JMIR Publications

ABSTRACT

Background:

Adequate ventilation in mechanically ventilated patients during general anesthesia is contingent upon the monitoring of the arterial partial pressure of carbon dioxide (PaCO2). Despite its significance, continuous monitoring remains challenging due to the limitations of intermittent invasive measurements, such as arterial blood gas analysis (ABGA), and the imprecision of estimates and often unpredictable gradient between PaCO2 and ETCO2 may lead to misinterpretation of a patient’s ventilator status, potentially resulting in adverse outcomes.

Objective:

This study aimed to develop a machine learning model to continuously estimate PaCO2 in mechanically ventilated patients using a comprehensive set of readily available noninvasive parameters, thereby improving intraoperative monitoring accuracy under general anesthesia.

Methods:

This retrospective study used the VitalDB dataset, a public database from Seoul National University Hospital, containing records from 6,388 noncardiac surgery patients between August 2016 and June 2017. After applying inclusion and exclusion criteria, data from 2,304 surgical cases, yielding 4,651 PaCO2 measurement event points, were included in this analysis. The CatBoostRegressor model was employed to predict PaCO2. A total of 19 noninvasive features were used, comprising intraoperative vital signs such as ETCO2, body temperature, the SpO2/FiO2 ratio, respiratory rate, and airway pressures, along with preoperative clinical information including age, gender, and pulmonary function test results. The model’s performance was evaluated using a nested cross-validation scheme to ensure robust and generalizable results. Performance was assessed across hypocapnic (<35 mmHg), normocapnic (35-45 mmHg), and hypercapnic (>45 mmHg) subgroups and compared to two conventional baseline methods: a simple offset (ETCO2 + 5 mmHg) and linear regression with ETCOw as the sole predictor.

Results:

The developed model demonstrated superior overall performance compared to both traditional estimation methods. It achieved a mean absolute error (MAE) of 2.38 mmHg and a root mean squared error (RMSE) of 3.26 mmHg. The model showed excellent agreement with actual PaCO2 measurements, with an average intraclass correlation coefficient (ICC) of 0.87 (95% CI: 0.86-0.87). In terms of clinical utility, 90.02% of the model’s estimations fell within the highly acceptable range of ±5 mmHg error, a substantial improvement from the 80.43% achieved by the linear regression baseline. Furthermore, clinically unacceptable errors (> ±10 mmHg) were reduced to 1.20%, less than half the rate of the baseline model (2.64%). These performance improvements were consistently observed across all PaCO2 subgroups, including the more challenging hypocapnic and hypercapnic ranges.

Conclusions:

The developed machine learning-based model provides more accurate and reliable estimates of PaCO2 than traditional ETCO2-based methods. This approach shows potential for enhancing continuous respiratory monitoring, facilitating timely and precise clinical interventions, and serving as a valuable supplementary tool for anesthetic management. Further validation, including prospective studies to assess its impact on clinical decision-making and patient outcomes, is necessary to fully realize its clinical integration.