The prognostic value of intraoperative HRV during anesthesia in patients presenting for non-cardiac surgery

Abstract

Abstract Objective To examine the prognostic value of HRV measurements during anesthesia for postoperative clinical outcomes prediction using machine learning models. Data sources VitalDB, a comprehensive database of 6388 surgical patients admitted to Seoul National University Hospital. Eligibility criteria for study selection Cases with ECG lead II recording duration of less than one hour were excluded. Cases with more than 20% of missing HRV measurements were also excluded. A total of 5641 cases were eligible for the analyses. Methods Six machine learning models including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Decision Trees (GBT), Extreme Gradient Boosting (XGB), and an ensemble of the five baseline models were developed to predict postoperative clinical outcomes. The prediction models were trained using only clinical information, and using both clinical information and HRV features, respectively. Feature importance based on the SHAP method was used to assess the contribution of the HRV measurements to the outcome predictions. Subgroup analysis was also performed to evaluate the risk association between postoperative ICU stay and various HRV measurements such as heart rate, low-frequency power (LFP), and short-term fluctuation DFA $${\alpha }_{1}$$ α 1 . Result The final cohort included 5641 unique cases, among whom 4678 (83.0%) cases had ages over 40, 2877 (51.0%) were male, 1073 (19.0%) stayed in ICU after surgery, 52 (0.9%) suffered in-hospital death, and 3167(56.1%) had a total length of hospital stay longer than 7 days. In the final test set, the highest AUROC performance with only clinical information was 0.79 for postoperative ICU stay, 0.58 for in-hospital mortality, and 0.76 for the total length of hospital stay prediction. Importantly, using both clinical information and HRV features, the AUROC performance was 0.83, 0.70, and 0.76 for the three clinical outcome predictions, respectively. Subgroup analysis found that patients with an average heart rate higher than 70, low-frequency power (LFP) < 33, and short-term fluctuation DFA $${\alpha }_{1}$$ α 1 < 0.95 during anesthesia, had a significantly higher risk of entering the ICU after surgery. Conclusion This study suggested that HRV measurements during anesthesia are feasible and effective for predicting postoperative clinical outcomes.

Keywords

• Anesthesia
• HRV
• Machine learning
• Surgery
• Clinical outcome