A Multimodal Stacked Ensemble Model for Cardiac Output Prediction Utilizing Cardiorespiratory Interactions during General Anesthesia

doi:10.21203/rs.3.rs-2990056/v1

Download PDF

Article

A Multimodal Stacked Ensemble Model for Cardiac Output Prediction Utilizing Cardiorespiratory Interactions during General Anesthesia

https://doi.org/10.21203/rs.3.rs-2990056/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 29 Mar, 2024

Read the published version in Scientific Reports →

You are reading this latest preprint version

Background:
This study aimed to examine the possibility of estimating cardiac output (CO) using a multimodal stacking model that utilizes cardiopulmonary interactions during general anesthesia.
Materials and methods:
This study included 469 adult patients who underwent normal pulmonary function tests under general anesthesia. We used the publicly available VitalDB, a multiparameter biosignal prospectively collected database containing the data of non-cardiac surgical patients from Seoul National University Hospital.

The hemodynamic data in this study included non-invasive blood pressure, plethysmographic heart rate, and SpO₂. CO was recorded using the pulse contour technique devices Vigileo and EV1000. Respiratory data included mechanical ventilation parameters and end-tidal CO₂ levels.

A generalized linear regression model was used as the base learner for the multimodal stacking ensemble method. Random forest, generalized linear regression, gradient boosting machine, and XGBoost are meta-learners that use base learner predictions. Multimodal stacking ensembles combine predictions from two modalities, hemodynamic and respiratory numerical data obtained from multimodal intraoperative monitoring.
Results:
A Bland–Altman plot revealed that a multimodal stacked ensemble model for CO prediction from 327 patients had a bias of −0.001 L/min (±1.96 SD, 0.611 L/min, and −0.614 L/min) and −0.271% (±1.96 SD, 12.94%; SD, −13.488%) when calculating the percentage of difference using the EV1000 device.

Agreement of model CO prediction and measured Vigileo CO in 142 patients reported a bias of −0.01 (±1.96 SD, 0.464 L/min, and −0.477 L/min) and −0.333% (±1.96 SD, 9.924%, and −10.59%).

Conclusions:

This study demonstrated that our model predicts CO compared to data obtained by the pulse contour technique cardiac output monitors with good agreement.

Biological sciences/Biotechnology

Biological sciences/Computational biology and bioinformatics

Biological sciences/Physiology

Health sciences/Cardiology

Health sciences/Medical research

Mechanical ventilation (MV) has a predictable impact on circulation [1][2].

Cardiorespiratory interactions are clinically important because MV can lead to cardiac instability [3]. MV typically uses positive airway pressure, thereby increasing intrathoracic pressure (ITP), causing a reduction in venous return, and increasing pulmonary vascular resistance, which can decrease preload and subsequently reduce cardiac output (CO). However, increased ITP can cause a decrease in the afterload on the heart, leading to increased stroke volume and CO.

In addition, excessive tidal volume and lung hyperinflation caused by overstimulation of sensory nerve endings located within the alveolar walls can lead to reflex bradycardia and depression of the somatic nervous system [4].

Variations in arterial pulse and systolic pressure in mechanically ventilated patients with adjusted tidal volumes can predict fluid responsiveness during acute circulatory failure related to sepsis [5].

A decrease in CO₂ concentration at the end-tidal concentration (EtCO₂) in humans and animals correlates with a reduction in pulmonary blood flow/CO [6][7]. This relationship is significant and is currently implemented in anesthesia monitoring for non-invasive and minimally invasive breath-by-breath CO monitoring in patients ventilated during anesthesia and critical care [8].

Presently, various methods for monitoring cardiovascular systems are available, including non-invasive, minimally invasive, and invasive techniques for measuring CO. Among these, thermodilution (TD) is considered the gold standard method, and pulse contour analysis is widely used [9][10].

A non-calibrated pulse pressure analysis device has been demonstrated to be clinically and statistically acceptable under hypo- and normodynamic conditions[11]. Notably, considerable overestimation and underestimation of hemodynamic parameters have been reported when using a peripherally inserted central catheter [12].

Moreover, more than a dozen non-invasive methods have been proposed and developed to estimate CO. The simplest of these methods involves calculating CO by multiplying the stroke volume (SV) by the heart rate (HR), where SV is obtained by multiplying the pulse pressure (systolic blood pressure (SBP)–diastolic blood pressure (DBP)) by a constant value (k = 2). This method has been evaluated and observed to have a moderate correlation between the measured and estimated CO (r = 0.60, p < .001) [13]. Furthermore, machine learning algorithms have been employed in animal models to predict CO accurately by utilizing waveform arterial blood pressure and HR, with a difference of − 0.13 (0.69 L/min) between the standard and predicted CO [14].

Despite real-time ventilation monitoring, non-invasive standard measures, such as blood pressure, HR, electrocardiogram, and oxygen saturation, provide limited insights into CO. Recent research suggests that estimating CO using cardiorespiratory interactions is possible; however, it requires precise physiological data integrated with a clinical decision support system. Therefore, an accurate and non-invasive CO estimation method to optimize intraoperative hemodynamics during general anesthesia is required. To address this challenge, we proposed a multimodal stacking scheme for CO prediction using cardiorespiratory interactions during general anesthesia. Reportedly, this is the first study to use this approach.

Data source

In this study, de-identified data were used from an open database of non-cardiac surgery patients who underwent routine or emergency operations at Seoul National University Hospital, Seoul, Korea, from August 2016 to June 2017 [15]. This database contains prospectively collected with the formal approval of an ethics committee/IRB (H-1408-101-605) and registered at clinicaltrials.gov (NCT02914444) intraoperative vital sign data from 6,388 general, thoracic, urological, and gynecological surgery cases. However, perioperative clinical information was retrospectively obtained. In addition, several anesthesia devices recorded up to 12 waveforms and 184 numeric data tracks during surgery using the Vital Recorder program.

Monitoring parameters and data structure

Hemodynamic data, obtained as numeric values at 2-s intervals, included non-invasive blood pressure (NIBP), plethysmographic HR, and SpO₂ data (Solar™ 8000 M, GE healthcare, Wauwatosa, WI, USA). In addition, CO was recorded using pulse contour technique monitors such as Vigileo and EV1000 (Edwards Lifesciences, Irvine, CA, USA).

Respiratory data collected using the anesthesia machine (Primus, Dräger, Lübeck, Germany) were recorded at 7-s intervals. MV was determined by estimating the fraction of inspired oxygen (FiO₂), expiratory TV, expiratory minute volume (Vm), positive end-expiratory pressure (PEEP), peak inspiratory pressure (PIP), respiratory rate based on capnography, and infrared spectrometry capnography, which measures EtCO₂), thereby ensuring adequate and accurate ventilation per period.

The data were structured in nonconstant time intervals [15]. Consequently, the irregular multivariate data were synchronized at intervals of 14 s, and CO marked the target parameter. Subsequently, the rows with missing values were deleted to ensure the development of a robust model.

In addition to describing the parameters, descriptive statistics were used to describe them in terms of minimum, maximum, mean, standard deviation, median, 25th–75th quartiles, and 95% confidence intervals (see Supplemental Tables S2 and S3).

Model development

For multisystem CO prediction, we used multimodal stacking-based ensemble learning regression techniques (Fig. 1). We split the multimodal data into training, validation, and test sets at a ratio of 7.5:1.25:1.25.

After preprocessing, the training and validation data were used to train and validate the model using both hemodynamic and respiratory parameters (NIBP, HR, MV, SpO₂, and EtCO₂), to calculate the relationship between cardiorespiratory interactions and CO.

Further training and validation data were separated into hemodynamic (NIBP and HR subsets) and respiratory parameters (MV, SpO₂, and EtCO₂ subsets). Three data subsets were constructed using the demographic variables (age, height, weight, and sex).

Four base learner models were used for the multimodal stacking ensemble in this study (Level-0): a generalized linear model (GLM) [16], Random Forest (RF)[17], Gradient Boosting Machine (GBM) [18], and extreme gradient boosting (XGBoost) [19]. Excluding GLM, all the models were nonlinear regressions.

A grid search method was used in the H2O R package [20], with the h2o.grid function adjusted to optimize the parameters of the base models and obtain stable and accurate predictions. To optimize the hyperparameters of the GLM, distributed RF, GBM, and XGBoost regression models in the three data subsets, a random search was conducted by splitting the training set 5-fold to optimize the hyperparameters and enhance the model performance by lowering the predicted value error measured through the mean absolute error (MAE). The hyperparameters of the base models are summarized in Supplementary Table 4.

The best optimized regression predictions of the 12 base models from the three data subsets were subsequently used as input features for the multimodal stacking ensemble method (Level-1) [15]. Using the h2o.stackedEnsemble function of the H2O package, an automatic meta-learner algorithm determined the best regression parameters in favor of a GLM.

The comparison results of the base models and multimodal stacked assembly with the performance metrics in the regression are listed in Table 1.

Performance metrics in regression

The mean square error (MSE), root mean square error (RMSE), MAE, and coefficient of determination (R²) were used as performance indicators to evaluate the regression algorithms. The MSE and RMSE are commonly used regression metrics.

$$RMSE=\sqrt{\frac{1}{N}}\sum _{i = 1}^{N}{({y}_{i}-{y}_{i}^{p})}^{2}$$

By taking the square root of MSE, the RMSE measures the difference between the predicted CO ${y}_{i}^{p}$and measured CO values ${y}_{i}$. The RMSE was calculated as the square root of the sum of all regression errors per row divided by the total number of observations. The regression performance improved, with lower RMSE and MSE values.

$$MAE=\frac{1}{N}\sum _{i = 1}^{N}\left[{y}_{i}-{y}_{i}^{p}\right]$$

The MAE measures the difference between the measured CO ${y}_{i}$ and predicted CO ${y}_{i}^{p}$ values divided by the total number of observations. Low MAEs indicate high model accuracy.

$${R}^{2}=1-\frac{\sum _{i=1}^{N}{\left({y}_{i}-{y}_{i}^{p}\right)}^{2}}{\sum _{i=1}^{N}{\left({y}_{i}-\overline{y}\right)}^{2}}$$

R² represents how well a regression model explains the variability between the measured and predicted CO. From 0 to 1, the higher the value, the better is the model.

R² represents the variability explained by the model (squared difference between the target ${y}_{i}$ and the predicted value${y}_{i}^{p}$) divided by the total deviance y value (squared difference between the target ${y}_{i}$ and the mean target values $\overline{y}$).

Model evaluation

We used the Bland–Altman method as a statistical standard to compare the measures of CO ${y}_{i}$ from both EV 1000 and Vigileo with the multimodal stacking ensemble model prediction CO ${y}_{i}^{p}$[21]. The Bland–Altman plot was introduced to describe the agreement, where the y-axis shows the difference between the measured and model-predicted CO (${y}_{i}-{y}_{i}^{p}$), and the x-axis represents the average of the measured and model-predicted CO ((${y}_{i}$+${y}_{i}^{p}$)/2). In summary, the absolute mean difference between the measured and predicted CO ($\overline{d}=\frac{1}{N}\sum _{i=1}^{N}{(y}_{i}-{y}_{i}^{p}$) can be used to estimate the constant bias, and the limits of agreement (LoAs) lie between $\overline{d}-$1.96S_d and $\overline{d}$+1.96S_d, where S_d is the standard deviation. Percentage difference plots and analyses were used to assess the proportional differences between the measured and model-predicted CO. This shows how this error influences lower CO measurements, whereas for higher CO values, the percentage bias is decreased [22]. In the proportional Bland–Altman plots, bias changes over the measuring range, and are presented as a proportional mean difference$\stackrel{̿}{d }$= $\frac{1}{N}\sum _{i = 1}^{N}\frac{({y}_{i} - {y}_{i}^{p})}{({y}_{i}+{y}_{i}^{p})/2 }$, and the LoA lay between $\stackrel{̿}{d}-$1.96S_d and $\stackrel{̿}{d}$+1.96S_d.

Among the 6388 patients in the VitalDB Dataset, which measured hemodynamics including CO from the Vigileo and EV1000 devices and respiratory monitoring data, 722 were eligible for the study (Fig. 1). Patients < 18 years old were excluded (n = 8). Data were selected from the beginning to the end of surgery to ensure hands-free, automatic, and constant ventilation. Additionally, the absence of lung disease in patients undergoing general anesthesia could be achieved by selecting patients with normal pulmonary function testing from their clinical information (n = 110). Participants with unsynchronized parameter measurements and an absence of intraoperative NIBP measurements were excluded (n = 135). After data preprocessing, 469 (EV1000, n = 327; Vigileo, n = 142) patients met the data requirements and covered 1732.4 h (EV1000, 1288.03 h; Vigileo 444.1 h) of 49007 pairs of row measurement data. A summary of the patients and their clinical characteristics is presented in Supplementary Tables S1–3.

3.1 Results of the base and multimodal stacked ensemble model regression

We evaluated the individual base learners and multimodal stacking ensemble regressions on three data subsets. We first trained and validated the hemodynamic response to cardiopulmonary interactions during MV using parameters from the hemodynamic and respiratory subsets (NIBP, HR, MV, SpO₂, and EtCO₂).

Second, the hemodynamic data subsets (NIBP and HR) were used for the training and validation of the base models to predict CO from arterial blood pressure and heart rate.

Third, respiratory data subsets (MV, EtCO₂, and SpO₂) were used for the training and validation of the base models and to calculate the hemodynamic effects of MV.

Table 1

presents the best results of the four base models, which were derived from the regression performance metrics of GLM, RF, GBM, XGBoost, and multimodal stacked ensemble models. For both CO monitoring devices, multimodal stacking outperformed all its base models in terms of the MSE, RMSE, and MAE, measuring 0.096, 0.31, and 0.186 for EV1000 and 0.057, 0.239, and 0.139 for Vigileo, respectively. In addition, based on MAE, average errors were evaluated. Compared to base models, RMSE was more sensitive to significant errors and multimodal stacking models predicted CO more accurately.
Best Model	Data	MSE	RMSE	R2	MAE
Multimodal Stacked Ensemble	NIBP/HR/MV/SpO2/EtCO2/Target CO-EV1000	0.096	0.31	0.97	0.186
RF	NIBP/HR/MV/SpO2/EtCO2/Target CO-EV1000	0.12	0.346	0.968	0.197
XGBoost	NIBP/HR/MV/SpO2/EtCO2/Target CO-EV1000	0.123	0.35	0.966	0.211
XGBoost	NIBP/HR/ Target CO-EV1000	0.154	0.393	0.956	0.232
RF	NIBP/HR/ Target CO-EV1000	0.148	0.385	0.955	0.241
GBM	NIBP/HR/ Target CO-EV1000	0.166	0.408	0.953	0.256
GBM	NIBP/HR/ MV/SpO2/EtCO2/Target CO-EV1000	0.182	0.426	0.95	0.275
GBM	MV/SpO2/EtCO2/Target CO-EV1000	0.181	0.425	0.948	0.27
RF	MV/SpO2/EtCO2/Target CO-EV1000	0.192	0.438	0.948	0.262
XGBoost	MV/SpO2/EtCO2/Target CO-EV1000	0.251	0.501	0.927	0.343
GLM	NIBP/HR/MV/SpO2/EtCO2/Target CO-EV1000	1.81	1.345	0.444	1.064
GLM	NIBP/HR/ Target CO-EV1000	1.97	1.403	0.395	1.105
GLM	MV/SpO2/EtCO2/Target CO-EV1000	2.105	1.451	0.353	1.149
Multimodal Stacked Ensemble	NIBP/HR/MV/SpO2/EtCO2/Target CO-Vigileo	0.057	0.239	0.974	0.139
RF	NIBP/HR/MV/SpO2/EtCO2/Target CO-Vigileo	0.074	0.273	0.973	0.157
GBM	NIBP/HR/MV/SpO2/EtCO2/Target CO-Vigileo	0.078	0.28	0.972	0.164
XGBoost	NIBP/HR/MV/SpO2/EtCO2/Target CO-Vigileo	0.078	0.28	0.972	0.163
GBM	NIBP/HR/ Target CO-Vigileo	0.101	0.318	0.963	0.186
XGBoost	NIBP/HR/ Target CO-Vigileo	0.108	0.329	0.962	0.188
RF	NIBP/HR/ Target CO-Vigileo	0.1	0.316	0.959	0.198
RF	MV/SpO2/EtCO2/Target CO-Vigileo	0.115	0.34	0.956	0.2
GBM	MV/SpO2/EtCO2/Target CO-Vigileo	0.116	0.34	0.954	0.206
XGBoost	MV/SpO2/EtCO2/Target CO-Vigileo	0.132	0.363	0.951	0.219
GLM	NIBP/HR/MV/SpO2/EtCO2/Target CO-Vigileo	1.072	1.035	0.524	0.787
GLM	NIBP/HR/ Target CO-Vigileo	1.14	1.068	0.494	0.826
GLM	MV/SpO2/EtCO2/Target CO-Vigileo	1.599	1.264	0.291	0.981

Table 1 The performance of the base and multimodal stacked ensemble models. NIBP, non-invasive blood pressure; HR, heart rate; MV, mechanical ventilation; SpO₂, Oxygen saturation; EtCO₂, end-tidal CO₂. Target CO measurements from the EV1000 and Vigileo monitoring devices. Machine learning algorithms: Generalized Linear Model (GLM), Random Forest (RF), Gradient Boosting Machine (GBM), and extreme gradient boosting (XGBoost).

3.2 Multimodal stacked ensemble model CO prediction vs. measured CO from the arterial waveform analysis device

Figure 2 and Table 2 present the baseline agreement between the multimodal stacked ensemble model for CO prediction and the CO measured using the EV1000 device. The difference between the methods was r = 0.985, and R² was 0.97 (Fig. 2a). Bland–Altman analysis revealed that the mean difference between measurements and prediction was − 0.001 L/min (± 1.96 SD, 0.611, and − 0.614 L/min; Fig. 2b). The proportional mean difference was − 0.271% (± 1.96 SD, 12.94%, and − 13.488%; Fig. 2c).

The agreement between the multimodal stacked ensemble model prediction of CO and the CO measured using Vigileo was r = 0.987, and R² was 0.974 (Fig. 2d). The overall mean bias for agreement in CO was − 0.01 (± 1.96 SD, 0.464, and − 0.477 L/min; Fig. 2e). The proportional mean difference was − 0.333% (± 1.96 SD, 9.924%, and − 10.59%; Fig. 2f).

Table 2

Bland–Altman analysis. Lower and upper limits of agreement (LLoA and ULoA) and standard deviation (SD)
			CI at 95%
	Bland-Altman Analysis	Mean Value	Lower CI	Upper CI
Agreement between Multimodal Stacked Ensemble Model CO prediction and measured EV1000 CO.	Mean differences (Bias) (L/min)	-0.001	-0.01	0.007
	SD of Bias(L/min)	0.312
	ULoA (+ 1.96 SD) (L/min)	0.611	0.595	0.626
	LLoA (-1.96 SD) (L/min)	-0.614	-0.629	-0.598
	Percent differences (Bias) (%)	-0.271	-0.466	-0.076
	SD of Bias (%)	6.743
	ULoA (+ 1.96 SD) (%)	12.94	12.608	13.284
	LLoA (-1.96 SD) (%)	-13.488	-13.826	-13.15
Agreement between Multimodal Stacked Ensemble Model CO prediction and measured Vigileo CO.	Mean differences (Bias) (L/min)	-0.01	-0.018	0.005
	SD of Bias(L/min)	0.24
	ULoA (+ 1.96 SD) (L/min)	0.464	0.443	0.485
	LLoA (-1.96 SD) (L/min)	-0.477	-0.497	-0.456
	Percent differences (Bias) (%)	-0.333	-0.810	-0.188
	SD of Bias (%)	5.233
	ULoA (+ 1.96 SD) (%)	9.924	9.474	10.373
	LLoA (-1.96 SD) (%)	-10.59	-11.04	-10.141

In this study, we proposed a multimodal stacking ensemble that combines data from non-invasive cardiovascular monitoring and MV parameters, including SpO₂ and EtCO₂. A fundamental principle of the proposed model is that stacking makes the prediction accuracy better than that of a single machine learning algorithm, and stacking several algorithms significantly improves the prediction accuracy. We demonstrate that the multimodal stacked ensemble model predicts accurate and valid CO values with marginal bias and a narrow CO limit of agreement compared with those obtained using pulse contour technique devices.

When using a clinical CO measurement device, Bland–Altman plots do not indicate whether the LoAs are acceptable [24]. For example, an agreement limitation of ± 1 L/min may not be acceptable in patients with low CO syndrome. Additionally, our results include percentage difference plots demonstrating that multimodal stacked ensemble models accurately predict CO, with predictions falling within the acceptable clinical criteria (± 30%) of the proportional mean difference when compared those obtained using the Vigileo and EV1000 devices.

In previous studies, the calibrated pulse wave analysis device EV1000, which was used for our reference CO measurement, has proven to be accurate and consistent. The results showed good agreement and interchangeability with TD CO measurement, with a bias of − 0.07 L/min, LoA of 2.0 L/min, and a percentage of 29% [25]. In addition, the uncalibrated FloTrac/Vigileo provides clinically acceptable accuracy under stable hemodynamic conditions, with an average error below 30% for CO compared with that obtained via TD [11]. However, severe sepsis and septic shock uncalibrated FloTrac/Vigileo vs. TD reveals no clinically acceptable tracking capability with a bias of − 0.86 L/min, LoA of − 4.48 to 2.77 L/min, and a percentage error of 48% [26].

Our study was based entirely on non-cardiac surgery. Accordingly, NIBP was selected because it is a standard measurement for patients with ASA I and II and for intermediate-risk surgery. In addition, NIBP appears to be in acceptable agreement with invasively measured BP in patients with cardiogenic shock [27], MV, and arrhythmia [28]. However, NIBP is not always well calibrated with invasive BP measurement, particularly in hypothermia and pronounced hypotension [27]. Although invasive BP, known as beat-by-beat measurement, is considered the gold standard method of diagnosis, NIBP is associated with fewer complications, particularly catheter-associated artery pseudoaneurysms, occlusions, and infections [29]. Occasionally, a measurement can be inaccurate owing to kinking or damping of the arterial line.

The HR was extracted from finger photoplethysmography and may represent acceptable accuracy based on electrocardiography (ECG) during normal breathing. Photoplethysmography and ECG-derived heart rates can differ moderately, and photoplethysmography shows an advantage in monitoring changes in ITP caused by ventilation, sleep apnea, and even changes in respiratory rate during deep breathing [30][31].

Using the respiratory rate based on capnography, the expiratory tidal volume, and the expiratory Vm enabled us to obtain the exact delivered volume per breathing cycle recorded in the anesthesia machine (Fig. 3n,o,p). Noteworthy differences between the set and delivered tidal volumes have been demonstrated in several clinical situations, such as patient lung size, lung compliance, airway resistance, and maintenance of spontaneous breathing during general anesthesia through invasively assisted spontaneous ventilation [32][33].

Figure 3. a. The two-way interaction strengths (Vint) between variables are represented by connecting lines in the random forest model. The stronger the interaction, the thicker and darker the indigo line. Both the size of the node and the intensity of green indicate the importance of the variable (Vimp). b-p. Partial dependence plots for variables in the multimodal stacking ensemble model. Partial Dependence Multimodel Plot gives a graphical depiction of the distributed random forest (DRF), gradient boosting machine (GBM), generalized linear model (GLM), and extreme gradient boosting (XGBoost). The effect of a variable is measured as the change in the mean cardiac output. NIBP–SBP and NIBP-DBP, systolic and diastolic non-invasive blood pressure; HR, plethysmographic heart rate; FiO₂, fraction of inspired oxygen; TV, expiratory tidal volume; Vm, expiratory minute volume; PEEP, positive end-expiratory pressure; PIP, peak inspiratory pressure; EtCO₂, infrared spectrometry capnography, which measures end-tidal CO₂.

4.1 Visualizing cardiopulmonary interactions and variable importance in a multimodal stacked ensemble model

Providing decision support using a functional hemodynamic machine learning model based on the complex relationship between the heart and lungs during general anesthesia should be understood by the medical environment. The predictability of the model was quantified in our work using partial dependence plots (PDPs) [20], model parameter’ importance, and interaction variables [34].

Figure 3a shows a network graph displaying the importance of training data variables and interactions fitted with a RF. Consequently, this plot helps identify cardiopulmonary interactions affecting CO prediction. The network plot shows a strong interaction between HR, NIBP–SBP, age, weight, and height, represented by thick and intense purple lines. The most robust relationship exists between HR and weight. Furthermore, strong interactions between HR and EtCO₂ were observed. HR and age are the most significant predictor and is represented by a large, intensely green node.

The results of our study were consistent with well-established data demonstrating that CO levels decrease with age by approximately 1% per year after the third decade (Fig. 3b). Age-related decline in the stroke index is accompanied by decreased body size and HR, which reduces CO [35]. We found the exact relationship between body size and CO in a straight-line regression, as observed in the last century [36] (Fig. 3d,e). According to our findings, in females, one-way PDPs from the RF, GBM, and XGBoost models showed a decrease in CO of approximately 10% compared to those in males during intraoperative measurements. However, this difference was smaller than the 22% difference reported during the resting state [37] (Fig. 3c).

The HR is crucial in determining the diastolic filling time, influencing the SV via the Frank–Starling mechanism. For cardiopulmonary interactions during MV, venous return can be reduced, which can further compromise diastolic filling, particularly at high heart rates. Our study revealed a linear relationship between CO and HR up to 90/min, where deceleration began (Fig. 3f). Early curve deceleration is well documented in impaired right heart filling [38]. However, here, this may have been influenced by factors, such as autonomic nervous system activity, blood volume, and heart contractility, which were beyond the scope of this study.

The relationships between SBP, DBP, and CO during general anesthesia are complex and dynamic.

In our study, we observed an increase in SBP corresponding to an increase in CO of up to 120 mmHg following the onset of the deceleration curve (Fig. 3g). The decreased CO level during high intraoperative SBP may be caused by increased vascular resistance, stiffened large arteries [39], and reduced SV owing to elevated afterload. Our study demonstrated a decrease in DBP with a marginal increase in CO (Fig. 3h). An increase in pulse pressure might elucidate the observed increase in CO. An increase in SV owing to volume substitution results in increased CO, causing an increase in pulse pressure. Cardiopulmonary interactions and additional interventions such as vasopressor administration or adjustments to ventilator settings may play a substantial role. Additionally, the nonlinear relationship between the pulse pressure, cardiac index (CI), and deceleration curve starting at a CI of 3 L/min/m² has been well documented [40].

One-way PDPs revealed an inverse relationship between CO and airway pressure (Fig. 3l). A decrease in SV and venous return is the primary mechanism by which increasing airway pressure reduces CO. The application of airway pressure levels at 10, 20, and 30 cm H₂O led to a variation in the CI between + 6% and − 43%, which was associated with corresponding changes in the SV index (p < 0001, r2 = 0.89) [41]. Our findings align with those of earlier studies, as they indicated an increase in airway pressure during lung inflation and a reduction in CO at a rate of 0.5 L/min per 10 mbar increase in PIP.

PEEP increases ITP during the entire respiratory cycle to restore normal end-expiratory lung volume during MV. Increasing the PEEP levels allowed for greater lung expansion. PEEP during MV may also displace blood from the pulmonary circulation, increase mean systemic pressure, reduce venous return, and decrease CO and tissue perfusion [42]. Our model exhibits a decrease in the CO rate of 0.1 L/min by raising PEEP to 2.5 mbar (Fig. 3m). This decrease in CO with increasing PEEP in a curvilinear relationship has been previously reported [43].

A reduction in TV increases CO; however, the degree of improvement in hemodynamics depends largely on ITP [44]. Reducing the tidal volume increases chest wall compliance by decreasing ITP during MV and increasing venous return, leading to increased left ventricular preload and CO. This is consistent with our finding; our model showed an increase in CO of 0.03 L/min per 1 mL/kg of TV reduction (Fig. 3o). A tidal volume > 15 mL/kg markedly decreases HR and blood pressure and reduces CO [45]. However, we could not evaluate this observation with limited training data for tidal volumes > 15 mL/kg, and a machine learning model could not make meaningful predictions.

Changes in exhaled carbon dioxide during general anesthesia with stable ventilation correspond to changes in CO and metabolic CO₂ production [46]. At ETCO₂ levels > 30 mmHg, RF, GBM, and XGBoost models predict a satisfactory CO increase of 0.5 L/min per 10 mmHg of ETCO₂ (Fig. 3k). A similar correlation between ETCO₂ and CO has been reported in previous studies [47]. An animal model during cardiopulmonary resuscitation showed a correlation coefficient of 0.79 between EtCO₂ and CI. [48]. This finding is consistent with that of the GLM model. However, the GLM model had a lower performance than that of the RF, GBM, and XGBoost models and had fewer training data with EtCO₂ < 30 mmHg.

A decline in SpO₂ was observed with decreasing CO in all base models in our study (Fig. 3j). Decreased CO caused by cardiopulmonary interactions is the primary factor in the reduction of arterial oxygen content observed during MV [49]. Hypovolemia and vasodilation, which are commonly observed during general anesthesia, may also contribute to this phenomenon. However, our data did not allow us to determine whether the increased inspired O₂ fraction reflected an increase in CO (Fig. 3i). It is widely recognized that increases in FiO₂ at fixed values of CO fail to detect conditions of low oxygen supply during central venous oxygen saturation [50].

This study has some limitations. First, we analyzed data from one source alone and focused solely on adult patients. During data mining, we could not find synchronized records of sudden blood loss or vasoactive infusions. However, the perioperative clinical information dataset contained information on the estimated intraoperative blood loss and vasoactive medications, which were not included in this study. In addition, the small number of patients with obstructive or restrictive lung diseases made it difficult to include them in the data subset. Although constant ventilation was ensured during surgery in this study, it is important to recognize that the period from the onset of anesthesia to the start of surgery and the time between the end of surgery and extubation are important for comprehending the influence of cardiopulmonary interactions on hemodynamics. During extubation or weaning, spontaneous inspiratory efforts in patients with obstructive and restrictive lung disease may strongly decrease CO by increasing the left ventricular afterload, especially if left ventricular function is already impaired [44]. Our model should be improved in the future to address these cardiopulmonary interactions.

This study introduced a multimodal stacking ensemble algorithm for predicting CO using cardiorespiratory interactions during general anesthesia. Numerical data were used to train a stacking ensemble model that involved two-component regression based on hemodynamic and respiratory monitoring inputs. Pulse contour technique CO monitors were used to obtain data for comparison with the CO predicted using the model. The results showed a bias of − 0.271% (± 1.96 SD, 12.94%, and − 13.488%) from the EV1000 device and a bias of − 0.333% (± 1.96 SD, 9.924%, and − 10.59%) from the Vigileo device. This novel approach could help meet the increasing demand for non-invasive CO measurements and improve intraoperative hemodynamic management in ventilated patients without chronic obstructive or restrictive lung disease.

Data availability

The datasets generated and/or analysed during the current study are available in the VitalDB repository, https://physionet.org/content/vitaldb/1.0.0/

Appendix A. Supplementary data

The R-based interactive web applications for this study can be accessed at the address provided below: https://albiondervishi.shinyapps.io/CO_Shiny/

N. Karamolegkos, A. Albanese, og N. W. Chbat, “Heart-lung interactions during mechanical ventilation: Analysis via a cardiopulmonary simulation model”, IEEE Open J. Eng. Med. Biol., b. 2, 324–341, 2021.
C. Ngo, R. Schlözer, T. Vollmer, S. Winter, B. Misgeld, og S. Leonhardt, “A simulative model approach of cardiopulmonary interaction”, í IFMBE Proc., b. 51, 1679–1682, 2015.
A. Vieillard-Baron o.fl., “Acute cor pulmonale in acute respiratory distress syndrome submitted to protective ventilation: Incidence, clinical implications, and prognosis”, Crit. Care Med., b. 29, tbl. 8, bls. 1551–1555, 2001.
J. T. Shepherd, “The lungs as receptor sites for cardiovascular regulation”, Circulation, b. 63, tbl. 1., 1–10, 1981.
F. Michard o.fl., “Relation between respiratory changes in arterial pulse pressure and fluid responsiveness in septic patients with acute circulatory failure”, Am. J. Respir. Crit. Care Med., b. 162, tbl. 1, 134–138, 2000.
M. S. Siobal, “Monitoring exhaled carbon dioxide”, Respir. Care, b. 61, tbl. 10, 1397–1416, 2016.
M. J. Harrison, R. Scott-Weekly, og M. Zacharias, “The qualitative detection of decreases in cardiac output”, Comput. Biol. Med., b. 58, bls. 85–90, mar. 2015.
P. J. Peyton, “Continuous minimally invasive peri-operative monitoring of cardiac output by pulmonary capnotracking: Comparison with thermodilution and transesophageal echocardiography”, J. Clin. Monit. Comput., b. 26, tbl. 2, 121–132, 2012.
X. Monnet o.fl., “Comparison of pulse contour analysis by Pulsioflex and Vigileo to measure and track changes of cardiac output in critically ill patients”, Br. J. Anaesth., b. 114, tbl. 2, 235–243, 2015.
N. Kiefer o.fl., “Clinical validation of a new thermodilution system for the assessment of cardiac output and volumetric parameters”, Crit. Care, b. 16, tbl. 3, 2012.
C. Slagt, I. Malagon, og A. B. J. Groeneveld, “Systematic review of uncalibrated arterial pressure waveform analysis to determine cardiac output and stroke volume variation”, Br. J. Anaesth, b. 112, tbl. 4. 626–637, 2014.
S. D’Arrigo o.fl., “Are peripherally inserted central catheters suitable for cardiac output assessment with transpulmonary thermodilution?”, Crit. Care Med., b. 47, tbl. 10, bls. 1356–1361, okt. 2019.
L. B. K. Hill, J. J. Sollers, og J. F. Thayer, “Evaluation of a simple estimation method for the derivation of cardiac output from arterial blood pressure and heart rate”, Biomed. Sci. Instrum., b. 48, 165–170, 2011.
N. T. Liu, G. C. Kramer, M. N. Khan, M. P. Kinsky, og J. Salinas, “Blood pressure and heart rate from the arterial blood pressure waveform can reliably estimate cardiac output in a conscious sheep model of multiple hemorrhages and resuscitation using computer machine learning approaches”, J. Trauma Acute Care Surg., b. 79, tbl. 4, S85–S92, 2015.
H.-C. Lee, Y. Park, S. Bin Yoon, S. M. Yang, D. Park, og C.-W. Jung, “VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients”, Sci. Data, b. 9, tbl. 1, bls. 279, 2022.
N. E. Breslow, “Generalized linear models : Checking assumptions and strengthening conclusions”, Transformation, b. 8, 23–41, 1996.
P. Geurts, D. Ernst, og L. Wehenkel, “Extremely randomized trees”, Mach. Learn., b. 63, tbl. 1, 3–42, 2006.
J. H. Friedman, “Greedy function approximation: A gradient boosting machine”, Ann. Stat., b. 29, tbl. 5, 1189–1232, 2001.
T. Chen og C. Guestrin, “XGBoost: A scalable tree boosting system”, Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, 785–794, b. 13-17-August-2016.
E. LeDell o.fl., “Package ‘h2o’”, April, 2020.
J. M. Bland og D. G. Altman, “Measuring agreement in method comparison studies”, Stat. Methods Med. Res., b. 8, tbl. 2., 135–160, 1999.
D. Giavarina, “Understanding Bland Altman analysis”, Biochem. Medica, b. 25, tbl. 2, 141–151, 2015.
B. Ripley, B. Venables, D. M. Bates, K. Hornik, A. Gebhardt, og D. Firth, “Package ‘MASS’ (Version 7.3–51.4)”, Cran-R Proj., 7–3, 2019.
P. M. Odor, S. Bampoe, og M. Cecconi, “Cardiac output monitoring: Validation studies–how results should be presented”, Curr. Anesthesiol. Rep., b. 7, tbl. 4, 410–415, 2017.
K. Bendjelid o.fl., “Performance of a new pulse contour method for continuous cardiac output monitoring: validation in critically ill patients.”, Br. J. Anaesth., b. 111, tbl. 4, bls. 573–579, okt. 2013.
C. Slagt, M. Helmi, I. Malagon, og A. B. J. Groeneveld, “Calibrated versus uncalibrated arterial pressure waveform analysis in monitoring cardiac output with transpulmonary thermodilution in patients with severe sepsis and septic shock: An observational study”, Eur. J. Anaesthesiol., b. 32, tbl. 1, 5–12, 2015.
J. Seidlerová, P. Tůmová, R. Rokyta, og M. Hromadka, “Factors influencing the accuracy of non-invasive blood pressure measurements in patients admitted for cardiogenic shock”, BMC Cardiovasc. Disord., b. 19, tbl. 1, 609–616, 2019.
K. Lakhal o.fl., “Blood pressure monitoring during arrhythmia: Agreement between automated brachial cuff and intra-arterial measurements”, Br. J. Anaesth., b. 115, tbl. 4, 540–549, 2015.
R. Tosti, S. Özkan, R. M. Schainfeld, og K. R. Eberlin, “Radial artery pseudoaneurysm”, J. Hand Surg. Am., b. 42, tbl. 4, 295e1–295e6, 2017.
A. H. Khandoker, C. K. Karmakar, og M. Palaniswami, “Comparison of pulse rate variability with heart rate variability during obstructive sleep apnea”, Med. Eng. Phys., b. 33, tbl. 2, 204–209, 2011.
A. Schäfer og J. Vagedes, “How accurate is pulse rate variability as an estimate of heart rate variability?: A review on studies comparing photoplethysmographic technology with an electrocardiogram”, Int. J. Cardiol., b. 166, tbl. 1, 15–29, 2013.
Y. Yamaguchi o.fl., “The difference between set and delivered tidal volume: A lung simulation study”, Med. Devices Evid. Res., b. 13, bls. 205–211, 2020.
Z. Ruszkai og Z. Szabó, “Maintaining spontaneous ventilation during surgery—a review article”, J. Emerg. Crit. Care Med., b. 4, 2020.
A. Inglis, A. Parnell, og C. B. Hurley, “Visualizing variable importance and variable interaction effects in machine learning models”, J. Comput. Graph. Stat., b. 31, tbl. 3, 766–778, 2022.
M. Brandfonbrener, M. Landownel, og N. W. Shock, “Changes in cardiac output with age.”, Circulation, b. 12, tbl. 4, 557–566, 1955.
W. Jegier, P. Sekelj, P. A. Auld, R. Simpson, og M. Mcgregor, “The relation between cardiac output and body size.”, Br. Heart J., b. 25, bls. 425–430, 1963.
K. Forton, Y. Motoji, S. Caravita, V. Faoro, og R. Naeije, “Exercise stress echocardiography of the pulmonary circulation and right ventricular-arterial coupling in healthy adolescents”, Eur. Heart J. Cardiovasc. Imaging, b. 22, tbl. 6, 1071–1078, 2021.
T. Sugimoto, K. Sagawa, og A. C. Guyton, “Effect of tachycardia on cardiac output during normal and increased venous return.”, Am. J. Physiol., b. 211, tbl. 2, 288–292, 1966.
A. M. Dart og B. A. Kingwell, “Pulse pressure - A review of mechanisms and clinical relevance”, J. Am. Coll. Cardiol., b. 37, tbl. 4, 975–984, 2001.
C. J. Petrie, K. Damman, P. S. Jhund, H. L. Hillege, D. J. Van Veldhuisen, og A. A. Voors, “Low pulse pressure as a poor-man’s indicator of a low cardiac index in patients with severe cardiac dysfunction”, J. Cardiovasc. Med., b. 15, tbl. 4, 315–321, 2014.
H. Jellinek, P. Krafft, R. D. Fitzgerald, S. Schwarz, og M. R. Pinsky, “Right atrial pressure predicts hemodynamic response to apneic positive airway pressure”, Crit. Care Med., b. 28, tbl. 3, 672–678, 2000.
B. T. Kuhn, L. A. Bradley, T. M. Dempsey, A. C. Puro, og J. Y. Adams, “Management of mechanical ventilation in decompensated heart failure”, J. Cardiovasc. Dev. Dis., b. 3, tbl. 4. MDPI, 01-des-2016.
J. F. Dhainaut, J. Y. Devaux, J. F. Monsallier, F. Brunet, D. Villemant, og M. F. Huyghebaert, “Mechanisms of decreased left ventricular preload during continuous positive pressure ventilation in ARDS”, Chest, b. 90, tbl. 1, 74–80, 1986.
M. R. Pinsky, “Cardiopulmonary interactions: Physiologic basis and clinical applications”, Ann. Am. Thorac. Soc., b. 15, S45–S48, 2018.
I. M. Cheifetz o.fl., “Increasing tidal volumes and pulmonary overdistention adversely affect pulmonary vascular mechanics and cardiac output in a pediatric swine model”, Crit. Care Med., b. 26, tbl. 4, 710–716, 1998.
M. I. Monge García, A. G. Cano, M. G. Romero, R. M. Pintado, V. P. Madueño, og J. C. Díaz Monrové, “Non-invasive assessment of fluid responsiveness by changes in partial end-tidal CO 2 pressure during a passive leg-raising maneuver”, Ann. Intensive Care, b. 2, tbl. 1, bls. 2–9, 1–10, 2012.
A. S. Baraka, M. T. Aouad, M. I. Jalbout, R. N. Kaddoum, M. F. Khatib, og S. T. Haroun-Bizri, “End-tidal CO2 for prediction of cardiac output following weaning from cardiopulmonary bypass”, J. Extra. Corpor. Technol., b. 36, tbl. 3, bls. 255–257, sep. 2004.
M. H. Weil, J. Bisera, R. P. Trevino, og E. C. Rackow, “Cardiac output and end-tidal carbon dioxide”, Crit. Care Med., b. 13, tbl. 11, 907–909, 1985.
G. R. Kelman, J. F. Nunn, C. Prys-roberts, og R. Greenbaum, “The influence of cardiac output on arterial oxygenation: A theoretical study”, Br. J. Anaesth., b. 39, tbl. 6, 450–458, 1967.
F. G. Zampieri, M. Park, L. C. P. Azevedo, M. B. P. Amato, og E. L. V. Costa, “Effects of arterial oxygen tension and cardiac output on venous saturation: A mathematical modeling approach”, Clinics, b. 67, tbl. 8, bls. 897–900, ágú. 2012.

No competing interests reported.

Download PDF

Journal Publication

published 29 Mar, 2024

Read the published version in Scientific Reports →

Editorial decision: Revision requested
02 Jan, 2024
Reviews received at journal
30 Dec, 2023
Reviewers agreed at journal
13 Dec, 2023
Reviews received at journal
06 Nov, 2023
Reviewers agreed at journal
27 Oct, 2023
Reviews received at journal
03 Sep, 2023
Reviewers agreed at journal
24 Aug, 2023
Reviewers invited by journal
18 Aug, 2023
Editor assigned by journal
18 Aug, 2023
Editor invited by journal
09 Jun, 2023
Submission checks completed at journal
09 Jun, 2023
First submitted to journal
27 May, 2023

You are reading this latest preprint version

A Multimodal Stacked Ensemble Model for Cardiac Output Prediction Utilizing Cardiorespiratory Interactions during General Anesthesia

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Methods

Data source

Monitoring parameters and data structure

Model development

Performance metrics in regression

Model evaluation

Results

3.1 Results of the base and multimodal stacked ensemble model regression

Discussion

4.1 Visualizing cardiopulmonary interactions and variable importance in a multimodal stacked ensemble model

Conclusion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1