Machine learning-based models predict postoperative cardiovascular and neurological complications after pneumonectomy: A 10-year retrospective observational study

doi:10.21203/rs.3.rs-4164223/v1

Download PDF

Research Article

Machine learning-based models predict postoperative cardiovascular and neurological complications after pneumonectomy: A 10-year retrospective observational study

https://doi.org/10.21203/rs.3.rs-4164223/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Reducing postoperative cardiovascular and neurological complications (PCNC) in thoracic surgery is key for improving postoperative survival. Therefore, we aimed to investigate the independent predictors of PCNC, develop machine learning models, and construct a predictive nomogram for PCNC in patients undergoing thoracic surgery for lung cancer.

Methods

This study used data from a previous retrospective study of 16,368 lung cancer patients with American Standards Association physical status I-IV who underwent surgery. Postoperative information was collected from electronic medical records; the optimal model was analyzed and filtered using multiple machine learning models (Logistic regression, eXtreme Gradient Boosting, Random Forest, Light Gradient Boosting Machine, and Naïve Bayes). The predictive nomogram was built, and the efficacy, accuracy, discriminatory power, and clinical validity were assessed using receiver operator characteristics, calibration curves, and decision curve analysis.

Results

Multivariate logistic regression analysis showed that age, duration of surgery, intraoperative intercostal nerve block, postoperative patient-controlled analgesia, bronchial blocker, and sufentanil were independent predictors of PCNC. Random forest was identified as the optimal model with an area under the curve of 0.898 in the training set and 0.752 in the validation set, confirming the excellent prediction accuracy of the nomogram. All the net benefits of five machine learning models in the training and validation sets demonstrated excellent clinical applicability, and calibration curves also showed good agreement between the predicted and observed risks.

Conclusion

The combination of machine learning models and nomograms may contribute to the early prediction and reduction of the incidence of PCNC.

Lung cancer

nomogram

machine learning

postoperative cardiovascular and neurological complications

thoracic surgery

video-assisted thoracoscopic surgery

Cardiovascular and neurological complications following thoracic surgery are critical postoperative concerns that seriously affect patient prognosis, quality of life, and hospital stay duration and endanger patient lives ^[1]. Existing studies have demonstrated that major cardiovascular and neurological adverse events, including death, acute ischemic stroke, or myocardial infarction, occur, with an incidence rate of 0.8–3% ^{[2, 3]}. Nevertheless, minor postoperative cardiovascular and neurological complications (PCNC), such as postoperative atrial fibrillation and transient cerebral ischemia, are independent risk factors for severe postoperative cardiovascular and neurological adverse events and death ^[4]. Therefore, postoperative patients with cardiovascular and neurological complications should be closely monitored regardless of severity.

Video-assisted thoracoscopic surgery (VATS) lobectomy has proven to correlate with reduced pain, minimized complications, and improved life quality without affecting oncologic outcomes in patients with lung cancer in previous studies ^{[5, 6]}. However, based on the data obtained from our hospital’s information system, postoperative cardiovascular and neurological complications after VATS have an incidence of approximately 20% ^[7–10], which is not lower than the incidence of open surgery. Furthermore, data from a European research group indicated that a longer duration of operation correlates with an increased incidence of complications in the postoperative period ^[11]. The operating time and anesthetic techniques may influence the outcome of cardiovascular and neurological complications.

Reducing the occurrence of PCNC in thoracic surgery remains a key issue in improving the quality of postoperative survival. Recently, multiple machine learning models such as random forest, eXtreme Gradient Boosting (XGboost), Light Gradient Boosting Machine ( LightGBM), and Naive Bayers can be used to construct information models based on big data and effectively reveal the potential relationship between multiple clinical predictors and outcomes, and the nomogram has been increasingly familiar to clinical doctors as an intuitive and concise predictive method ^{[12, 13]}. Hence, this study aimed to develop a predictive model for predicting PCNC.

Patients

The data of patients who have undergone thoracic surgery for lung tumors at the A and B Divisions of the Hospital between September 2012 and September 2022 were reviewed. Patients at the Zhongshan Division comprised the training set, while patients at the Hunnan Division comprised the validation set. The study samples and treatment data were obtained from the hospital’s electronic information systems. Patients who (1) had been pathologically diagnosed with malignant lung cancer and (2) underwent lung tumor resection were enrolled in the research. In contrast, patients (1) in whom lung cancer was not the only primary malignancy, (2) those with insufficient medical record data, and (3) those who had a previous history of cardiovascular and neurological disease were excluded (since this study focused on patients with new-onset PCNC, which are often overlooked preoperatively, increasing the likelihood of medical disputes once new-onset postoperative complications occur). The PCNC were described based on evidence shown on postoperative electrocardiogram (ECG), computed tomography (CT) scan, or cranial magnetic resonance imaging (MRI); findings following a consultation with a cardiologist or neurologist; or a diagnosis of new-onset cardiovascular or neurological-related diseases including (1) cardiac-related: atrial flutter, atrial fibrillation, supraventricular tachycardia, rapid ventricular arrhythmias, sustained ventricular tachycardia, coronary artery disease, congestive heart failure; percutaneous coronary intervention or coronary artery bypass grafting, obstructive coronary artery disease, myocardial ischaemia (presence or absence of clinical symptoms of ECG abnormalities) or myocardial infarction (suppression of tumorigenicity [ST] segment elevation, non-ST segment elevation, or troponin elevation); or (2) neurological-related: transient ischaemic attack, stroke, cerebral embolism, ischaemic/ haemorrhagic cerebral infarction, or cerebral artery stenosis (intracranial arteries or carotid arteries) during the postoperative period. The Ethics committee of the First Hospital of China Medical University approved this study (No: 2022 − 449, November 4, 2022) and did not contain personal identifiers; Therefore, there was no requirement to obtain individual patient consent. Trial registration: Chinese Clinical Trial Registry, 2023.08.01, No. ChiCTR2300074188 (https://www.chictr.org.cn/showproj.html?proj=200744), Principal investigator's name is Wen-fei Tan. For this study, the principles of the Declaration of Helsinki were followed.

Data collection

Gender, age, body mass index, history of smoking, alcohol consumption, chemotherapy, previous surgery, and disease history were collected from the Hospital's electronic medical record system. Data on the time of surgery, operative technique, interoperative intercostal nerve block used, postoperative patient-controlled analgesia (PCA), method of intubation, and dosage of anesthetic drugs were obtained from the electronic records of the surgical anesthesiologists.

Statistical analysis

Categorical variables were expressed using counts and percentages, and the Pearson chi-square (χ2) test or Fisher exact test was used for comparison of data between groups. Continuous variables were expressed as mean ± standard deviation, and data between groups was compared by Student's t-test or Mann–Whitney U-test.

Logistic regression analysis was performed to predict the incidence of PCNC. The significant candidate predictors and factors related to the dose of anesthetic drugs were selected using univariate logistic regression analysis and enrolled into multivariate logistic regression analysis. All these candidate predictors were checked for collinearity using the variance inflation factor (VIF). These characteristics are expressed as odds ratios (ORs) and 95% confidence intervals (CIs). The p-value of < 0.05 for two-tailed was regarded as statistically significant.

Multiple machine-learning models, including Logistic regression, eXtreme Gradient Boosting (XGBoost), RandomForest, Light Gradient Boosting Machine (LightGBM), and Naïve Bayes, were built using R software. The models were repeated 10 samples for training and validating to analyze the importance of the training and validation set metrics, and optimal models were selected.

The performance of the predictive nomogram was validated in both training and test sets, and calibration curves were used to assess the calibration accuracy of the predictive model. Calibration plots with 1,000 bootstrap weights were performed to assess the model's calibration accuracy. The nomogram’s specific performance was quantified by calculating the area under the curve (AUC) value of the receiver operator characteristics (ROC) curve and by calculating the specificity, sensitivity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). Determination of nomogram clinical utility and net benefit was then performed using decision curve analysis (DCA). The x-axis of the decision curve was the threshold for the predicted probability. Net benefit to patients of clinical decisions based on the classification of outcomes in this threshold is shown on the y-axis. To perform all statistical analyses, SPSS version 25 (IBM Corporation, Armonk, NY, USA) and R software version 4.3.1 (http://www.rproject.org) were used.

Patient characteristics

Between September 2012 and September 2022, 16,368 patients who underwent thoracic surgery for lung tumors at the Hospital's Department of Thoracic Surgery were enrolled in this study (Table 1). A total of 3,528 patients with missing electronic medical records and 2,106 patients with previous histories of cardiovascular or neurological disease prior to surgery were excluded. Of the included patients, 3,398 developed postoperative cardiovascular and neurological complications. The age of patients in the PCNC group (64.10 ± 10.63) was remarkably high compared to the non-PCNC group (59.84 ± 11.60); the same trend was observed regarding the history of smoking and alcohol consumption. Additionally, the average duration of surgery was also noticeably higher in the patients in the PCNC group (136.36 ± 73.06) when compared with that in the non-PCNC group (116.28 ± 63.19); there were also statistically considerable differences between the two groups regarding PCA and anesthetic drug dose of sufentanil. The 11,458 patients from Hospital Division A made up the training set, while 4,910 patients from Hospital Division B made up the validation set (Fig. 1). All indicators were distributed equally between the two groups (Table 2).

Table 1

Characteristics of patients with postoperative cardiovascular and neurological complications (PCNC)
Characteristics	No PCNC patients (n = 12970)	PCNC patients (n = 3398)	P value
Gender			0.164
Male	6819 (52.6)	1832 (53.9)
Female	6151 (47.4)	1566 (46.1)
Age, years	59.84 ± 11.60	64.10 ± 10.63	< 0.001
BMI (kg/m²)			0.225
≤ 22.9	8463 (65.3)	2255 (66.4)
> 22.9	4507 (34.7)	1143 (33.6)
History of smoking			0.007
Yes	4080 (31.5)	1151 (33.9)
No	8890 (68.5)	2247 (66.1)
History of alcohol			< 0.001
Yes	2401 (18.5)	739 (21.7)
No	10569 (81.5)	2659 (78.3)
History of chemotherapy			0.998
Yes	1496 (11.5)	392 (11.5)
No	11474 (88.5)	3006 (88.5)
History of surgery			0.807
Yes	5657 (43.6)	1490 (43.8)
No	7313 (56.4)	1908 (56.2)
Duration of surgery, min	116.28 ± 63.19	136.36 ± 73.06	< 0.001
Operative technique			0.322
Thoracotomy	4941 (38.1)	1263 (37.2)
VATS	8029 (61.9)	2135 (62.8)
Intraoperative intercostal nerve block			< 0.001
Yes	2714 (20.9)	120 (3.5)
No	10256 (79.1)	3278 (96.5)
Type of PCA			< 0.001
None	5920 (45.6)	1754 (51.6)
Epidural PCA	3869 (29.8)	998 (29.4)
Intravenous PCA	3181 (24.5)	646 (19.0)
Method of intubation			< 0.001
Double-lumen endobronchial tubes	10863 (83.8)	3126 (92.0)
Bronchial blockers	2107 (16.2)	272 (8.0)
Dose of Sufentanil, µg	33.48 ± 13.67	33.71 ± 12.79	0.375
Abbreviation: PCNC: postoperative cardiovascular and neurological complications; BMI: body mass index; VATS: video-assisted thoracoscopic surgery; PCA: patient-controlled analgesia

Table 2

Characteristics of patients with postoperative cardiovascular and neurological complications (PCNC) in the training and validation cohort
Characteristics	Training cohort (n = 11458)	Validation (n = 4910)	P value
Gender			0.635
Male	6042 (52.7)	2609 (53.1)
Female	5416 (47.3)	2301 (46.9)
Age, years	60.70 ± 11.63	60.80 ± 11.31	0.621
BMI (kg/m²)			0.637
≤ 22.9	7516 (65.6)	3202 (65.2)
> 22.9	3942 (34.4)	1708 (34.8)
History of smoking			0.425
Yes	3640 (31.8)	1591 (32.4)
No	7818 (68.2)	3319 (67.6)
History of alcohol			0.056
Yes	2154 (18.8)	986 (20.1)
No	9304 (81.2)	3924 (79.9)
History of chemotherapy			0.403
Yes	1306 (11.4)	582 (11.9)
No	10152 (88.6)	4328 (88.1)
History of surgery			0.893
Yes	5007(43.7)	2140 (43.6)
No	6451 (56.3)	2770 (56.4)
Duration of surgery, min	120.40 ± 65.75	120.58 ± 66.14	0.868
Operative technique			0.859
Thoracotomy	4348 (37.9)	1856 (37.8)
VATS	7110 (62.1)	3054 (62.2)
Intraoperative intercostal nerve block			0.189
Yes	2013 (17.6)	821 (16.7)
No	9445 (82.4)	4089 (83.3)
Type of PCA			0.684
None	5394 (47.1)	2280 (46.4)
Epidural PCA	3404 (29.7)	1463(29.8)
Intravenous PCA	2660 (23.2)	1167 (23.8)
Method of intubation			0.748
Double-lumen endobronchial tubes	9786 (85.4)	4203 (85.6)
Bronchial blockers	1672 (14.6)	707 (14.4)
Dose of Sufentanil, µg	33.47 ± 13.52	33.65 ± 13.43	0.437
Abbreviation: PCNC: postoperative cardiovascular and neurological complications; BMI: body mass index; VATS: video-assisted thoracoscopic surgery; PCA: patient-controlled analgesia

Feature Selection

Among all the assessed variables, univariate logistic regression analysis revealed that age, alcohol consumption history, duration of surgery, intraoperative intercostal nerve block, postoperative PCA, and method of anesthetic intubation were significantly associated with the occurrence of postoperative cardiovascular and neurological complications (Table 3). The candidate variables were filtered, and the anesthetic drug dose-related variable of particular interest to the anesthesiologists (such as sufentanil dose) was incorporated in the multivariate logistic regression analysis; only six variables were individually relevant to PCNC: age (OR = 1.030, 95% CI = 1.026–1.035, p < 0.001), duration of surgery (OR = 1.003, 95% CI = 1.003–1.004, p < 0.001), intraoperative intercostal anesthesia (OR = 0.091, 95% CI = 0.072–0.115, p < 0.001), type of PCA (epidural PCA: OR = 0.452, 95% CI = 0.402–0.508; intravenous PCA: OR = 0.452, 95% CI = 0.398–0.513; p < 0.001), method of intubation (endobronchial blocker tube: OR = 0.553, 95% CI = 0.470–0.651, p < 0.001), and dose of sufentanil (OR = 0.988, 95% CI = 0.984–0.992, p < 0.001) (Table 4). All the above variables were tested for multicollinearity and the VIF < 10. Ultimately, these six independent factors were used to construct predictive models for PCNC.

Table 3

Univariate logistic regression analysis for the PCNC risk factors.
Intercept and variables	Prediction model
Intercept and variables	β	OR (95%CI)	P-value
Gender	-0.013	0.988 (0.902–1.081)	0.785
Age, years	0.037	1.037 (1.033–1.042)	< 0.001
BMI(≤ 22.9/>22.9)	-0.048	0.953 (0.866–1.048)	0.322
History of smoking	0.078	1.081 (0.982–1.191)	0.111
History of alcohol	0.137	1.147 (1.024–1.284)	0.017
History of chemotherapy	0.006	1.006 (0.873–1.160)	0.932
History of surgery	0.008	1.008 (0.920–1.104)	0.863
Duration of surgery, min	0.004	1.004 (1.003–1.004)	< 0.001
Operative technique
Thoracotomy	Reference
VATS	0.008	1.008 (0.919–1.107)	0.863
Intraoperative intercostal nerve block	-2.035	0.131 (0.104–0.164)	< 0.001
Type of PCA
None	Reference
Epidural PCA	-0.141	0.868 (0.782–0.965)	0.008
Intravenous PCA	-0.343	0.710 (0.630–0.799)	< 0.001
Method of intubation
Double-lumen endobronchial tubes	Reference
Bronchial blockers	-0.763	0.466 (0.399–0.545)	< 0.001
Dose of Sufentanil, µg	-0.001	0.999 (0.996–1.003)	0.762
Abbreviation: PCNC: postoperative cardiovascular and neurological complications; BMI: body mass index; VATS: video-assisted thoracoscopic surgery; PCA: patient-controlled analgesia

Table 4

Multivariate logistic regression analysis for the PCNC risk factors.
Intercept and variables	Prediction model
Intercept and variables	β	OR (95%CI)	P-value
Intercept	-2.474	0.084 (0.061–0.117)	< 0.001
Age, years	0.030	1.030 (1.026–1.035)	< 0.001
History of alcohol	-0.095	0.909 (0.806–1.026)	0.122
Duration of surgery, min	0.003	1.003 (1.003–1.004)	< 0.001
Intraoperative intercostal nerve block	-2.402	0.091 (0.072–0.115)	< 0.001
Type of PCA
None	Reference
Epidural PCA	-0.795	0.452 (0.402–0.508)	< 0.001
Intravenous PCA	-0.794	0.452 (0.398–0.513)	< 0.001
Method of intubation
Double-lumen endobronchial tubes	Reference
Bronchial blockers	-0.592	0.553 (0.470–0.651)	< 0.001
Dose of Sufentanil, µg	-0.012	0.988 (0.984–0.992)	< 0.001
Abbreviation: PCNC: postoperative cardiovascular and neurological complications; PCA: patient-controlled analgesia

Model Evaluation and Comparison

Several machine learning models (Logistic regression, XGBoost, Random Forest, Light GBM, and Naïve Bayes), which are more widely used in the medical field, were trained to model PCNC and repeated 10 times. The effectiveness of the models in the training and validation sets was evaluated by calculating AUC values of ROC, as well as specificity, sensitivity, accuracy, PPV, and NPV (Fig. 2) (Table 5). In the training set, Random Forest had the best performance with an AUC value of 0.898, 95% CI of 0.892–0.904, sensitivity value of 0.905, specificity value of 0.712, PPV value of 0.451, and NPV value of 0.966. In the testing set, both Random Forest and Light GBM showed good effectiveness with an AUC value of 0.752. In the validation set, both Random Forest and Light GBM showed good effectiveness with an AUC value of 0.752. XGBoost and logistic regression follow; Naïve Bayes had the lowest AUC value but also exceeded 0.7, demonstrating that all models performed satisfactorily. We also constructed calibration plots for the five models, and we found that the line of each model is around the ideal line both in the training and validation sets (Fig. 3). The Brier values are logistic regression (0.148), XGBoost (0.150), Random Forest (0.114), Light GBM (0.132), and Naïve Bayes (0.142) in the training set and logistic regression (0.149), XGBoost (0.145), Random Forest (0.142), Light GBM (0.142), and Naïve Bayes (0.151) in the validation set. The DCA curves exhibited excellent clinical applicability for all five machine-learning models (Fig. 4).

Table 5

A The values of ROC curves in the training set.
	AUC	95%CI	Accuracy	PPV	NPV	Sensitivity	Specificity
Light GBM	0.795	0.785–0.805	0.694	0.380	0.911	0.745	0.681
Logistic Regression	0.724	0.713–0.735	0.676	0.351	0.884	0.659	0.681
Naïve bayes	0.714	0.703–0.725	0.575	0.304	0.912	0.811	0.514
Randon forest	0.898	0.892–0.904	0.752	0.451	0.966	0.905	0.712
XG Boost	0.758	0.748–0.768	0.644	0.342	0.913	0.779	0.608

Table 5

B The values of ROC curves in the validation set.
	AUC	95%CI	Accuracy	PPV	NPV	Sensitivity	Specificity
Light GBM	0.752	0.736–0.768	0.631	0.333	0.908	0.770	0.595
Logistic Regression	0.718	0.701–0.735	0.646	0.333	0.889	0.699	0.632
Naïve bayes	0.710	0.693–0.727	0.634	0.3253	0.888	0.704	0.616
Randon forest	0.752	0.737–0.768	0.695	0.371	0.889	0.665	0.703
XG Boost	0.737	0.721–0.752	0.677	0.356	0.890	0.681	0.676
Abbreviation: ROC: receiver operator characteristics; AUC: area under the curve; CI: confidence interval; PPV: positive predictive value; NPV: negative predictive value; Light GBM: Light Gradient Boosting Machine; XG Boost: eXtreme Gradient Boosting.

Development of nomogram

Since the predictive nomogram based on logistic regression is more intuitive and broadly used in the clinic, this study developed the nomogram of PCNC in addition to multiple machine learning models for clinicians' convenience (Fig. 5).

In this retrospective study, multivariate logistic regression analysis demonstrated that age and duration of surgery were risk factors independently. In contrast, PCA, intraoperative intercostal nerve block, anesthetic intubation method, and sufentanil use were all independent protective factors. The five machine learning models constructed from these candidate indicators demonstrated excellent predictive efficacy in both training and validation sets; consequently, it is highly applicable to clinical settings. Moreover, accurately identifying protective and disruptive factors via these machine learning models and nomograms may reduce the incidence of PCNC and improve postoperative patients’ quality of life.

In the present study, the incidence of PCNC was approximately 20%, which was different from the results of previous studies, which focused more on severe perioperative cardiovascular and neurological complications such as perioperative acute ischemic stroke and acute myocardial infarction ^{[2, 3]}. However, cardiovascular and neurological events, such as postoperative atrial fibrillation and transient cerebral ischemia, have also been shown in previous studies to be independent risk factors for severe postoperative cardiovascular and neurological adverse events and death ^[4]. Therefore, our study was differentiated from previous studies by including all perioperative cardiovascular and neurological complications, regardless of severity, as outcome variables to be explored in the study and analyzed to derive independent factors affecting the outcome. The criteria for defining PCNC in this study included evidence on ECG, cranial MRI, or CT scans, findings following consultation with a cardiologist or a neurologist, and a diagnosis of new-onset cardiovascular or neurological-related diseases during the postoperative period. These inclusion criteria included almost all the additional PCNC-related diseases as observations, making the study more detailed and comprehensive.

It has been confirmed that age has a critical role as a risk indicator for PCNC ^[14–17]. Additionally, several studies have confirmed the duration of surgery in relation to the high postoperative complications incidence ^{[11, 18, 19]}. The findings of this study confirmed this.

With the increasing trend of opioid-free anesthesia recently, many anesthesiologists consider the opioid analgesics used in the past as unideal and possibly even counterproductive, contributing to the severity of postoperative pain and being strongly associated with adverse postoperative outcomes ^[20–23]. Regarding the use of opioid drugs, the existing studies demonstrate contradictory evidence ^{[24, 25]}. However, most studies examined whether focusing an investigation on opioid-related adverse outcomes such as nausea and vomiting alone ^[26–28] while ignoring postoperative adverse cardiovascular and neurological events may adversely affect patient survival and prognostic outcomes. In real-world practice, experienced anesthesiologists prefer to use a higher dose of sufentanil when a patient’s hemodynamic state becomes unstable. In addition, the duration of surgery is positively correlated with the incidence of PCNC; as the duration of surgery increases, the dosage of sufentanil increases, which may lead to serious bias that a higher dose of sufentanil may be related to the occurrence of PCNC. In contrast, our data analysis showed that sufentanil is an independent protective rather than a risk factor for PCNC. Hence, we are more concerned with the conclusion that the appropriate intraoperative administration of sufentanil helps improve patients’ prognosis during the early postoperative period from the anesthesiologist’s point of view. This may be because maintaining perioperative hemodynamic stability improves perioperative cardiovascular and neurological complications.

In thoracic surgical procedures, an intraoperative intercostal nerve block is an intercostal nerve block performed by thoracic surgeons under direct thoracoscopic view. Compared with the traditional intercostal nerve blocks performed under ultrasound guidance with a more defined block extent, the intercostal nerve block performed by thoracic surgeons under direct thoracoscopic view has not been fully implemented. In this study, the intraoperative intercostal nerve block was shown to be an independent protective factor and accounted for a larger predictive score in the nomogram; therefore, it has a stronger influence on the risk of PCNC. Therefore, this method can be recommended to thoracic surgeons to reduce postoperative patient pain, accelerate recovery, and reduce the incidence of PCNC.

As a standard postoperative analgesic modality used in recent years with individualized, on-demand, and self-controlled drug delivery, PCA ^[29] is receiving increasing attention and popularity among patients. However, postoperative patient-controlled epidural analgesia (PCEA) is associated with adverse postoperative outcomes in surgery, such as hypotension ^[30]. Additionally, patient-controlled intravenous analgesia (PCIA) may cause a marked increment in the incidence of nausea and vomiting in patients ^{[31, 32]}. Although the impact of PCA on PCNC is unclear, it is possible that both PCEA and PCIA are protective because PCA reduces postoperative pain. All of the above three perioperative analgesic modes exerted a significant protective effect on PCNC, thus, we hypothesized that by alleviating perioperative pain, the aim of maintaining stable perioperative hemodynamics and improving perioperative cardiovascular and neurological complications could be achieved ^[33].

Achieving lung separation in thoracic surgery primarily involves using double-lumen endobronchial tubes (DLTs) or bronchial blockers (BBs). DLTs can be easily and precisely inserted, while BBs can reduce the incidence and degree of airway injuries ^{[34, 35]}. However, no studies have investigated the association between the intubation method and PCNC. The present study confirmed that BB is more protective in reducing the risk of PCNC compared with DLT. In contrast to the results of previous studies, we hypothesize that the inflammatory response due to more severe airway injury caused by DLT may be responsible for the higher susceptibility to PCNC ^[35]. Outcomes after VATS pneumonectomy are comparable to open approaches.

While this study developed machine learning models and predictive nomograms for PCNC, it has limitations. A relatively minor sample size was included in the current study. The inclusion of more patients in future studies will reduce bias and increase the accuracy and effectiveness of the prediction model. In addition, in a strict sense, the training and validation sets established in this study cannot be called external validation sets. Hence, multi-center studies are warranted to fill this gap.

This study established machine learning models and nomograms for predicting PCNC, showing that while elderly patients are at risk of developing PCNC adverse events, surgeons can reduce the incidence of PCNC, reduce postoperative adverse outcomes, and improve the patient prognosis by shortening the duration of surgery and providing intraoperative intercostal nerve blocks under direct thoracoscopic view. Meanwhile, anesthesiologists can reduce incidence by performing bronchial blocker intubation, providing additional postoperative PCA, and administering appropriate sufentanil doses intraoperatively. In summary, the concerted effort of the entire operation team is the most effective protective factor and may decrease the risk of PCNC among elderly patients.

AUC Area under the curve

CIs Confidence intervals

CT computed tomography scan

DCA Decision curve analysis

ECG Electrocardiogram

LightGBM Light Gradient Boosting Machine

MRI Cranial magnetic resonance imaging

NPV Negative predictive value

ORs odds ratios

PCA Postoperative patient-controlled analgesia

PCNC Postoperative cardiovascular and neurological complications

PPV Positive predictive value

ROC Receiver operator characteristics curve

ST Suppression of tumorigenicity

VATS Video-assisted thoracoscopic surgery

VIF Variance inflation factor

XGboost eXtreme Gradient Boosting

Ethics approval and consent to participate

The Ethics committee of the First Hospital of China Medical University approved this study (No: 2022-449, November 4, 2022) and Trial registration: Chinese Clinical Trial Registry, 2023.08.01, No. ChiCTR2300074188 (https://www.chictr.org.cn/showproj.html?proj=200744), Principal investigator's name is Wen-fei Tan. For this study, the principles of the Declaration of Helsinki were followed.

Consent for publication

This study did not contain personal identifiers; Therefore, there was no requirement to obtain individual patient consent.

Availability of data and materials

The datasets used and analysed during the current study are available from the corresponding author on reasonable request.

Competing interests

The authors have declared that no confict of interest

Funding

This work was supported by a grant from the National Natural Science Foundation of China (No. 82171187 to T.W.), the Natural Science Foundation of Liaoning Province of China (No. 2022-MS-191 to S.X.) and the Natural Science Foundation of Liaoning Province of China (No. 2022-MS-222 to Y.W.).

Author contributions

WFT and SX contributed to the conceptualization and design of the study. YXW, SYX, JYL, HW and AG collected the data. YXW, YC and WYL conducted the analysis. YXW and SYX led the writing of the original draft. SX, YC, JYL, and WYL edited the manuscript, discussed results, and provided feedback regarding the manuscript. WFT supervised the study and acquired funding. SX has verifed the underlying data. All authors had full access to the data and approved the manuscript for publication.

Acknowledgments

Not applicable

Smilowitz NR, Gupta N, Ramakrishna H, Guo Y, Berger JS, Bangalore S. Perioperative Major Adverse Cardiovascular and Cerebrovascular Events Associated With Noncardiac Surgery. JAMA Cardiol. 2017;2(2):181–7.
Smilowitz NR, Gupta N, Guo Y, Berger JS, Bangalore S. Perioperative acute myocardial infarction associated with non-cardiac surgery. Eur Heart J. 2017;38(31):2409–17.
Bateman BT, Schumacher HC, Wang S, Shaefi S, Berman MF. Perioperative acute ischemic stroke in noncardiac and nonvascular surgery: incidence, risk factors, and outcomes. Anesthesiology. 2009;110(2):231–8.
Lin MH, Kamel H, Singer DE, Wu YL, Lee M, Ovbiagele B. Perioperative/Postoperative Atrial Fibrillation and Risk of Subsequent Stroke and/or Mortality. Stroke. 2019;50(6):1364–71.
Lim E, Harris RA, McKeon HE, Batchelor TJ, Dunning J, Shackcloth M, Anikin V, Naidu B, Belcher E, Loubani M, et al. Impact of video-assisted thoracoscopic lobectomy versus open lobectomy for lung cancer on recovery assessed using self-reported physical function: VIOLET RCT. Health Technol Assess (Winchester Eng). 2022;26(48):1–162.
Alban J, Kennedy K, Hulbert A, Lighter M, Pasquinelli M, Rubinstein I, Ghelani S, Clayburn A, Feldman LE. Surgery for early-stage lung cancer with video-assisted thoracoscopic surgery versus open thoracotomy: A narrative review. Semin Oncol 2022.
Riddersholm S, Tayal B, Kragholm K, Andreasen JJ, Rasmussen BS, Sogaard P, Torp-Pedersen C, Riahi S. Incidence of Stroke After Pneumonectomy and Lobectomy. Stroke. 2019;50(5):1052–9.
Giambrone GP, Wu X, Gaber-Baylis LK, Bhat AU, Zabih R, Altorki NK, Fleischut PM, Stiles BM. Incidence and implications of postoperative supraventricular tachycardia after pulmonary lobectomy. J Thorac Cardiovasc Surg. 2016;151(4):982–8.
Frendl G, Sodickson AC, Chung MK, Waldo AL, Gersh BJ, Tisdale JE, Calkins H, Aranki S, Kaneko T, Cassivi S, et al. 2014 AATS guidelines for the prevention and management of perioperative atrial fibrillation and flutter for thoracic surgical procedures. J Thorac Cardiovasc Surg. 2014;148(3):e153–193.
Vaporciyan AA, Correa AM, Rice DC, Roth JA, Smythe WR, Swisher SG, Walsh GL, Putnam JB Jr.. Risk factors associated with atrial fibrillation after noncardiac thoracic surgery: analysis of 2588 patients. J Thorac Cardiovasc Surg. 2004;127(3):779–86.
Gomez-Hernandez MT, Forcada C, Varela G, Jimenez MF, Spanish Group of Video-assisted Thoracic S. : Operating time: an independent and modifiable risk factor for short-term complications after video-thoracoscopic pulmonary lobectomy. Eur J Cardiothorac Surg 2022, 62(6).
Li W, Liu Y, Liu W, Tang ZR, Dong S, Li W, Zhang K, Xu C, Hu Z, Wang H, et al. Machine Learning-Based Prediction of Lymph Node Metastasis Among Osteosarcoma Patients. Front Oncol. 2022;12:797103.
Li W, Dong S, Wang H, Wu R, Wu H, Tang ZR, Zhang J, Hu Z, Yin C. Risk analysis of pulmonary metastasis of chondrosarcoma by establishing and validating a new clinical prediction model: a clinical study based on SEER database. BMC Musculoskelet Disord. 2021;22(1):529.
Vlisides P, Mashour GA. Perioperative stroke. Can J Anaesth. 2016;63(2):193–204.
Mashour GA, Shanks AM, Kheterpal S. Perioperative stroke and associated mortality after noncardiac, nonneurologic surgery. Anesthesiology. 2011;114(6):1289–96.
Wilcox T, Smilowitz NR, Xia Y, Beckman JA, Berger JS. Cardiovascular Risk Factors and Perioperative Myocardial Infarction After Noncardiac Surgery. Can J Cardiol. 2021;37(2):224–31.
Smilowitz NR, Berger JS. Perioperative Cardiovascular Risk Assessment and Management for Noncardiac Surgery: A Review. JAMA. 2020;324(3):279–90.
Heise D, Bednarsch J, Kroh A, Schipper S, Eickhoff R, Lang S, Neumann U, Ulmer F. Operative Time, Age, and Serum Albumin Predict Surgical Morbidity After Laparoscopic Liver Surgery. Surg Innov. 2021;28(6):714–22.
Surace P, Sultan AA, George J, Samuel LT, Khlopas A, Molloy RM, Stearns KL, Mont MA. The Association Between Operative Time and Short-Term Complications in Total Hip Arthroplasty: An Analysis of 89,802 Surgeries. J Arthroplasty. 2019;34(3):426–32.
Bugada D, Lorini LF, Lavand'homme P. Opioid free anesthesia: evidence for short and long-term outcome. Minerva Anestesiol. 2021;87(2):230–7.
Massoth C, Schwellenbach J, Saadat-Gilani K, Weiss R, Popping D, Kullmar M, Wenk M. Impact of opioid-free anaesthesia on postoperative nausea, vomiting and pain after gynaecological laparoscopy - A randomised controlled trial. J Clin Anesth. 2021;75:110437.
Ziemann-Gimmel P, Goldfarb AA, Koppman J, Marema RT. Opioid-free total intravenous anaesthesia reduces postoperative nausea and vomiting in bariatric surgery beyond triple prophylaxis. Br J Anaesth. 2014;112(5):906–11.
Cozowicz C, Olson A, Poeran J, Morwald EE, Zubizarreta N, Girardi FP, Hughes AP, Mazumdar M, Memtsoudis SG. Opioid prescription levels and postoperative outcomes in orthopedic surgery. Pain. 2017;158(12):2422–30.
D'Amico F, Barucco G, Licheri M, Valsecchi G, Zaraca L, Mucchetti M, Zangrillo A, Monaco F. Opioid Free Anesthesia in Thoracic Surgery: A Systematic Review and Meta Analysis. J Clin Med 2022, 11(23).
Shanthanna H, Ladha KS, Kehlet H, Joshi GP. Perioperative Opioid Administration. Anesthesiology. 2021;134(4):645–59.
Frauenknecht J, Kirkham KR, Jacot-Guillarmod A, Albrecht E. Analgesic impact of intra-operative opioids vs. opioid-free anaesthesia: a systematic review and meta-analysis. Anaesthesia. 2019;74(5):651–62.
Szczepaniak A, Machelak W, Fichna J, Zielińska M. The role of kappa opioid receptors in immune system - An overview. Eur J Pharmacol. 2022;933:175214.
Wang Q, Liu Z, Tang S, Wu Z. Morphine suppresses the immune function of lung cancer by up-regulating MAEL expression. BMC Pharmacol Toxicol. 2022;23(1):92.
Grass JA. Patient-controlled analgesia. Anesth Analg. 2005;101(5 Suppl):S44–61.
Salicath JH, Yeoh EC, Bennett MH. Epidural analgesia versus patient-controlled intravenous analgesia for pain following intra-abdominal surgery in adults. Cochrane Database Syst Rev. 2018;8(8):CD010434.
Kissin I. Patient-controlled-analgesia analgesimetry and its problems. Anesth Analg. 2009;108(6):1945–9.
Momeni M, Crucitti M, De Kock M. Patient-controlled analgesia in the management of postoperative pain. Drugs. 2006;66(18):2321–37.
van Boekel RLM, Warle MC, Nielen RGC, Vissers KCP, van der Sande R, Bronkhorst EM, Lerou JGC, Steegers MAH. Relationship Between Postoperative Pain and Overall 30-Day Complications in a Broad Surgical Population: An Observational Study. Ann Surg. 2019;269(5):856–65.
Knoll H, Ziegeler S, Schreiber JU, Buchinger H, Bialas P, Semyonov K, Graeter T, Mencke T. Airway injuries after one-lung ventilation: a comparison between double-lumen tube and endobronchial blocker: a randomized, prospective, controlled trial. Anesthesiology. 2006;105(3):471–7.
Clayton-Smith A, Bennett K, Alston RP, Adams G, Brown G, Hawthorne T, Hu M, Sinclair A, Tan J. A Comparison of the Efficacy and Adverse Effects of Double-Lumen Endobronchial Tubes and Bronchial Blockers in Thoracic Surgery: A Systematic Review and Meta-analysis of Randomized Controlled Trials. J Cardiothorac Vasc Anesth. 2015;29(4):955–66.

Download PDF

Version 1

posted

You are reading this latest preprint version

Machine learning-based models predict postoperative cardiovascular and neurological complications after pneumonectomy: A 10-year retrospective observational study

Status:

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

Introduction

Methods

Patients

Data collection

Statistical analysis

Results

Patient characteristics

Feature Selection

Model Evaluation and Comparison

Development of nomogram

Discussion

Summary

Abbreviations

Declarations

References

Status:

Version 1