionized calcium threshold for providers, being that hypocalcemia might be a specific modifiable way to reduce PRF risk. Using AI to help predict a patient's level of risk can help providers more clearly tailor an approach given the specific markers showcased in this article.