Following double-lung transplantations, grade 3 primary graft dysfunction at 72 h (PGD3-T72) is associated with increased risks of graft failure, bronchiolitis obliterans syndrome, and higher one-year mortality [1, 2]. Its incidence varies widely across centers, ranging from 3% to 25%, underscoring the need to reevaluate its risk factors while considering the evolving clinical practices. For instance, ex vivo lung perfusion has expanded the lung donor pool, extending the grafts’ ischemic times, with favorable outcomes [3, 4]. Likewise, tremendous strides have been made with the wider use of intraoperative extracorporeal membrane oxygenation (ECMO) [5] and its extension into the postoperative period [6]. Such dynamic changes in clinical practice, while beneficial for patients, can pose challenges in identifying risk factors for PGD3-T72 development using mathematical models. In fact, the complex interrelationships among these factors often complicate their integration into traditional linear regression models.

Emerging machine learning techniques are promising tools, offering the capacity to detect complex, non-linear relationships among numerous variables associated with PGD3-T72. These approaches have been successfully employed to predict outcomes in kidney [7], liver [8], and pediatric heart transplantation [9]. Yet, their application to lung transplantation remains limited [10–13], particularly in the perioperative setting. Recently, Michelson et al. compared four algorithms to predict PGD3-T72, using features selected via LASSO regression to guide graft selection [14]. Such tools hold potential for informing bedside decisions, though further development is needed to adapt intraoperative strategies dynamically as the surgical procedure progresses.

Building on this foundation, our study leverages a large, prospectively collected dataset with detailed, step-by-step intraoperative data from patients undergoing double-lung transplantation (DLT). We aimed to identify risk factors for PGD3-T72 and develop a simplified, clinically practical, risk scoring system.

The patient inclusion flowchart is depicted in Figure 1. Of the 510 patients who underwent double-lung transplantation (DLT) at our institution during the study period, 477 met the inclusion criteria and were analyzed (83 in the PGD3 group and 394 in the No PGD3 group).

Of these, in 455 cases the organs were sourced from brain-dead donors, while 22 cases involved donation after circulatory death.

Table 1 summarizes the data collected at each step. Our cohort reflected a large portion of cystic fibrosis patients (252, 52.7%) and no patients with primary pulmonary hypertension. Notably, 83 patients (17.3%) who developed a PGD3-T72 had a higher body mass index 21 [18–25] vs. 20 [18–22], p = 0.001, more elevated lactate at all time points (p < 0.001, expect p = 0.003 at step 5), but lower total lung capacity (TLC) 4.9 [3.2–6.3] vs. 6 [4.9–7.5], p = <0.001. Additionally, patients who met the criteria for the French High Emergency Lung Transplantation (HELT) program were overrepresented in the PGD3-T72 group (13 (15.7%) vs. 32 (8.1%) p = 0.03).

ECMO was not used in 251 patients, 7 (8.4%) in the PGD3+ group and 244 (61.2%) in the No PGD3 group (p < 0.001). On the other hand, 27 patients had ECMO in place upon arrival to the operating room: 11 (13.3%) in the PGD3+ group and 18 (4.6%) in the No PGD3 group (p = 0.003). The timing of ECMO cannulation, shown in Figure 2, differed significantly between groups (p = 0.005). Postoperatively, ECMO was continued in 62 (74.7%) patients in the PGD3+ group and 47 (11.9%) in the No PGD3 group (p < 0.001). Primary and secondary postoperative complications are detailed in Table 2.

Machine learning algorithms such as XGBoost offer a contemporary approach to clinical challenges [22]. Through automated variable selection, this method uncovered nonlinear relationships [23], adjusted for confounding factors, and delivered accurate, well-calibrated risk estimates. This study utilized such strengths of the XGBoost machine learning algorithm to predict primary graft dysfunction (PGD3) at 72 h (PGD3-T72) following lung transplantation. A distinctive feature of this research was the sequential development of predictive models at distinct stages of surgery, spanning from the assessment of recipient and donor characteristics to the transfer to the ICU. By progressively integrating intraoperative data, we determined that the highest predictive AUROC for PGD3-T72 was achieved after the second graft was implanted. We identified six key predictive features: recipient TLC and its mismatch with donor TLC, blood lactate levels (reflecting microcirculation), use of ECMO at any point (particularly for hypoxemia), and the PaO2/FiO2 ratio. These factors highlight the complex interplay of recipient characteristics, donor attributes, and intraoperative variables. Additionally, we developed a practical risk score based on these top six features to aid clinicians in assessing PGD3-T72 risk.

Importantly, the top predictors identified by our XGBoost model, including ECMO use, elevated lactate levels, impaired PaO₂/FiO₂ ratio, and donor-recipient total lung capacity mismatch, are consistent with previously published risk factors for primary graft dysfunction [5, 6, 24–26]. Our contribution lies in confirming these variables in a large, granular intraoperative dataset and integrating them into a unified, interpretable risk score with strong predictive performance. Since this scoring system can be implemented mid-surgery immediately after the second graft implantation, it can serve as an early prediction tool that provides clinicians with critical prognostic information, potentially allowing for timely adjustments in intraoperative or immediate postoperative management.

While our findings align with prior studies on PGD3-T72 risk factors, it also revealed novel associations, likely due to variations in institutional practices, evolving definitions of PGD3-T72, graft selection criteria, and intraoperative management [27–29]. In our study cohort, early predictors of PGD included elevated blood lactate at step 7, the PaO2/FiO2 ratio at step 8, and the use of ECMO for hypoxemia. These findings suggest that the pathophysiological mechanism driving the development of PGD likely begins at the stage of initial graft-host interaction, consistent with studies linking biomarker emergence to second graft implantation [24, 25, 30]. Additionally, it is worth noting that blood lactate was particularly predictive in patients who did not require ECMO, possibly underscoring the importance of maintaining adequate microcirculation during surgery.

Consistent with findings from a previous large retrospective cohort study [5], ECMO use was associated with increased incidence of PGD3-T72, regardless of timing. To further investigate the role of ECMO and its impact on model performance, we conducted subgroup analyses stratified by ECMO exposure. In the subgroup of patients who did not receive ECMO, the model achieved strong discriminatory performance (AUROC 0.82), with early intraoperative features, such as elevated lactate and low PaO₂/FiO₂ ratios after anesthetic induction, emerging as key predictors. These findings support the notion that early physiologic deterioration may represent a critical window for intervention, possibly advocating for a lower threshold for ECMO initiation to maintain cellular oxygen delivery in at-risk patients.

In contrast, in patients who received ECMO at any time (preoperative or intraoperative), the model’s performance was substantially reduced (AUROC ∼0.64). This diminished accuracy likely reflects the greater clinical heterogeneity in this subgroup, including variation in ECMO indications, timing of ECMO initiation, and preexisting severity of illness. In this context, the model may be confounded by complex decision-making patterns. Importantly, ECMO initiated specifically for hypoxemia (PaO₂/FiO₂ < 100 mmHg) remained a strong risk factor for PGD3, often occurring after second graft reperfusion, suggesting it may serve as an early clinical surrogate for emerging graft dysfunction.

Taken together, these findings indicate that the current risk score is best suited for use in non-ECMO patients or prior to ECMO initiation. In ECMO-supported patients, its interpretability and predictive power are more limited, and dedicated models tailored to this subgroup may be needed in future work [31, 32].

Another notable discovery is that recipient TLC emerged as a significant risk factor, independent of the type of end-stage lung disease. This may be attributed to the challenging surgical manipulation of severely retracted lungs in pulmonary fibrosis patients or the compromised nutritional status of cystic fibrosis patients [33]. However, TLC was not normalized to patient height in this analysis. Further research is needed to explore these specific patient groups, particularly to identify restrictive subpopulations with elevated chest wall elastance and to develop strategies for accelerating postoperative recovery of chest wall compliance [34].

In line with Tague et al., we found an optimal donor-recipient TLC ratio of 1.2–1.6, which prompted a practice shift following their publication, post-dating this cohort [26]. Such nonlinear relationships, obscured in traditional LR, underscore the value of machine learning.

Michelson et al. introduced a tool to support preoperative graft selection [14]. Our simplified score demonstrates superior discriminatory power, likely due to the inclusion of intraoperative factors affecting outcomes. Consequently, it serves as an effective instrument at the end of surgery for refining early postoperative approaches. Future studies could build on this foundation, developing tools with even greater AUROC values at later time points to optimize ICU postoperative care.

A key strength of this study lies in the detailed granularity of intraoperative data within our database, notably the comprehensive dataset organized around nine surgical steps, with systematic patient assessments at these specific time points. This structure enabled standardized data collection and its alignment with critical clinical moments. Another advantage is the use of a gradient boosting method, which, unlike LR, accommodates missing data without imputation, captures non-linear relationships, and delivers superior discrimination and calibration performance. Additionally, the application of state-of-the-art SHAP analysis provided an in-depth evaluation of how model features influence the risk of PGD3-T72, including the identification of clinically meaningful thresholds. Finally, we developed a simplified risk prediction score that avoids reliance on institution-specific variables, providing a practical tool for any transplantation center to assess PGD3-T72 risk effectively.

Our cohort predominantly featured cystic fibrosis patients, with primary pulmonary hypertension underrepresented due to recruitment patterns. While comprehensive, our dataset lacks variables such as immunologic compatibility and frailty. Unlike other studies, we prioritized early predictive factors to enable rapid clinical responses as primary graft dysfunction mechanisms emerge. Transfusion and fluid balance, introduced at step 9, did not enhance model performance [35, 36]. The repeated inclusion of ECMO-related variables, though unconventional in linear models, improved AUROC and was validated by supplementary analysis. A potential limitation of our study is the inability to explicitly account for variability in intraoperative decision-making, including differences in surgical technique, ECMO cannulation strategy, or operator-specific thresholds for intervention. Although our single-center setting with standardized surgical protocols helps mitigate this variability, some residual confounding is likely. Our model partially addresses this by encoding ECMO timing as a categorical feature, which may act as a surrogate for certain intraoperative choices. Nonetheless, future multi-center studies with access to surgeon- or institution-level metadata could benefit from hierarchical modeling frameworks to isolate operator-driven variability and better understand its impact on model generalizability. Another limitation is that, aside from LR, we did not evaluate a broader range of machine learning algorithms. While many supervised methods (e.g., random forests, support vector machines, deep neural networks) could potentially be applied, we selected XGBoost due to its strong empirical performance on structured data, built-in handling of missing values, and compatibility with SHAP-based interpretability. These characteristics make it well suited for real-time intraoperative applications. Future studies could compare alternative modeling strategies, including ensemble or hybrid architectures, to optimize performance and generalizability.

A key limitation of this study is the moderate sample size (n = 477), which may increase the risk of overfitting. To address this, we employed conservative hyperparameter settings and repeated random split validation, but future studies with larger multicenter cohorts are essential for external validation and generalizability.

Finally, an important limitation of this study is the absence of external validation. Despite outreach to several international centers through the ISHLT network, no collaborating institution was able to provide a dataset with comparable intraoperative granularity, particularly for stepwise modeling around second graft implantation. As a result, the model’s performance has only been demonstrated within a single center, and its generalizability to other clinical environments remains untested. Given known variability in transplant practices (including graft selection, ECMO initiation strategies, anesthetic techniques, and changing indications such as the increasing prevalence of pulmonary fibrosis), model performance may differ across settings. Thus, this model should be viewed as hypothesis-generating. We strongly advocate for prospective, multicenter cohort studies to validate perioperative machine learning models in diverse clinical contexts. To support reproducibility and facilitate such efforts, our full codebase has been made publicly available.

After validation of such models by a multicentric prospective study, the score could be implemented in a simple application or to the electronic record to alert clinicians on the possible risk of PGD3-T72. Therefore, it would suggest discussing within a preventive strategy. Furthermore, it could help to build future studies on prophylactic strategies to reduce PGD3-T72.

In conclusion, gradient boosting effectively predicted PGD3-T72 with an AUROC of 84% immediately after second graft implantation using routine intraoperative data. Further studies are needed to solidify machine learning’s role in primary graft dysfunction prediction and clinical practice. This tool could identify high-risk patients, enabling aggressive preventive measures to improve outcomes [37].