Predicting Reintubation in Postoperative Pediatric Cardiac Surgery: A Machine Learning Approach

Abstract

Background:

Accurate prediction of reintubation in pediatric patients following cardiac surgery is vital for enhancing postoperative care. This study aimed to identify key predictors of reintubation and train a multilayer perceptron (MLP) neural network model for prediction.

Methods:

This retrospective analysis included clinical data from 294 pediatric patients (1–24 months of age) who underwent cardiac surgery and postoperative mechanical ventilation between January and December 2024. Patients who were successfully extubated and monitored for reintubation were included. Significant predictors were identified using Pearson Chi-square (PC²) test and binomial logistic regression analysis (BLRA). An MLP neural network was trained using clinical covariates to predict reintubation.

Results:

Significant predictors of reintubation included low BMI (0.1–1 percentile, P < 0.01, PC²), emergency surgery (P < 0.01, PC²), previous infection (P < 0.01, PC²), pre-reintubation ABG levels (P < 0.001, PC²), and procedure type (aortoplasty, P = 0.05, PC²). Additionally, the duration of ventilation (P = 0.014, BLRA) and the RACHS2 score (P = 0.006, BLRA) were significant predictors. The MLP model achieved a sensitivity of 93.7% and a specificity of 90.5%, with an F1-score of 0.94. The sum of squared error was 0.152, the root mean squared error was 0.248, and the area under the receiver operating characteristic curve was 0.94 for both training and testing datasets.

Conclusion:

The MLP neural network exhibited excellent predictive accuracy for identifying risk factors associated with reintubation.

INTRODUCTION

Advances in detection, diagnosis, and management have significantly improved survival rates, with over 91% of pediatric patients successfully recovering from cardiac surgery.[1] Postoperative mechanical ventilation is a standard intervention for these patients, with most being weaned within 24 hours.[2] However, approximately 18% experience extubation failure, which is defined as reintubation within 48 hours of extubation.[3] Reintubation is associated with unfavorable clinical outcomes, extended hospitalizations, and increased healthcare expenses.[4] Reintubation frequently leads to extended mechanical ventilation duration, increased incidence of ventilator-associated infections, perioperative morbidity, and a greater risk of mortality.[5] These challenges are particularly pronounced in neonates and infants, emphasizing the importance of early recognition and mitigation of risk factors to improve clinical outcomes.[6] Early identification of predictors for reintubation is vital to enhance postoperative care and reduce associated risks.

Reintubation risk is believed to be influenced by specific perioperative factors, and understanding these can guide optimal extubation timing and refine ventilatory weaning protocols.[7] Traditional predictive approaches often struggle to account for the complex, non-linear interplay of clinical variables.[8] In contrast, artificial neural networks (ANNs), particularly the multilayer perceptron (MLP) model, can capture complex nonlinear relationships between clinical variables, outperforming traditional linear models such as logistic regression. Compared to decision trees and support vector machines, MLP offers greater flexibility and predictive power in handling multiple interacting features.[9]

This study focuses on the development and evaluation of an ANN-based model to predict reintubation risk in pediatric patients undergoing cardiac surgery.

SUBJECTS AND METHODS

This retrospective study was conducted at the Department of Cardiac Anesthesia at Sri Jayadeva Institute of Cardiovascular Sciences in Bangalore. The primary aim of the study was to identify predictors of reintubation in pediatric patients undergoing cardiac surgery and to train an ANN model by using this data. Clinical data from January 2024 to December 2024 were analyzed.

The inclusion criteria encompassed pediatric patients aged 1–24 months who underwent either open or closed cardiac surgery, required postoperative mechanical ventilation, and were successfully extubated. Neonates (under 1 month) were excluded due to their complex cardiac physiology, poor compliance with pulmonary rehabilitation, and higher risk of reintubation. Patients were then monitored for any need for reintubation. Exclusion criteria included those requiring prolonged ventilation beyond 150 hours (after which elective tracheostomy is performed per our institutional protocol), patients unable to be weaned off cardiopulmonary bypass (CPB), those needing re-exploration or tracheostomy, and patients with an open chest. Following extubation, children with complex congenital heart disease were typically transitioned to high-flow nasal oxygenation, while others received oxygen via nasal prongs or masks. Additionally, it is standard practice to initiate diuretics (such as furosemide infusions) on postoperative day 1, particularly in high-risk cases such as ventricular septal defect (VSD) with pulmonary hypertension, total anomalous pulmonary venous connection (TAPVC), or tetralogy of Fallot (TOF).[10]

The following clinical covariates were assessed: age, gender, BMI, type of surgery (emergency or elective), preoperative white blood cell (WBC) count, cyanotic disease, preoperative albumin level, chest X-ray findings, previous infections, duration of ventilation, Risk Stratification for Congenital Heart Surgery (RACHS-2) score,[11] procedure intent, CPB time, and aortic clamp time. These variables were retrospectively extracted from the patient records. Missing data were addressed using a bootstrapping method. To mitigate overfitting, we employed dropout regularization and cross-validation.

Statistical analysis

1. Test to assess construct validity: Pearson Chi-square (PC²) tests evaluated the associations between categorical variables[12] and reintubation rates. Binomial logistic regression analysis (BLRA) was utilized to identify predictors of continuous variables and reintubation rates.[13] Sensitivity and specificity were analyzed using the receiver operator characteristic (ROC) curve, while the Youden statistic was applied to determine optimal cutoff values.

2. ANN for prediction: An MLP neural network was used to model the likelihood of reintubation based on the clinical covariates. The MLP model consisted of three layers: input, hidden, and output. [Figure 1]

   • The input layer included 14 units corresponding to the covariates, which were standardized before inclusion in the model.

   • The hidden layer contained eight units, using a hyperbolic tangent (tanh) activation function to capture non-linear relationships.

   • The output layer had two units, representing the dependent variable (reintubation). The activation function used was identity, and the error function was based on the sum of squares to minimize residuals.

   • The MLP model was trained using the Adam optimizer and a binary cross-entropy loss function. Training was conducted over 10 generations, with automated batch training and testing samples created using scaled conjugate gradient optimization. Each generation employed varying train-test splits, optimized through internal validation to minimize the risk of overfitting.

3. Test to assess content validity: The performance of the MLP model was evaluated by calculating the sum of squared error, the percentage of incorrect predictions in both the training and testing datasets, and the root mean squared error (RMSE).[14] Additionally, the area under the ROC curve was computed to evaluate the model’s discriminatory ability.

Figure 1.

MLP neural network

All procedures performed in this study were in accordance with the ethical standards outlined in the Declaration of Helsinki (2013). Data reporting followed the EQUATOR - Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) checklist.[15] The latest TRIPOD-AI checklist is more appropriate for advanced explainable AI models [Supplementary Material 1 ^{(5MB, tif)} ].

RESULTS

The study included 294 participants, with ages ranging from 1 to 24 months (mean = 9.50, SD = 6.58). Of the total participants, 59.9% were male, and 40.1% were female. The weight of participants ranged from 1.5 to 15.0 kg (mean = 5.98, SD = 2.17), while their length ranged from 41.0 to 105.0 cm (mean = 65.30, SD = 9.76). BMI values ranged from 4.5 to 32.0 (mean = 13.70, SD = 2.80). The World Health Organization (WHO) BMI percentiles showed a mean of 12.62 (SD = 23.10), with 37.8% in the 0.1–1 percentile, 18.7% in the 1–3 percentile, and 17.7% in the 3–25 percentile.

Among the procedures, 96.6% were elective, while 3.4% were emergency surgeries. The preoperative WBC count ranged from 4,011 to 25,180 (mean = 8,039.05, SD = 2,778.46), and albumin levels ranged from 2.2 to 4.5 (mean = 3.28, SD = 0.49). Chest X-ray findings revealed pulmonary plethora in 75.5% of cases, oligemia in 20.4%, and cardiomegaly in a few cases. A history of previous infection was reported in 9.2% of participants, and 40.5% had cyanotic heart disease.

The CPB time extended up to 530 minutes (mean = 112.63, SD = 68.48), and the aortic clamp time limited to 413 minutes (mean = 71.74, SD = 49.18). Reintubation was necessary in 8.5% of cases, most commonly due to respiratory distress (3.1%) and PCO₂ retention (2.4%). The time gap to reintubation ranged from 1 to 120 hours (mean = 26.96, SD = 37.15), and the duration of ventilation varied from 1.5 to 134.0 hours (mean = 26.90, SD = 31.67).

Most surgeries (92.9%) were corrective, with 7.1% being palliative. The most common procedures included VSD repair (33%), TOF repair (19.4%), and TAPVC repair (12.6%). Other procedures, such as arterial switch with VSD repair and isolated patent ductus arteriosus closure, were less frequent. The RASCH2 score, which assesses procedural complexity, ranged from 1 to 4, with a mean score of 2.05 (SD = 1.07) [Table 1].

Table 1.

Descriptive statistics

Variable	Frequency	Percent	n	Minimum	Maximum	Mean	Std. Deviation	P
Gender
M	176	59.9						0.659
F	118	40.1
BMI WHO Percentile
0.1–1	111	37.8						<0.01 (PC2)
1–3	55	18.7
3–25	52	17.7
25–50	36	12.2
50–75	22	7.5
75–97	12	4.1
97–99	3	1
99–99.9	3	1
Type
Emergency	10	3.4						<0.01 (PC2)
Elective	284	96.6
Cyanotic Disease
Yes	119	40.5						0.45 (PC2)
No	175	59.5
Reintubation
Yes	25	8.5
No	269	91.5
Pre-Reintubation ABG
Yes	20	6.8						<0.001 (PC2)
No	5	1.7
Not Applicable	269	91.5
AGE			294	1	24	9.5	6.582	0.5 (BLRA)
WEIGHT			294	1.5	15	5.976	2.1652	0.11 (BLRA)
LENGTH			294	41	105	65.303	9.755	0.11 (BLRA)
BMI			294	4.5	32	13.703	2.8041	0.57 (BLRA)
WBC COUNT			294	4011	25180	8039.05	2778.46	0.17 (BLRA)
ALBUMIN LEVEL			294	2.2	4.5	3.283	0.4929	0.26 (BLRA)
CPB TIME			294	0	530	112.63	68.476	0.12 (BLRA)
AORTIC CLAMP TIME			294	0	413	71.74	49.176	0.43 (BLRA)
TIME GAP REINTUBATION			25	1	120	26.96	37.152	0.23 (BLRA)
DURATION OF VENTILATION			294	1.5	134	26.9	31.6718	0.014 (BLRA)
RASCH2 SCORE			294	1	4	2.05	1.065	0.006 (BLRA)

Descriptive statistics and group comparisons of demographic, clinical, and perioperative variables

Conventional statistics to determine the construct validity

Conventional statistical methods were used to assess the construct validity of various predictors of reintubation. A significant association was found between low BMI (0.1–1 percentile) and higher reintubation rates (P < 0.01, PC²). Emergency surgeries had notably higher reintubation rates compared to elective surgeries, with a significant difference (P < 0.01, PC²). A history of previous infection was also linked to increased reintubation rates (P < 0.01, PC²). Pre re-intubation ABG levels were significant predictors of reintubation (P < 0.001, PC²). Among different procedures, aortoplasty showed significantly higher reintubation rates (P = 0.05, PC²), while VSD repair was associated with significantly lower reintubation rates (P < 0.012, PC²). The duration of ventilation (P = 0.014, BLRA) and the RASCH2 score (P = 0.006, BLRA) were also identified as significant predictors of reintubation.

ROC analysis was performed to determine optimal cutoff values for these predictors: the area under the receiver operating characteristic curve (AUC) for ventilation duration was 0.692 (P = 0.002), with an optimal cutoff of 27 hours, yielding 64% sensitivity and 74.3% specificity; for the RASCH2 score, the AUC was 0.69 (P = 0.001), with a cutoff value greater than 2, showing 68.4% sensitivity and 60% specificity.

MULTILAYER PERCEPTRON NEURAL NETWORK MODEL

Model performance

The model achieved 95.12% average correct predictions on the training dataset and 92.11% correct predictions on the testing dataset. Its sensitivity (true positive rate) was 93.7%, and its specificity (true negative rate) was 90.5% [Figures 2-3]. The model attained an F1-score of 0.94, indicating balanced performance. Both the training and testing AUC were 0.94. The RMSE on the testing set was 0.248, indicating a small prediction error [Table 2].

Figure 2.

ROC analysis. Panel A – ROC for duration of ventilation; Panel B – ROC for RASCHS2

Figure 3.

Accuracy of 10 epochs of MLP. Panels A–J depict the ROC curve of 10 epochs of MLP in order

Table 2.

Summary of MLP-ANN

Iteration	Training Accuracy (%)	Testing Accuracy (%)	Training AUC	Testing AUC	Training SSE	Testing SSE	Training RMSE	Testing RMSE
E1	95.3	91.2	0.94	0.94	8.8	5.26	0.21	0.25
E2	95.1	92.1	0.94	0.94	9	5.15	0.21	0.24
E3	95	94.7	0.95	0.95	9.12	4.98	0.21	0.24
E4	95.2	91.9	0.94	0.94	8.86	5.18	0.21	0.25
E5	95.3	92.4	0.94	0.94	8.8	5.08	0.21	0.24
E6	95	87.5	0.94	0.93	9.12	5.62	0.21	0.26
E7	95.8	91	0.95	0.93	3.15	5.74	0.13	0.31
E8	94.9	92.6	0.94	0.94	9.19	5.05	0.21	0.24
E9	95.4	93.8	0.95	0.98	8.74	3.98	0.2	0.23
E10	95.6	93.3	0.95	0.95	8.67	4.53	0.2	0.23
Average	95.12	92.11	0.94	0.94	9.06	5.14	0.21	0.24

Performance summary of the MLP-based ANN model across 10 iterations, showing consistent training and testing accuracy, AUC, and error metrics. The model demonstrates stable generalizability with minimal variation across runs.

Factor importance using MLP

Factor importance analysis revealed that WBC count was the most influential variable (15.40%), followed by the type of surgery (13.30%) and BMI (11.80%). Other key variables included albumin level (9.20%), duration of ventilation (8.00%), and chest X-ray findings (7.50%). Less influential variables included RASCH2 score (4.20%), procedure intent (3.80%), cyanotic disease (2.90%), age (2.60%), and gender (2.30%) [Supplementary Material 2 ^{(4.7MB, tif)} ].

DISCUSSION

This study aimed to identify the predictors of reintubation in pediatric patients undergoing cardiac surgeries and develop an ANN model to predict reintubation likelihood. Our findings revealed that clinical factors such as BMI, type of surgery (emergency vs. elective), preoperative WBC count, cyanotic disease, albumin level, chest X-ray findings, previous infections, and the duration of ventilation were significant predictors of reintubation in pediatric cardiac surgery patients.

Conventional statistical methods and ANN models differed in predicting reintubation factors. Conventional statistics identified key predictors such as low BMI (P < 0.01), emergency surgeries (P < 0.01), previous infections (P < 0.01), and ABG levels (P < 0.001), with moderate predictive power (AUC: 0.692 for ventilation duration and 0.69 for RASCH2 score). In contrast, the ANN model demonstrated superior performance, achieving 92.11% accuracy on the test dataset, with sensitivity at 93.7%, specificity at 90.5%, an F1-score of 0.94, and an AUC of 0.94. The ANN identified WBC count (15.4%), type of surgery (13.3%), and BMI (11.8%) as top predictors. While conventional methods offered interpretability, the ANN provided higher accuracy, capturing complex interdependencies and non-linear relationships.

Brown et al.[16] found that higher preoperative WBC counts (>10,000/mm³) were associated with a higher risk of readmission due to infection and inflammatory diseases. Moynihan et al.[17] found that pediatric cardiac surgical patients with symptomatic viral respiratory infections had longer PICU stays and intubation durations, but no increased mortality. Gupta et al.[18] analyzed 27,398 pediatric cardiac surgery cases and found that 6% of patients extubated in the operating room (OR) required reintubation, compared to 10% in the ICU. High-complexity surgeries, lower-volume centers, and absence of dedicated cardiac ICUs were linked to higher reintubation rates.

The ANN model showed that WBC count of >12,000 were associated with increased risk of reintubation. Emergency surgeries were linked to higher reintubation rates compared to elective surgeries. This may be due to inadequate optimization, lack of prehabilitation, and greater hemodynamic instability in the postoperative period.[19] In our study, a low BMI, with a WHO percentile less than 0.1 to 1, was identified as a significant factor for reintubation (P < 0.01). Complex cardiac surgeries, extended CPB duration, and hemodynamic instability frequently lead to prolonged mechanical ventilation and risk of reintubation.[20] Duration of postoperative mechanical ventilation is a key measure of quality care in pediatric cardiac surgery, with prolonged ventilation increasing risks such as infection, airway injury, and failed extubation.[21] The results from the BLRA suggest that duration of ventilation and RASCH2 score were strong predictors of reintubation. Ventilation lasting over 27 hours showed a sensitivity of 64% and a specificity of 74.3% in predicting reintubation in our dataset. Rooney et al.[3] analyzed extubation failure and found younger age, higher surgical complexity, and specific comorbidities were key extubation failure predictors, while greater nursing hours and critical care certification were associated with lower extubation failure odds. Saengsin et al.[22] developed a clinical risk score to predict extubation failure in pediatric cardiac patients. They identified key risk factors, including a history of pneumonia, intubation at admission, and cyanosis as causes of reintubation. Their score demonstrated good discrimination (AUC = 0.77), and further external validation is needed before clinical implementation.

Elevated WBC count may reflect an underlying systemic inflammatory response or early infection, both of which can adversely affect pulmonary function and extubation readiness. The type of surgery likely serves as a composite indicator of operative complexity, CPB duration, and associated physiological stress, all of which are known contributors to extubation failure. BMI, while not routinely emphasized in neonatal or infant populations compared to absolute body weight, may act as a proxy for nutritional reserves, metabolic stress tolerance, and the adequacy of postoperative recovery. A low BMI may reflect underlying malnutrition or fluid imbalance, both of which increase the vulnerability to respiratory fatigue.[23,24] We also highlight concern regarding the interpretability of predictors identified in narrow percentile ranges (e.g., 0.1–1 percentile). Given the sample size, such thresholds may reflect sample-specific association rather than true generalizable patterns. We need to stress the need for external validation on larger datasets to confirm the clinical utility of these findings.

Interestingly, several conventionally important clinical predictors such as duration of ventilation, age, and preoperative oxygenation indices showed lower feature importance in the ANN model. This may be explained by non-linear interactions between variables or multicollinearity, particularly where BMI or surgical type may encapsulate related physiological domains. Neural networks tend to distribute importance across interacting variables, which can obscure the standalone predictive value of certain features. These findings emphasize the importance of interpreting model outputs in the context of clinical reasoning and the need for transparency tools in AI-based modeling.[25]

Lin et al.[26] developed an explainable AI model to predict successful weaning in prolonged mechanical ventilation patients, with XG Boost achieving the highest AUC (0.908). It provided ventilation (31%), physiology (26.5%), fluid (20.1%), and lab data (18.2%) domains as important factors. Model interpretability remains a key limitation, as most readers perceive AI as a black box. Clear understanding of how the model integrates into clinical workflows is essential for practical adoption. The use of an ANN, MLP model, allowed us to assess the combined influence of multiple covariates on reintubation. This is in line with studies that have identified prolonged mechanical ventilation as a key factor in reintubation risk, as prolonged ventilation often reflects more severe respiratory dysfunction or complications.[22,27,28]

Zhao et al.[29] developed and validated a machine-learning model (CatBoost) to predict extubation failure in ICU patients by using 19 key clinical and laboratory features, achieving high predictive accuracy (AUC: 0.835 internal, 0.803 external). A web-based tool was also created to aid clinicians in risk assessment.

Limitations

The MLP model achieved an optimal balance between sensitivity and specificity, suggesting that it can reliably distinguish between patients at high and low risk for reintubation. While the study presents a novel approach to post-extubation outcomes, reintubation is a multifactorial event. With only 25 reintubations studied across 14 variables, the limited event rate should be acknowledged and highlighted to guide future research. Hemodynamic parameters are crucial in predicting reintubation, but due to the retrospective nature of our study, these parameters were not available for analysis. A common cause of reintubation—residual lesions—was not included in the analysis. The absence of certain key clinical variables, such as the Horowitz index, may limit the model’s clinical robustness. Including such variables could enhance predictive accuracy and real-world applicability. Outlier analysis was not performed in our study. Repeated iterations with varying distributions were employed to reduce overfitting, though some residual risk may remain.

The issue of fairness across age groups, particularly between neonates and older infants, is an inherent limitation of MLP-based ANN models. These models do not inherently adjust for subgroup disparities, which may affect predictive accuracy across age ranges.

Evaluating the utility of ANN in predicting extubation outcomes requires prospective studies comparing its accuracy with clinical decisions made by physicians. Because it is ethically challenging to base extubation decisions solely on ANN predictions, studies should assess outcomes when extubation is determined by physicians blinded to the ANN results. Additionally, understanding how ANN predictions can enhance existing clinical protocols is vital to maximize their utility. ANN models complement physician decisions but require TRIPOD-guided reporting, prospective validation, and ethical consideration of their clinical utility. These would involve comparing outcomes between patients whose extubation decisions are guided by ANN predictions and those managed using standard clinical practices. Beyond reducing extubation failure, such studies should examine whether ANN impacts ventilator duration, rates of tracheostomy, ventilator-associated complications, or mortality.[30]

Moreover, it is critical to acknowledge that respiratory failure and reintubation often result from overall clinical deterioration rather than isolated respiratory factors.

The retrospective design is subject to biases such as missing data or incomplete clinical records, which could potentially affect the generalizability of our findings. Additionally, the data were collected from a single center, limiting the external validity of our results. The lack of external validation is a notable limitation, particularly in the context of strict exclusion criteria. This may further constrain the generalizability to diverse clinical populations. Future prospective studies with multicenter data are required to validate these findings and refine the prediction model.

CONCLUSION

Our study provides important insights into the clinical factors associated with reintubation in pediatric patients following cardiac surgery. The development of an ANN-based predictive model offers a promising tool for clinicians to identify high-risk patients early and tailor postoperative care strategies accordingly.

The use of ANN aimed to predict reintubation and to identify unknown factors responsible for failure of extubation that could not be identified by conventional statistics. Further prospective study with external validation is needed to validate and refine this model for broader clinical use.

This manuscript was presented by the first author at the Annual Conference of Cardiac Anesthetists in India on February 22, 2025, in Kochi, where it received the Janak Mehta Young Scientist Award.

Conflicts of interest

SH and KNR are married

SUPPLEMENTARY MATERIAL

Supplementary Material 1

TRIPOD checklist: prediction model development

Supplementary Material 2

Factor importance over 10 epochs of ANN

Funding Statement

Nil.