This retrospective cohort study included adult patients ($\ge$18 years) who underwent cardiopulmonary bypass (CPB)-assisted heart valve surgery at a tertiary medical center between January 2013 and January 2023. Data were extracted from electronic perioperative clinical databases, imaging archives, and surgical repositories. Patient demographics, clinical characteristics, and perioperative variables were systematically collected from electronic health records. Anesthesia documentation was independently verified by two reviewers blinded to postoperative outcomes, using de-identified records prior to PMV adjudication. Inclusion criteria were: (1) primary valvular disease requiring surgical intervention; (2) age $\ge$18 years (no upper age limit); (3) ASA physical status I–III. Exclusion criteria included prior valve surgery, advanced heart failure therapy (ventricular assist device or transplantation), contraindications or intolerance to general anesthesia, and incomplete key perioperative data. Concurrent radiofrequency ablation for atrial fibrillation was permitted. After exclusions (n = 85), 498 patients were included in the final cohort (Fig. 1). Ethics approval was granted by the Institutional Review Board (IRB) of Zhangzhou Affiliated Hospital of Fujian Medical University (Approval No. 2025LWB257), which waived the requirement for informed consent due to the retrospective study design. Standardized protocols were used to ensure a comprehensive capture of demographic characteristics, anesthetic management, and postoperative outcomes, particularly ventilation duration and weaning parameters.

Fig. 1.

Flowchart of patient enrollment and cohort allocation. A schematic overview of inclusion and exclusion criteria and cohort stratification. A total of 583 patients who underwent CPB-assisted heart valve surgery between January 2013 and January 2023. After applying exclusion criteria (n = 85), 498 patients were included and randomly assigned to the training cohort (n = 349) or internal validation cohort (n = 149).

All heart valve procedures were performed under CPB by the center’s surgical team, typically composed of 3 to 4 members, including chief physicians, attending surgeons, and residents. CPB was established via femoral arterial and venous cannulation. Valve interventions (aortic, mitral, tricuspid, or multiple) involved either replacement or repair, with mitral and tricuspid repairs performed whenever anatomically feasible. Concomitant atrial fibrillation ablation and prosthesis selection (mechanical vs. bioprosthetic) were guided by preoperative and intraoperative findings, as well as patient-specific factors such as age, comorbidities, and preferences. Intraoperative efforts focused on minimizing CPB duration and ensuring meticulous hemostasis. Transesophageal echocardiography was routinely employed to assess valve function and confirm procedural success. Postoperatively, patients were transferred to the ICU for continuous hemodynamic monitoring, fluid management, and inotropic support as needed. Multimodal analgesia—combining opioids, non-opioid agents, and regional techniques such as epidural or paravertebral blocks—was prioritized to promote early mobilization and facilitate timely extubation. Standardized ICU protocols guided the management of prolonged ventilation, including daily weaning assessments and prevention of complications (e.g., ventilator-associated pneumonia). Clinical guidelines informed the detection and management of postoperative complications, including bleeding, infection, and arrhythmias. Patients were closely monitored for respiratory distress, hemodynamic instability, and other factors potentially contributing to prolonged ventilation.

PMV was defined as invasive mechanical ventilation lasting more than 48 h after surgery. This included patients who remained continuously intubated for $>$48 h, as well as those who were extubated but subsequently reintubated, with a cumulative ventilation duration exceeding 48 h. Total mechanical ventilation time, measured from the end of surgery to final successful extubation, was also analyzed as a secondary outcome. Arterial partial pressure of oxygen (PaO₂) was assessed using the first arterial blood gas sample obtained within 10 minutes of ICU admission. Pain intensity was measured preoperatively and postoperatively using the visual analog scale (VAS) [14], scored from 0 (no pain) to 10 (worst imaginable pain). All measurements adhered to standardized perioperative assessment protocols, and data were recorded by trained clinical personnel.

The study cohort comprised 498 patients from the Zhangzhou Affiliated Hospital of Fujian Medical University. Patients were randomly assigned to a training cohort (70%) and an internal validation cohort (30%). Comparative analyses between patients requiring PMV and those not requiring PMV (non-PMV) were performed using Fisher’s exact test or the $\chi$² test for categorical variables and the Wilcoxon rank-sum test for non-normally distributed continuous variables. Significant predictors identified via multivariate logistic regression in the training cohort were used to construct a postoperative PMV prediction nomogram. Model discrimination was assessed using the area under the receiver operating characteristic (ROC) curve (AUC). Calibration, defined as the agreement between predicted and observed outcomes, was evaluated using the Hosmer–Lemeshow goodness-of-fit test. Clinical utility was quantified using decision curve analysis (DCA) and clinical impact curves (CIC), which estimate net benefit across a range of predicted probability thresholds. Continuous measures are presented as mean $\pm$ standard deviation if normally distributed and as median with interquartile range (median [IQR]) if non-normally distributed. Categorical variables are summarized as frequencies and percentages. Normality was tested using the Shapiro–Wilk test. A two-tailed p-value 0.05 was considered statistically significant. All analyses were performed using R (version 4.4.0, Boston, MA, USA) and SPSS (version 26.0, IBM SPSS statistics, Chicago, IL, USA).

This study included 498 adult patients ($\ge$18 years) who underwent CPB-assisted heart valve surgery at a tertiary medical center between January 2013 and January 2023. The cohort was randomly assigned to either the training set (n = 349, 70%) or the internal validation set (n = 149, 30%). PMV, defined as ventilation lasting $>$48 h, was required in 162 patients (32.5%), with comparable incidences observed in the training (114 patients, 32.7%) and validation (48 patients, 32.2%) cohorts (Tables 1,2). In the training set, significant differences (all p 0.05, two-tailed) were observed between the PMV group (n = 114) and the non-PMV group (n = 235) across multiple baseline characteristics. These included demographic variables (age, sex distribution, BMI, current smoking status), comorbidities (COPD severity, pulmonary hypertension, atrial fibrillation), functional status (New York Heart Association [NYHA] class, left ventricular ejection fraction [LVEF]), respiratory function (predicted forced expiratory volume in 1 second [FEV1%]), surgical parameters (surgery type, operative time, CPB time), oxygenation status (PaO₂/FiO₂ ratio), and clinical outcomes (intra-aortic balloon pump use, extracorporeal membrane oxygenation [ECMO], tracheostomy, reintubation, 30-day mortality, and overall mortality) (Table 1).

Univariate analysis identified 13 factors significantly associated with PMV (all p 0.05) (Table 3). These included demographic variables (age, male sex, BMI), respiratory status (current smoking, COPD severity grade, FEV1%, pulmonary hypertension), cardiac function (NYHA class III/IV, LVEF%, atrial fibrillation), and procedural metrics (surgery type, operative time, CPB time). Six variables retained independent predictors of PMV in multivariate logistic regression: age (per 1-year increase), BMI (per 1 kg/m² increase), COPD severity (per 1-grade increase), FEV1% (per 1% increase), LVEF% (per 1% decrease), and CPB time (per 10-minute increase). A predictive nomogram incorporating these factors was developed to estimate individualized PMV risk (Fig. 2). Each variable was assigned a weighted score proportional to its regression coefficient. The total score corresponds to the patient-specific probability of requiring PMV $>$48 h, facilitating preoperative risk stratification.

Fig. 2.

Multivariable nomogram for predicting postoperative mechanical ventilation risk. The visual scoring tool incorporates six independent predictors: age (per 1-year increase), body mass index (BMI; per 1 kg/m² increase), chronic obstructive pulmonary disease (COPD) severity (per 1-grade increase), predicted forced expiratory volume in 1 second (FEV1%; per 1% increase), left ventricular ejection fraction (LVEF%; per 1% decrease), and cardiopulmonary bypass (CPB) time (per 10-minute increase). The total score points correspond to the patient-specific probability of requiring mechanical ventilation for $>$48 h postoperatively.

The predictive model for PMV demonstrated strong calibration performance. In the derivation cohort, the mean absolute error between predicted and observed probabilities was 0.011 (Fig. 3). The model also showed robust discriminatory ability, with an AUC of 0.847 (95% CI: 0.798–0.882) during initial testing (Fig. 4). Internal validation confirmed the model’s stability. The mean absolute error was 0.019 (Fig. 3), and the AUC remained high at 0.891 (95% CI: 0.858–0.927) (Fig. 4).

Fig. 3.

Calibration curves of the prolonged mechanical ventilation (PMV) prediction model. (A) Calibration in the derivation cohort (training set, n = 349) showed strong agreement between predicted probabilities and observed PMV rates, with a mean absolute error (MAE) of 0.011. (B) Internal validation cohort (n = 149) demonstrated comparable calibration performance (MAE = 0.019). The dashed diagonal line represents ideal prediction accuracy. Both curves indicate minimal systematic bias, supporting the nomogram’s reliability for clinical risk stratification. PMV was defined as mechanical ventilation lasting $>$48 hours postoperatively.

Fig. 4.

Receiver operating characteristic (ROC) curves evaluating the discriminatory performance of the prediction model for prolonged mechanical ventilation (PMV). (A) In the derivation cohort (training set, n = 349), the model achieved an area under the curve (AUC) of 0.847 (95% CI: 0.798–0.882), indicating excellent discrimination between patients requiring PMV and those extubated within 48 hours. (B) In the internal validation cohort (n = 149), model performance remained high, with an AUC of 0.891 (95% CI: 0.858–0.927). The diagonal reference line indicates chance-level prediction. PMV was defined as mechanical ventilation lasting $>$48 hours following cardiac valve surgery.

DCA was used to quantitatively assess the clinical utility of the PMV prediction model. As shown in Fig. 5, the model demonstrated superior net benefit compared with the strategies of treating all or no patients across a wide range of clinically relevant threshold probabilities. This suggests that applying the model to guide clinical decisions—such as targeted ICU resource allocation or early preventive measures for high-risk individuals—offers greater clinical value than indiscriminate or absent intervention.

Fig. 5.

Clinical utility of the prediction model for prolonged mechanical ventilation (PMV), assessed using decision curve analysis (DCA) and clinical impact curves (CIC). (A) DCA quantifies the net clinical benefit of the predictive nomogram (red curve) compared with treating all or no patients across a range of threshold probabilities (0–100%). The model shows superior clinical utility across nearly all thresholds range, with cost-benefit ratios indicated on the x-axis. (B) CIC evaluates the number of patients classified as high-risk (red curve) versus the number of actual PMV events (blue curve) at each threshold probability. Below the 40% threshold (dashed vertical line), high-risk classifications exceed observed PMV cases, indicating increased false positives. At the prespecified 40% cutoff, the model achieves optimal alignment between predicted and observed outcomes, maximizing true-positive identification while maintaining an acceptable false-positive rate. These findings support the model’s application in targeted ICU resource allocation. PMV was defined as mechanical ventilation lasting $>$48 hours following cardiac surgery.

Further assessment using a clinical impact curve revealed that, at thresholds below 40%, the number of false-positive classifications (i.e., patients identified as high-risk who did not experience PMV) increased disproportionately relative to true positives (Fig. 5). This underscores the need for careful threshold selection to balance the benefits of intervention with the risk of overtreatment. Calibration-guided threshold analysis identified a predicted risk probability of 0.40 as the optimal cutoff. At this threshold, true-positive identification was maximized while maintaining an acceptable false-positive rate, supporting its use for targeted postoperative respiratory monitoring. Overall, both DCA and the clinical impact curve confirmed that predicted probability thresholds between 40% and 100% yielded a meaningful clinical benefit. Within this range, the model achieved an optimal trade-off between minimizing false negatives and avoiding false positives. These findings support the model’s utility in enhancing perioperative decision-making and ICU resource allocation for patients undergoing valve surgery.

This study developed and internally validated a predictive nomogram incorporating six readily available clinical variables—age, BMI, COPD severity, FEV1%, LVEF%, and CPB time—to stratify the risk of PMV ($>$48 h) following heart valve surgery. The model demonstrated robust discrimination (AUC 0.847–0.891) and calibration across both the training and validation cohorts. DCA confirmed a clinically meaningful net benefit at the 40% risk threshold, which optimally balanced sensitivity and specificity for practical implementation. This threshold enables preemptive interventions for high-risk patients, including preoperative pulmonary optimization (e.g., inspiratory muscle training or bronchodilator therapy), ICU resource prioritization (e.g., allocation of advanced ventilators and respiratory therapists), and protocolized lung-protective ventilation strategies. The nomogram allows rapid bedside risk assessment using routinely collected variables, addressing the need for early identification of patients at increased risk for extended ventilatory support. Given the substantial morbidity and economic burden associated with PMV—as prior studies suggest hospital costs for affected patients may significantly exceed those of non-PMV patients [15]—timely identification and intervention may reduce complications such as ventilator-associated pneumonia and ICU length of stay.

Our findings align with previously identified predictors for PMV [4, 9, 11, 16, 17], including advanced age, impaired cardiac function (reduced LVEF%), and prolonged CPB duration. These factors reflect established pathophysiologic mechanisms linking cardiopulmonary stress to postoperative ventilatory dependency [18]. However, our model improves upon prior approaches in several ways. First, unlike models such as the Intensive Care Unit Respiratory Support Score (ICURSS), which rely solely on static preoperative variables, or the Simplified Acute Physiology Score II (SAPS II), which focuses on ICU parameters [19, 20], our nomogram incorporates intraoperative data (i.e., CPB duration), thereby improving temporal relevance to surgical insult. Second, the model is procedure-specific, targeting valve surgery rather than mixed cohorts dominated by coronary artery bypass grafting [1, 3, 7, 8], allowing greater sensitivity to the unique intraoperative and postoperative risks of valvular procedures. Third, the use of a $>$48-hour PMV definition aligns with contemporary critical care standards [9], offering a balance between overestimation ($>$24 hours) and delayed identification ($>$72 hours). Finally, the use of continuous variables (e.g., FEV1% instead of categorical COPD staging) allows more precise risk estimation. The empirically derived 40% cutoff further supports actionable, resource conscious-intervention planning.

Beyond risk prediction, this model offers a platform for real-time, actionable clinical decision-making. Patients identified as high-risk ($\ge$40% probability) can be prioritized for preoperative pulmonary rehabilitation, ICU bed allocation, and early implementation of postoperative interventions such as bronchoscopy readiness and secretion clearance protocols. These are not only clinically relevant but economically justified—Rajakaruna et al. [4] reported that PMV patients accrued 5.2-fold higher costs than non-PMV patients ($\$$14,286 vs. $\$$2761 per case). Targeting high-intensity resources toward those most likely to benefit may mitigate overall expenditure. Although this was a single-center study, its cohort included heterogeneous surgical profiles: 38% of patients had atrial fibrillation, 32% underwent combined valve procedures, and repeat operations were included. This heterogeneity enhances the generalizability of the findings. Importantly, the model excludes postoperative variables (e.g., vasopressor requirements), enabling application at the point of surgical scheduling. While this limits adaptability to intraoperative complications, it enhances early planning value. Future prospective studies should evaluate whether model-guided interventions reduce PMV rates, ICU stays, and related costs. Integration into electronic health records could automate risk alerts, prompting intraoperative strategies such as bronchospasm prophylaxis or intensified recruitment maneuvers.

This study has several limitations that merit consideration. Its retrospective, single-center design introduces potential selection bias and limits generalizability to centers with different surgical practices or case mixes. Although internal validation was rigorous, external verification in multicenter populations—particularly those with higher extracorporeal membrane oxygenation or urgent procedures—is needed. The exclusion of dynamic postoperative data (e.g., serial blood gases or sedation depth) limits the model’s capacity for recalibration, which could be addressed in future iterations using machine learning algorithms. While the nomogram is simple and clinically accessible, its real-world impact depends on integration into hospital informatics systems for automated risk scoring. Future studies should assess whether model-based preoperative interventions (e.g., preoperative physiotherapy for high-risk patients) translate into reduced PMV incidence and better resource utilization. Despite these limitations, the current model provides a physiologically informed, clinically actionable tool for preoperative risk stratification and resource allocation in heart valve surgery.