Skip to main content
BJA: British Journal of Anaesthesia logoLink to BJA: British Journal of Anaesthesia
. 2025 Jun 4;135(5):1415–1427. doi: 10.1016/j.bja.2025.04.025

Prediction models for postoperative pulmonary complications: a systematic review and meta-analysis

Zhiyu Huang 1,†, Yang Han 1,†, Huijia Zhuang 1, Jingyao Jiang 2, Cheng Zhou 2,⁎, Hai Yu 1,⁎
PMCID: PMC12597335  PMID: 40473567

Abstract

Background

Postoperative pulmonary complications (PPCs) increase mortality, hospital stays, and healthcare costs. Multivariable prediction models can guide patient care by identifying high-risk patients. The discriminative ability and potential for clinical impact of PPC prediction models remains unclear.

Methods

We systematically searched Cochrane, Embase, and PubMed (up to June 2024) for studies developing or validating prediction models for PPCs that reported c-statistic. The primary outcome was the c-statistic of prediction models for composite PPCs, and the secondary outcome was the c-statistic for individual PPCs, including pneumonia, respiratory failure, reintubation, and others. Data were extracted using the CHARMS checklist, and bias was assessed with PROBAST. For models with data from three or more cohorts, discrimination was synthesised by pooling c-statistic using Bayesian meta-analysis, with heterogeneity assessed through prediction intervals.

Results

A total of 123 studies were included, covering 116 prediction models for PPCs, with 14 models (1 004 029 patients) eligible for meta-analysis. The c-statistic of all models ranged from 0.614 to 0.996 (median 0.80), with 50% of models self-reporting good (c-statistic >0.8) discrimination. In meta-analysis, the ARISCAT PPC score (summary c-statistic 0.76, 95% CI 0.67–0.86), Xue's model (0.82, 0.75–0.89), and CARDOT score (0.73, 0.61–0.85) demonstrated moderate (c-statistic 0.7–0.8) to good discrimination for composite PPCs; the DAGDA score (0.81, 0.74–0.88), Wang's model (0.78, 0.70–0.86), and Jin's model (0.75, 0.68–0.82) for postoperative pneumonia; and Yoon's model (0.90, 0.84–0.96) and Nizamuddin's model (0.85, 0.78–0.92) for postoperative respiratory failure. The reliability of these models, however, is currently limited by the lack of external validation cohorts. Overall, 90.2% of models were assessed as having a high risk of bias.

Conclusions

Many prediction models postoperative pulmonary complications have been developed, but the clinical utility of the vast majority remains uncertain.

Clinical trial registration.

PROSPERO database (CRD42024580216).

Keywords: meta-analysis, prediction models, predictors, postoperative pulmonary complications, systematic review


Editor's key points.

  • •

    Many models intended to predict postoperative pulmonary complications exist, but no comprehensive systematic review and meta-analysis has been performed.

  • •

    This study systematically reviewed existing models, evaluated their quality, and quantitatively synthesised their performance through meta-analysis.

  • •

    The clinical utility of the vast majority of these models remains uncertain. Future research should focus on conducting high-quality external validation of existing models rather than developing new ones.

Postoperative pulmonary complications (PPCs) represent an important clinical challenge in patients undergoing general anaesthesia. Reported incidence rates vary substantially (4.2–64.6%) depending on outcome definitions and study populations, in particular the surgical specialty.1, 2, 3, 4, 5, 6, 7, 8 Although milder events such as atelectasis (30–60%8,9) are relatively common, severe complications such as acute respiratory distress syndrome (ARDS; <2%10) and respiratory failure (3.4–4.2%11,12) occur less frequently but carry grave consequences. The global volume of major surgeries is projected to increase from 230 million13 in 2012 to 310 million14 by 2030, increasing the burden of PPCs. Even rare complications carry significant consequences when they occur, as PPCs are strongly associated with increased patient mortality across short- and long-term15, 16, 17 follow-ups, prolonged hospital stays,16,18 and substantial healthcare costs.19 Addressing PPCs is critical to improving surgical outcomes and reducing healthcare burdens.

Given the significant impact of PPCs, early risk stratification of surgical patients is critical.20 Clinically, it allows physicians to identify high-risk patients and implement preventive measures proactively, while economically, it optimises resource allocation and reduces healthcare costs.

One effective approach for risk stratification is the development of prediction models, a field that has been explored since the 1990s.21 More than 100 models have been developed, yet most studies prioritise creating new models over validating existing ones.22 Furthermore, variability in surgical types and outcome definitions, ranging from composite outcomes to individual complications, has led to a fragmented literature with limited consensus, hindering their clinical adoption.8

Thus, a comprehensive review and synthesis of existing models is necessary. To date, only the STARSurg Collaborative and TASMAN Collaborative have systematically reviewed prediction models for pulmonary complications after abdominal surgery.23 However, their review was limited to abdominal surgeries (29 models were included) and did not include a meta-analysis.

We conducted a systematic review and meta-analysis with the following objectives: (1) to systematically evaluate the bias of prediction models using the standardised PROBAST bias assessment tool; and (2) to extract and synthesise model performance data, including discrimination (c-statistic) and calibration, in order to compare the effectiveness of different models as tools for PPCs risk assessment. Through these efforts, we aim to provide evidence-based and systematic insights to support the clinical application of these models. Furthermore, we hope to offer guidance for future model developers by summarising and discussing key characteristics of model construction.

Methods

We conducted a systematic review and meta-analysis following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.24 The review was registered on PROSPERO (CRD42024580216).

Identification of studies

We searched the Cochrane Central Register of Controlled Trials, Embase, and PubMed from their inception to June 20, 2024, using a combination of keywords and subject headings related to PPCs, including both composite and individual outcomes, and prediction models. As definitions of PPCs vary, we selected individual outcomes for inclusion based on the Standardized Endpoints for Perioperative Medicine-Core Outcome Measures in Perioperative and Anaesthetic Care (StEP-COMPAC) definitions25 and European Perioperative Clinical Outcome (EPCO) definitions.26 The outcomes included pneumonia, atelectasis, respiratory failure, pleural effusion, pulmonary oedema, ARDS, reintubation, pneumothorax, bronchospasm, and pulmonary embolism (Supplementary material 1). We also performed manual searches of citations and reference lists from relevant articles. Duplicates were removed using both automated and manual strategies in EndNote.

To be eligible for inclusion, studies had to develop or validate, or both, at least one prediction model for PPCs and be published in English. Studies were excluded if they only reported measures of association between predictors and PPCs without developing a complete prediction model, if they developed a model but did not report the area under the receiver operating characteristic curve (AUROC) or c-statistic, or if they developed or validated a model not specific to pulmonary complications, such as those for all postoperative complications without separate reporting of pulmonary outcomes. For inclusion in the meta-analysis, models had to provide c-statistic data from at least three cohorts.

Data extraction and quality assessment

Two reviewers (ZH and YH) independently extracted study and cohort characteristics used to develop and validate eligible models, adhering to the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) checklist.27 Two investigators (ZH and HZ) assessed each model's risk of bias and applicability using the Prediction Model Risk of Bias Assessment Tool (PROBAST), which evaluates four domains: participants, predictors, outcomes, and analysis, based on 20 signalling questions.28,29 The evaluation of the participants domain primarily involves assessing the type of data used (prospective, retrospective, or database-based) and the appropriateness of the inclusion and exclusion criteria. For the predictors domain, the evaluation focuses on whether the predictors are accessible when applying the model, and whether the measurements of these predictors were taken without prior knowledge of the outcomes. In the outcomes domain, key considerations include whether the outcome definitions were clearly reported, whether they are reasonable, and whether the predictors were appropriately excluded from the outcome assessment. The analysis domain evaluation addresses whether the outcome incidence is within expected ranges, how missing data were handled, whether univariate analyses were avoided in identifying predictors, and whether both the model's discrimination ability and calibration were calculated. Each domain was rated as having a high, low, or unclear risk of bias. Studies were considered low risk overall if all four domains were rated as low risk. Discrepancies were resolved through group discussions.

To enable quantitative synthesis of model performance, we extracted discrimination and calibration measures. Specifically, we collected AUROC or c-statistic and their corresponding 95% confidence intervals (CIs). If the 95% CI was not reported, we calculated it using methods described by Debray and colleagues.30 Calibration measures included the P-value from goodness-of-fit tests, calibration slope, and Brier score. Additionally, we checked for the reporting of the clinical utility of a model (net benefit in the form of decision curve analysis [DCA]). We also extracted the six highest-weight predictors from each model and classified these variables into three categories: patient factors, procedure factors, and laboratory testing (including laboratory, imaging, and functional tests) following the classification method used by Smetana and colleagues.31 We then summarised the frequency of these predictors across the models.

For studies reporting multiple models, we selected the best model, defined as either the model preferred by the authors or the one with the highest c-statistic. If a study included models for different outcomes, we extracted data for each outcome-specific model separately.

Outcomes

The primary outcome was the discriminative performance of prediction models for composite PPCs, measured by the summary c-statistic and prediction intervals (PIs). Secondary outcomes were the discriminative performance of prediction models for individual PPCs, including pneumonia, atelectasis, respiratory failure, pleural effusion, pulmonary oedema, ARDS, reintubation, pneumothorax, bronchospasm, and pulmonary embolism, based on the StEP-COMPAC25 and EPCO definitions.26

Data synthesis and statistical analysis

We reported continuous variables as means (standard deviations) and categorical variables as percentages. Statistical significance was set at 0.05. If the 95% CI of the c-statistic included 0.50, model discrimination was considered insufficient. A model was considered to have good calibration if the P-value of the Hosmer–Lemeshow goodness-of-fit test (H-L test) was >0.05 or if the calibration slope ranged between 0.95 and 1.05.

We performed a Pearson correlation analysis to assess whether model performance has improved over time. To compare the effects of different modelling approaches, we categorised models built using logistic regression (LR) and machine learning (ML) methods and compared their discriminatory abilities. Similarly, we examined whether the number of predictors included in the models had any effect on their performance using Pearson correlation analysis.

For models with c-statistic data from at least three cohorts, along with sample size and outcome event data, we performed a Bayesian meta-analysis to synthesise the c-statistic of models. This method is recommended for situations with sparse data on predictive performance measures,30,32 as it allows for the pooling of c-statistic and the derivation of PIs. The summary c-statistic provides an overall measure of model discrimination across multiple cohorts, whereas the PIs evaluate between-study heterogeneity and estimate the range of model discriminative performance in future cohorts.32,33 The c-statistic values were classified as follows: 0.60–0.70 (poor), 0.70–0.80 (fair), 0.80–0.90 (good), and >0.90 (excellent).

Meta-analyses were conducted using R (version 4.3.2; R-project; R Foundation for Statistical Computing, Vienna, Austria) with the metamisc packages. All statistical tests were two-tailed, and P-values <0.05 were considered statistically significant.

Discrimination performance versus reliability

We evaluated all 116 models using two key dimensions: (1) discriminative ability, measured by c-statistic, and (2) reliability, based on the number of validation cohorts and the assessment of bias using PROBAST. For the reliability dimension, we used the following scoring system: (1) each model receives 1 point for each validation cohort included in the study; and (2) each model is also assigned 1 point for each main domain in PROBAST where the original study was rated as having ‘low risk of bias’. The models were then plotted on a 2D coordinate system: the X-axis represents c-statistic, whereas the Y-axis represents reliability. The closer a model is to the top-right corner of the plot, the higher its recommendation level, indicating that it has both strong performance and high reliability.

Results

A total of 9880 citations were identified through the search, and 280 potentially eligible articles were retrieved for full-text review (Fig. 1). Of these, 161 studies were excluded as they did not meet the inclusion criteria. Additionally, five studies were excluded because their full texts were not available. An additional 11 studies meeting the inclusion criteria were identified by checking the references of relevant papers and searching for studies that cited these papers. In total, 123 studies conducted between 1997 and 2024, covering 116 models and involving 11 580 645 patients, were included in the systematic review. Fourteen models, which involved 1 004 029 patients and were validated in more than three cohorts, were eligible for meta-analysis.

Fig 1.

Fig 1

Flowchart of literature search for prediction models for postoperative pulmonary complications. AUROC, area under the receiver operating characteristic curve; PPCs, postoperative pulmonary complications.

Characteristics of included studies

The systematic review included a total of 123 studies with varying surgical types and predicted outcomes (Supplementary material 2). All studies were published in English and focused on either composite or individual PPCs, developing or validating one or more models. Most studies utilised retrospective cohort designs (68/123, 55.3%). The number of participants ranged from 8234 to 5 365 114,35 with a mean age of 0.936 to 83.737 yr, and the proportion of women varied from 16% to 71%.38 The percentage of patients with an ASA score of ≥3 ranged from 2%39 to 100%.40 Most of these studies aimed to develop a new prediction model (106/123, 86.2%), while only a few validated existing prediction models (17/123, 13.8%).

The time windows for prediction varied significantly, with the shortest being the PACU stay period41 and the longest extending to 35 days after surgery.42 Notably, a considerable number of studies did not clearly define the time window for outcomes (61/125, 49.6%). We summarised the characteristics of the included studies and cohorts (Supplementary material 3).

Characteristics of included models

A total of 116 prediction models have been developed for risk assessment of PPCs. The most common models are composite PPC prediction models (43/116, 37.1%) and postoperative pneumonia prediction models (45/116, 38.8%) (Table 1). LR is the most widely used modelling method (87/116, 75%), although the use of ML techniques has been growing in recent years. Only a few models have not undergone any form of validation (15/116, 12.9%), whereas 19 models (16.4%) have undergone both internal and external validation. Nearly half of the models (59/116, 42.2%) lack a clinically applicable format that facilitates their application in clinical practice. See model details in Supplementary material 4.

Table 1.

Summary characteristics of included studies and prediction models. ∗There are 111 studies in total for ‘development’ and ‘D+V’, resulting in 116 models owing to the following five studies that addressed multiple outcomes: Kiyatkin and colleagues43 (2023) developed their own PPC prediction model while externally validating the SPORC model44 for predicting PPCs; Han and colleagues45 (2019) predicted both pneumonia and PPCs; Yokoo and colleagues46 (2016) predicted pneumonia and reintubation; Xue and colleagues47 (2021) predicted pneumonia and pulmonary embolism; and Kunisaki and colleagues48 (2016) predicted pneumonia and reintubation. H-L test, Hosmer–Lemeshow goodness-of-fit test; PPCs, postoperative pulmonary complications.

Characteristic Models n (%)
All models∗ (n=116) Composite PPCs (n=43) Pneumonia (n=45) Reintubation (n=12) Respiratory failure (n=11) Others (n=5)
No. of predictive variables
 ≤6 46 (0.4) 19 (0.44) 17 (0.38) 5 (0.42) 2 (0.18) 3 (0.6)
 7–12 30 (0.26) 11 (0.26) 11 (0.24) 3 (0.25) 5 (0.45) 0
 ≥13 29 (0.25) 8 (0.19) 14 (0.31) 4 (0.33) 2 (0.18) 1 (0.2)
 No information 11(9.48) 5 (0.12) 3 (0.07) 0 2 (0.18) 1 (0.2)
Modelling algorithms
 Logistic regression 87 (0.75) 32 (0.74) 33 (0.73) 10 (0.83) 9 (0.82) 3 (0.6)
 Machine learning algorithms 27 (0.23) 10 (0.23) 12 (0.27) 1 (0.08) 2 (0.18) 2 (0.4)
 Neural network 2 (0.02) 1 (0.02) 0 1 (0.08) 0 0
Alternative presentation
 Nomogram 25 (0.22) 7 (0.16) 14 (0.31) 1 (0.08) 1 (0.09) 2 (0.4)
 Score system 36 (0.31) 11 (0.26) 13 (0.29) 6 (0.5) 5 (0.45) 1 (0.2)
 Online calculator 6 (0.05) 1 (0.02) 3 (0.07) 0 2 (0.18) 0
 None 49 (0.42) 24 (0.56) 15 (0.33) 5 (0.42) 3 (0.27) 2 (0.4)
Discrimination performance
 Excellent (c-statistics >0.9) 12 (0.1) 3 (0.03) 6 (0.05) 1 (0.01) 0 2 (0.02)
 Good (c-statistics 0.8–0.9) 46 (0.4) 14 (0.12) 17 (0.15) 6 (0.05) 7 (0.06) 2 (0.02)
 Fair (c-statistics 0.7–0.8) 50 (0.43) 22 (0.51) 18 (0.4) 5 (0.04) 4 (0.36) 1 (0.2)
 Poor (c-statistics 0.6–0.7) 8 (0.07) 4 (0.09) 4 (0.09) 0 0 0
Calibration measures
 Calibration plot 19 (0.16) 5 (0.12) 9 (0.2) 2 (0.17) 3 (0.27) 0
 H-L test 25 (0.22) 9 (0.21) 9 (0.2) 4 (0.33) 2 (0.18) 1 (0.2)
 Brier score 4 (0.03) 3 (0.07) 0 1 (0.08) 0 0
 Multiple calibration measures 30 (0.26) 11 (0.26) 13 (0.29) 3 (0.25) 3 (0.27) 0
 None 38 (0.33) 15 (0.35) 14 (0.31) 2 (0.17) 3 (0.27) 4 (0.8)
Decision curve analysis
 Yes 19 (0.16) 4 (0.09) 12 (0.27) 1 (0.08) 1 (0.09) 1 (0.2)
Validation
 Internal 56 (0.48) 21 (0.49) 24 (0.53) 4 (0.33) 5 (0.45) 2 (0.4)
 External 26 (0.22) 6 (0.14) 10 (0.22) 4 (0.33) 4 (0.36) 2 (0.4)
 Both 19 (0.16) 11 (0.26) 6 (0.13) 0 1 (0.09) 1 (0.2)
 None 15 (0.13) 5 (0.12) 5 (0.11) 4 (0.33) 1 (0.09) 0

We performed a Pearson correlation analysis between the modelling year and discriminative ability for all 116 models, 43 PPC prediction models, and 45 postoperative pneumonia prediction models. Although there was some improvement in model performance over time, this change was not statistically significant (all models r=0.045, P=0.634; PPC models r=0.118, P=0.451; postoperative pneumonia models r=0.306, P=0.158; Supplementary materials 5–7). There was no correlation between the number of predictors and model performance (all models r=–0.019, P=0.848; PPC models r=–0.069, P=0.681; postoperative pneumonia models r=–0.049, P=0.760; Supplementary materials 8–10). Additionally, models constructed using ML methods had higher c-statistic than those built using LR methods. All models: LR 0.785 (0.770–0.800) vs ML 0.852 (0.821–0.886), P<0.001; PPC models: LR 0.769 (0.745–0.793) vs ML 0.811 (0.739–0.883), P=0.239; postoperative pneumonia models: LR 0.781 (0.730–0.807) vs ML 0.875 (0.834–0.923), P=0.002.

A total of 14 prediction models were included in the meta-analysis, with five composite PPC outcome prediction models, six postoperative pneumonia prediction models, two postoperative respiratory failure prediction models, and one reintubation prediction models. Although some models had data from at least three cohorts, they could not be included in the meta-analysis for various reasons. These models and the reasons for their exclusion are detailed in Supplementary material 11.

For the 116 models included in the review, the six most frequently used predictors were age (n=39, 33.6%), type of surgery (n=36, 31.0%), ASA score (n=25, 21.6%), chronic obstructive pulmonary disease (COPD) status (n=25, 21.6%), duration of surgery (n=21, 18.1%), and albumin concentration (n=15, 12.9%) (Supplementary material 12). Additionally, these predictors remained the most widely used in models predicting either composite or individual PPC outcome. Figure. 2 illustrates the predictors used in models for each outcome and their intersections.

Fig 2.

Fig 2

Venn diagram of predictors used in prediction models categorised by postoperative pulmonary complication (PPC) outcomes. Each circle represents a predictor, with grey circles indicating predictors used across different outcome-specific models. The size of each circle reflects the frequency of use of these predictors in the models. APACHE-Ⅱ, Acute Physiology and Chronic Health Evaluation-Ⅱ; AT, anterior thoracic; COPD, chronic obstructive pulmonary disease; DLCO, diffusing capacity of the lung for carbon monoxide; eGFR, estimated glomerular filtration rate; FEV1, forced expiratory volume in 1 s; FVC, forced vital capacity; MP, mechanical power; NYHA class, New York Heart Association class; PT, posterior thoracic; PTCA, percutaneous transluminal coronary angioplasty; rATG, anti-human T-lymphocyte rabbit globulin; RF, respiratory failure; SIRS, systemic inflammatory response syndrome; STAT, the Society of Thoracic Surgeons-European Association for Cardio-Thoracic Surgery (score).

Model performance

Half of the models reported excellent or good discrimination (58/116, 50%), with a few models self-reporting poor discrimination. A total of 78 models (67.2%) used one or more methods to measure calibration (using calibration plots, goodness-of-fit tests, or Brier scores), among which only four reported insufficient calibration ability.49, 50, 51, 52 Only a small number of models (19/116, 16.4%) included DCA.

Risk of bias assessment

We used PROBAST to assess the risk of bias for all studies developing or externally validating a prediction model. Overall, 90.2% (n=111) of model results were assessed as having a high risk of bias (Fig. 3). In the participants and analysis domains, 77.2% (n=95) and 58.5% (n=72) of studies were evaluated as high risk, primarily owing to the lack of prospective cohorts and the absence of simultaneous analysis of model discrimination and calibration. In Supplementary material 13, we provide further details on the risk of bias assessment for all studies.

Fig 3.

Fig 3

Risk of bias across 123 included studies.

Meta-analysis

Prediction model for composite postoperative pulmonary complication outcome

Five prediction models, involving 86 352 patients, were included in the analysis: the Assess Respiratory Risk in Surgical Patients in Catalonia (ARISCAT) PPC score, LAS VEGAS risk score, Xue's model, Ferguson pulmonary risk score, and CARDOT score (Fig. 4). The Ferguson pulmonary risk score was developed specifically for oesophagectomy patients, the CARDOT score for thoracic surgery patients, and Xue's model for abdominal surgery patients. The ARISCAT PPC score and LAS VEGAS risk score were developed and validated across a broader range of surgical types.

Fig 4.

Fig 4

Pooled discrimination performance of composite postoperative pulmonary complication models: c-statistic and prediction intervals. C1, C2, C3, etc. represent Cohort 1, Cohort 2, Cohort 3, etc. within the same study. ∗We excluded the study by Zorrilla-Vaca and colleagues, which focused on patients undergoing lung resection surgery, the majority of whom were critically ill. References are as follows (from top to bottom in the forest plot): Canet53 (2010), Mazo54 (2014), Zorrilla-Vaca55 (2023), Hokari56 (2015), Aceto57 (2021), Kiyatkin43 (2023), Jin58 (2014), Menezes59 (2021), Neto60 (2018), Xue61 (2021), Ferguson62 (2011), Reinersman63 (2016), Xing64 (2016), Pipanmekaporn65 (2019), Pipanmekaporn66 (2024).

The PIs for all models were considerably wider than the 95% CIs for the summary c-statistic, indicating between-studies heterogeneity. Except for Xue's model, the PIs for all other models included 0.5, suggesting that these models may lack sufficient discriminative ability to differentiate between patients who will or who will not develop PPCs when applied to a new cohort.

The ARISCAT PPC score currently has the most external validations among the models, with a summary c-statistic of 0.76 (95% CI 0.67–0.86) and a PI of 0.49–0.86. The study by Zorrilla-Vaca and colleagues exclusively included patients who underwent lung resection, with 93% of patients having an ASA score of ≥3. After this study was excluded from the analysis, the ARISCAT PPC score demonstrated improved predictive performance, with a summary c-statistic of 0.79 (95% CI 0.70–0.87) and a narrower PI of 0.55–0.96. Similarly, after removing the same study, the CARDOT score improved its predictive ability, increasing the summary c-statistic from 0.73 (95% CI 0.61–0.85) to 0.76 (95% CI 0.68–0.86) and narrowing the PI from 0.48–0.97 to 0.60–0.92.

Prediction model for individual postoperative pulmonary complication outcome

Six prediction models for postoperative pneumonia (including DAGDA score, postoperative pulmonary risk index [PPRI], Allou's model, Kinlin's model, Wang's model, and Jin's model), involving a total of 380 730 patients, were included in the analysis. Additionally, two models for respiratory failure (including Yoon's model and Nizamuddin's model), encompassing 527 966 patients, and one model for reintubation (Freundlich's model), based on 8991 patients, were also evaluated.

Allou's model, Kinlin's model, Wang's model, and Freundlich's model were developed and validated specifically for patients undergoing cardiac surgery. The DAGDA score was focused solely on patients who experienced major abdominal surgery, whereas the remaining models included a wide variety of surgical types in their validation and development cohorts.

The discriminative performance of the models is illustrated in Fig. 5. Except for Allou's model, the CIs for the summary c-statistic of all other models did not include 0.5. However, half of the models have PIs that encompass 0.5, indicating the presence of heterogeneity between studies and suggesting that these models may not perform well in a new cohort. Notably, Yoon's model demonstrated excellent discriminative ability for postoperative respiratory failure in noncardiac surgery patients, with a summary c-statistic of 0.90 (95% CI 0.84–0.96). Detailed cohorts for each model and their forest plots can be found in Supplementary materials 14–16.

Fig 5.

Fig 5

Pooled discrimination performance of individual postoperative pulmonary complication models: c-statistic and prediction intervals. References are as follows (from top to bottom in the forest plot): DAGAD score,67 PPRI,56,68,69 Allou's model,70,71 Kinlin's model,72,73 Wang's model,38 Jin's model,74 Yoon's model,75 Nizamuddin's model,76 and Freundlich's model.40,77 More details can be found in Supplementary materials 14–16. PI, prediction interval; PPRI, postoperative pulmonary risk index.

After excluding studies with a high risk of bias, none of the models had enough cohorts to perform data pooling.

Discrimination performance versus reliability

The PPC composite outcome model with the highest recommendation level is the ARISCAT PPC score, whereas the model with the highest c-statistic is Zhou's model. Many models report high discriminative ability (c-statistic >0.8), but owing to the lack of external validation cohorts, their reliability is uncertain (Supplementary material 17). For postoperative pneumonia, the highest-rated model is the PPRI, whereas the model with the highest discriminative ability is Guo's model (Supplementary material 18). For summaries of other individual outcome models, please refer to Supplementary materials 19 and 20.

Discussion

Main findings

This systematic review evaluated 123 studies and 116 prediction models for PPCs. There are numerous models, but most have not undergone reliable validation, making their practical value uncertain. We identified models that had been validated in multiple cohorts and performed a Bayesian meta-analysis on 14 models. The pooled data showed that all models demonstrated varying levels of discriminative ability. Among them, the ARISCAT score had the most validation cohorts and showed stable performance across different cohorts, making it an effective tool for guiding patient care. There are some promising models, such as Yoon's model for postoperative respiratory failure, SPORC-2 (The Score for the Prediction of Postoperative Respiratory Complications) and Freundlich's model for reintubation, but their reliability still needs further validation through independent external cohorts.

The primary methods for constructing prediction models are LR and ML. LR has traditionally been the go-to method, whereas the first prediction model based on ML principles was developed in 2019.45 In the subsequent 5 yr, the use of ML for model development has gained traction, with 27 models emerging by our literature search date. The analysis results indicate that ML models exhibit superior discriminative ability compared with LR models. However, ML models typically incorporate numerous predictive factors and often lack a clinically applicable format. An example of a well-designed model is by Li and colleagues,78 which created a PPC prediction model based on intraoperative dynamic features and simplified it into an online calculator containing 20 predictive factors.

Among the 116 models, nearly half only presented the original mathematical form of the model rather than converting it in a clinically accessible format, such as nomograms or scoring systems. This limits the clinical applicability of these models and their subsequent external validation.

The predictors used in this model are largely consistent with those identified as risk factors for PPCs in previous studies. The six most frequently used predictors—age, type of surgery, ASA score, COPD status, duration of surgery, and albumin—are well-supported by evidence linking them to a modified risk of PPCs. Older age, higher ASA score, and longer surgery duration correlate with a higher risk of PPCs. Certain types of surgeries, such as thoracic surgery, upper abdominal surgery, neck surgery, neurosurgery, and major vascular surgery, are associated with an elevated risk of PPCs.8 COPD is an independent risk factor for PPCs, whereas albumin level acts as a protective factor. Low albumin concentrations, in contrast, increase the risk of PPCs.8 Some well-known risk factors, such as pulmonary function test results, recent respiratory infections, and radiological findings, are not among the most frequently used predictors in existing models. A likely explanation lies in the retrospective nature of data collection. Among the 123 studies included, only 19 used a prospective design. Consequently, model developers often could not access these variables, especially when they are not part of routine preoperative assessments. In addition to availability, interpretability also appears to influence the selection of predictors. It is notable that the six most frequently used predictors are not only easily obtainable but also easy to interpret, which may have influenced model developers' preferences. It is difficult to make a definitive judgement on whether this phenomenon is good or not. There may be an inherent trade-off between the usability of a model and its performance. Achieving an optimal balance between the two likely requires special designed studies and sophisticated analytical strategies.

In the meta-analysis, all models demonstrated varying degrees of discriminative ability. However, among the 14 models analysed, only the ARISCAT PPC score has sufficient independent external validation, whereas other models face issues of inadequate validation cohorts and a lack of diversity in cohort sources. Specifically, all other models were supported by only three to four cohorts, and the validation cohorts for some models (Xue's model, DAGDA score, Jin's model, Wang's model, Yoon's model, Nizamuddin's model, and Freundlich's model primarily originated from a single study). These studies were assessed as having a high risk of bias according to PROBAST, which undermines the strength of the evidence supporting these models. Owing to the lack of available cohorts, conducting subgroup analysis was challenging. In the meta-analyses of the ARISCAT PPC score and CARDOT score, we excluded the study by Zorrilla-Vaca and colleagues, which focused on patients undergoing lung resection surgery, the majority of whom were critically ill (93% of patients had an ASA score of ≥3). These surgeries involved direct intervention on the lungs, and the patients typically had baseline respiratory impairment.

The main source of bias in the included studies was the retrospective nature of data collection. Despite the high risk of bias in the original studies, data pooling can still provide valuable insights by reducing the overestimation of model performance. Typically, the best discrimination is reported in development studies, which often overestimate performance as a result of factors such as subjective bias and overfitting. In contrast, external validations usually show diminished performance. By pooling data from different cohorts and settings, our meta-analysis offers a more accurate and conservative estimate of model performance.

We attempted to provide recommendations for models based on their discriminative ability and reliability. It is worth noting that reliability tends to increase at the expense of performance, as self-reported performance is often better than external validation results from other studies. More external validation cohorts contribute to higher reliability but reduce the overall discriminative ability of models: the ARISCAT PPC score and PPRI are the most trusted models for composite PPC outcomes and postoperative pneumonia, respectively, but their c-statistic are only considered ‘fair’.

Comparison with other reviews

Canet and Gallart79 published a study in 2013 reviewing the predictors and prediction models for PPCs, summarising six prediction models. This review primarily focused on the clinical characteristics of the prediction models. A 2020 systematic review examined prediction models for PPCs in abdominal surgery,23 including 31 studies and 29 models up to March 2020. This review concentrated specifically on abdominal surgery patients and did not conduct a meta-analysis.

Our study expands on these previous works in several ways: (1) we broadened the scope of surgical types to include prediction models for all types of surgery, encompassing 123 studies and 116 prediction models; (2) we incorporated 27 prediction models based on ML, which were not included in the earlier reviews because of the timeframe of literature retrieval; and (3) we provided a quantitative assessment of model performance through meta-analysis.

Strengths and limitations

The primary contribution of this study is that it provides an overall mapping of the available research on prediction models for PPCs. We have made extensive efforts to retrieve and summarise all existing models, describing their characteristics and evaluating the current studies using PROBAST. This systematic approach facilitates further validation of these models and their implementation in clinical practice. Additionally, we conducted a meta-analysis for models with sufficient data.

However, this study has two main limitations. Firstly, there is a lack of uniformity in the definition of PPCs across various studies, which increases heterogeneity and limits the generalisability of the models. Secondly, among the models included in the meta-analysis, only the ARISCAT score has a sufficient validation cohort; the others have fewer validation studies. These two factors reduce the strength of the evidence from the meta-analysis. Furthermore, additional limitations include our inability to perform subgroup analyses to explore sources of heterogeneity as owing to insufficient validation cohorts. Most studies did not report quantitative O:E calibration metrics, preventing us from merging assessments of model calibration.

The inconsistent definitions of PPCs across studies have increased heterogeneity and reduced the reliability of the results. A globally standardised definition is essential for prediction models or any other research using PPCs as an outcome. Two reviews have attempted to consolidate and propose unified definitions for PPCs: the StEP-COMPAC definitions25 and EPCO definitions.26 The StEP-COMPAC definitions suggest that certain diagnoses sharing similar pathophysiological mechanisms should be combined into a composite PPC outcome, including atelectasis, pneumonia, ARDS, and pulmonary aspiration. Other PPCs should be considered separate outcomes. The EPCO definitions take a broader approach to defining PPCs, categorising seven individual outcomes under the composite PPC outcome: respiratory infection, respiratory failure, pleural effusion, atelectasis, pneumothorax, bronchospasm, and aspiration pneumonitis.

The StEP-COMPAC definitions also provide a grading system for the severity of composite PPC outcome, which could be more beneficial for the future development of more refined risk stratification models.

Although the term 'prediction' has been widely used, it is important to clarify that these models do not truly predict the occurrence of PPCs. Making an accurate prediction of future diseases based on current data is difficult, if not impossible. Moreover, for various reasons, the effectiveness of many models has been overreported. In external validation cohorts and real-world clinical practice, we often find significant discrepancies between the model's theoretical performance and its actual effectiveness, further complicating the concept of ‘prediction’. A more practical understanding of prediction models is to integrate multiple relevant risk factors to provide risk-level assessments for different patients, which can guide stratified management and prevention. This aligns with the needs of health economics and humanitarian considerations.

Conclusions and implications for future research

Over the past 30 yr, numerous PPC prediction models have been developed, but most of these models have not been widely used, and only a few have undergone multiple external validations. Among these models with multiple validation cohorts, we believe there is sufficient evidence to support the use of the ARISCAT score for assessing risk in surgical patients and guiding patient care. However, there is insufficient evidence to conclude that other models are adequate, and their performance requires further validation in future research.

We encourage researchers to focus on three areas. Firstly, more external validation studies on existing prediction models must be conducted, as high-quality external validation may prove more valuable than the development of new models, especially for those models in our coordinate system that have strong performance but still lack reliability (see Supplementary materials 17–20 for further details). Secondly. we consider ‘usability’ a critical attribute of effective prediction models; however, this aspect is often overlooked by researchers, with nearly half of the models failing to present the original model in a clinically accessible format, such as nomograms or scoring systems. The ultimate goal of developing prediction models is for clinical application, enabling efficient management of patients through risk stratification. Therefore, we recommend that future model developers consider converting mathematical models into more user-friendly formats. Thirdly, we recommend that future research in prediction models adhere to standardised definitions of outcomes.

Authors’ contributions

Study concept/design: ZH, YH, CZ, HY

Data collection/analysis/interpretation: ZH, YH, HZ, JJ

Drafting of paper: ZH, YH, CZ, HY

Critical revision of paper: ZH, YH, CZ, HY

Final approval of paper: all authors

Data availability statement

All data generated or analysed during this study are included in this article and its supplementary information files.

Funding

Key Research and Development Projects of Sichuan Provincial Department of Science and Technology (Grant number 2024YFFK0255; Received by HY).

Declaration of interest

CZ is a member of the associate editorial board of the British Journal of Anaesthesia.

Handling Editor: Jonathan Hardman

Footnotes

Appendix A

Supplementary data to this article can be found online at https://doi.org/10.1016/j.bja.2025.04.025.

Contributor Information

Cheng Zhou, Email: zhouc@163.com.

Hai Yu, Email: yuhai@scu.edu.cn.

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Multimedia component 1
mmc1.docx (7MB, docx)

References

  • 1.Li G., Freundlich R.E., Gupta R.K., et al. Postoperative pulmonary complications’ association with sugammadex versus neostigmine: a retrospective registry analysis. Anesthesiology. 2021;134:862–873. doi: 10.1097/ALN.0000000000003735. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Zhang Y.T., Han Y., Zhuang H.J., et al. Effect of inspiratory oxygen fraction during driving pressure-guided ventilation strategy on pulmonary complications following open abdominal surgery: a randomized controlled trial. J Clin Anesth. 2024;99 doi: 10.1016/j.jclinane.2024.111676. [DOI] [PubMed] [Google Scholar]
  • 3.Li X.F., Jin L., Yang J.M., Luo Q.S., Liu H.M., Yu H. Effect of ventilation mode on postoperative pulmonary complications following lung resection surgery: a randomised controlled trial. Anaesthesia. 2022;77:1219–1227. doi: 10.1111/anae.15848. [DOI] [PubMed] [Google Scholar]
  • 4.Campos N.S., Bluth T., Hemmes S.N.T., et al. Intraoperative positive end-expiratory pressure and postoperative pulmonary complications: a patient-level meta-analysis of three randomised clinical trials. Br J Anaesth. 2022;128:1040–1051. doi: 10.1016/j.bja.2022.02.039. [DOI] [PubMed] [Google Scholar]
  • 5.Colquhoun D.A., Leis A.M., Shanks A.M., et al. A lower tidal volume regimen during one-lung ventilation for lung resection surgery is not associated with reduced postoperative pulmonary complications. Anesthesiology. 2021;134:562–576. doi: 10.1097/ALN.0000000000003729. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Karalapillai D., Weinberg L., Peyton P., et al. Effect of intraoperative low tidal volume vs conventional tidal volume on postoperative pulmonary complications in patients undergoing major surgery: a randomized clinical trial. JAMA. 2020;324:848–858. doi: 10.1001/jama.2020.12866. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Kirmeier E., Eriksson L.I., Lewald H., et al. Post-anaesthesia pulmonary complications after use of muscle relaxants (POPULAR): a multicentre, prospective observational study. Lancet Respir Med. 2019;7:129–140. doi: 10.1016/S2213-2600(18)30294-7. [DOI] [PubMed] [Google Scholar]
  • 8.Miskovic A., Lumb A.B. Postoperative pulmonary complications. Br J Anaesth. 2017;118:317–334. doi: 10.1093/bja/aex002. [DOI] [PubMed] [Google Scholar]
  • 9.Ferreyra G.P., Baussano I., Squadrone V., et al. Continuous positive airway pressure for treatment of respiratory complications after abdominal surgery: a systematic review and meta-analysis. Ann Surg. 2008;247:617–626. doi: 10.1097/SLA.0b013e3181675829. [DOI] [PubMed] [Google Scholar]
  • 10.Bellani G., Laffey J.G., Pham T., et al. Epidemiology, patterns of care, and mortality for patients with acute respiratory distress syndrome in intensive care units in 50 countries. JAMA. 2016;315:788–800. doi: 10.1001/jama.2016.0291. [DOI] [PubMed] [Google Scholar]
  • 11.Canet J., Sabaté S., Mazo V., et al. Development and validation of a score to predict postoperative respiratory failure in a multicentre European cohort: a prospective, observational study. Eur J Anaesthesiol. 2015;32:458–470. doi: 10.1097/EJA.0000000000000223. [DOI] [PubMed] [Google Scholar]
  • 12.Arozullah A.M., Daley J., Henderson W.G., Khuri S.F. Multifactorial risk index for predicting postoperative respiratory failure in men after major noncardiac surgery. The National Veterans Administration Surgical Quality Improvement Program. Ann Surg. 2000;232:242–253. doi: 10.1097/00000658-200008000-00015. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Weiser T.G., Haynes A.B., Molina G., et al. Estimate of the global volume of surgery in 2012: an assessment supporting improved health outcomes. Lancet. 2015;385:S11. doi: 10.1016/S0140-6736(15)60806-6. [DOI] [PubMed] [Google Scholar]
  • 14.Meara J.G., Leather A.J.M., Hagander L., et al. Global surgery 2030: evidence and solutions for achieving health, welfare, and economic development. Lancet. 2015;386:569–624. doi: 10.1016/S0140-6736(15)60160-X. [DOI] [PubMed] [Google Scholar]
  • 15.Seese L., Sultan I., Gleason T.G., et al. The impact of major postoperative complications on long-term survival after cardiac surgery. Ann Thorac Surg. 2020;110:128–135. doi: 10.1016/j.athoracsur.2019.09.100. [DOI] [PubMed] [Google Scholar]
  • 16.Lugg S.T., Agostini P.J., Tikka T., et al. Long-term impact of developing a postoperative pulmonary complication after lung surgery. Thorax. 2016;71:171–176. doi: 10.1136/thoraxjnl-2015-207697. [DOI] [PubMed] [Google Scholar]
  • 17.Fernandez-Bustamante A., Frendl G., Sprung J., et al. Postoperative pulmonary complications, early mortality, and hospital stay following noncardiothoracic surgery: a multicenter study by the Perioperative Research Network Investigators. JAMA Surg. 2017;152:157. doi: 10.1001/jamasurg.2016.4065. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Fischer M.O., Brotons F., Briant A.R., et al. Postoperative pulmonary complications after cardiac surgery: the VENICE international cohort study. J Cardiothorac Vasc Anesth. 2022;36:2344–2351. doi: 10.1053/j.jvca.2021.12.024. [DOI] [PubMed] [Google Scholar]
  • 19.Fleisher L.A., Linde-Zwirble W.T. Incidence, outcome, and attributable resource use associated with pulmonary and cardiac complications after major small and large bowel procedures. Perioper Med. 2014;3:7. doi: 10.1186/2047-0525-3-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Carr Z.J. Further innovation is required to enhance risk prediction of postoperative pulmonary complications—comment on Anesthesiol Perioper Sci. 2023;1(4):34. Anesthesiol Perioper Sci. 2024;2:14. [Google Scholar]
  • 21.Brooks-Brunn J.A. Development of a predictive model for postoperative pulmonary complications after cholecystectomy. Clin Nurs Res. 1992;1:180–195. doi: 10.1177/105477389200100207. [DOI] [PubMed] [Google Scholar]
  • 22.Zhu Y., Liu X., Li Y., Yi B. The applications and prospects of big data in perioperative anesthetic management. Anesthesiol Perioper Sci. 2024;2:30. [Google Scholar]
  • 23.Kouli O., Murray V., Bhatia S., et al. Evaluation of prognostic risk models for postoperative pulmonary complications in adult patients undergoing major abdominal surgery: a systematic review and international external validation cohort study. Lancet Digit Health. 2022;4:e520–e531. doi: 10.1016/S2589-7500(22)00069-3. [DOI] [PubMed] [Google Scholar]
  • 24.Liberati A., Altman D.G., Tetzlaff J., et al. The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate healthcare interventions: explanation and elaboration. BMJ. 2009;339 doi: 10.1136/bmj.b2700. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Abbott T.E.F., Fowler A.J., Pelosi P., et al. A systematic review and consensus definitions for standardised end-points in perioperative medicine: pulmonary complications. Br J Anaesth. 2018;120:1066–1079. doi: 10.1016/j.bja.2018.02.007. [DOI] [PubMed] [Google Scholar]
  • 26.Jammer I., Wickboldt N., Sander M., et al. Standards for definitions and use of outcome measures for clinical effectiveness research in perioperative medicine: European Perioperative Clinical Outcome (EPCO) definitions. Eur J Anaesthesiol. 2015;32:88–105. doi: 10.1097/EJA.0000000000000118. [DOI] [PubMed] [Google Scholar]
  • 27.Moons K.G.M., Altman D.G., Reitsma J.B., Collins G.S. Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist. PLOS Med. 2014;11 doi: 10.1371/journal.pmed.1001744. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Wolff R.F., Moons K.G.M., Riley R.D., et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med. 2019;170:51. doi: 10.7326/M18-1376. [DOI] [PubMed] [Google Scholar]
  • 29.Moons K.G.M., Wolff R.F., Riley R.D., et al. PROBAST: a tool to assess risk of bias and applicability of prediction model studies: explanation and elaboration. Ann Intern Med. 2019;170:W1. doi: 10.7326/M18-1377. [DOI] [PubMed] [Google Scholar]
  • 30.Debray T.P., Damen J.A., Riley R.D., et al. A framework for meta-analysis of prediction model studies with binary and time-to-event outcomes. Stat Methods Med Res. 2019;28:2768–2786. doi: 10.1177/0962280218785504. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Smetana G.W., Lawrence V.A., Cornell J.E. Preoperative pulmonary risk stratification for noncardiothoracic surgery: systematic review for the American College of Physicians. Ann Intern Med. 2006;144:581. doi: 10.7326/0003-4819-144-8-200604180-00009. [DOI] [PubMed] [Google Scholar]
  • 32.Debray T.P.A., Jaag Damen, Snell K.I.E., et al. A guide to systematic review and meta-analysis of prediction model performance. BMJ. 2017;356 doi: 10.1136/bmj.i6460. [DOI] [PubMed] [Google Scholar]
  • 33.IntHout J., Ioannidis J.P.A., Rovers M.M., Goeman J.J. Plea for routinely presenting prediction intervals in meta-analysis. BMJ Open. 2016;6 doi: 10.1136/bmjopen-2015-010247. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Cheng Y., Zeng P., Wang L., Wu Q. Validating the predictors for pulmonary complications following abdominal surgery in intensive care units using monitored time-series physiological data: a retrospective cohort study. Asian J Surg. 2022;45:1077–1079. doi: 10.1016/j.asjsur.2022.01.064. [DOI] [PubMed] [Google Scholar]
  • 35.Koretsky M.J., Brovman E.Y., Urman R.D., Tsai M.H., Cheney N. A machine learning approach to predicting early and late postoperative reintubation. J Clin Monit Comput. 2023;37:501–508. doi: 10.1007/s10877-022-00908-z. [DOI] [PubMed] [Google Scholar]
  • 36.Tong C., Du X., Chen Y., et al. Machine learning prediction model of major adverse outcomes after pediatric congenital heart surgery: a retrospective cohort study. Int J Surg. 2024;110:2207–2216. doi: 10.1097/JS9.0000000000001112. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Huang T.S., Hu F.C., Fan C.W., Lee C.H., Jwo S.C., Chen H.Y. A simple novel model to predict hospital mortality, surgical site infection, and pneumonia in elderly patients undergoing operation. Dig Surg. 2010;27:224–231. doi: 10.1159/000274485. [DOI] [PubMed] [Google Scholar]
  • 38.Wang D., Li Y., Sheng W., et al. Development and validation of a nomogram model for pneumonia after redo cardiac surgery. J Cardiovasc Med Hagerstown. 2022;23:325–334. doi: 10.2459/JCM.0000000000001302. [DOI] [PubMed] [Google Scholar]
  • 39.Xie D., Chen Q., Zhang Y., et al. Development and validation of a prediction model for postoperative pneumonia in patients who received spinal surgery: a retrospective study. Heliyon. 2024;10 doi: 10.1016/j.heliyon.2024.e29845. [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]
  • 40.Freundlich R.E., Clifton J.C., Epstein R.H., et al. External validation of a predictive model for reintubation after cardiac surgery: a retrospective, observational study. J Clin Anesth. 2024;92 doi: 10.1016/j.jclinane.2023.111295. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Lin H.T., Ting P.C., Chang W.Y., Yang M.W., Chang C.J., Chou A.H. Predictive risk index and prognosis of postoperative reintubation after planned extubation during general anesthesia: a single-center retrospective case-controlled study in Taiwan from 2005 to 2009. Acta Anaesthesiol Taiwan. 2013;51:3–9. doi: 10.1016/j.aat.2013.03.004. [DOI] [PubMed] [Google Scholar]
  • 42.Russotto V., Sabaté S., Canet J. Development of a prediction model for postoperative pneumonia: a multicentre prospective observational study. Eur J Anaesthesiol. 2019;36:93–104. doi: 10.1097/EJA.0000000000000921. [DOI] [PubMed] [Google Scholar]
  • 43.Kiyatkin M.E., Aasman B., Fazzari M.J., et al. Development of an automated, general-purpose prediction tool for postoperative respiratory failure using machine learning: a retrospective cohort study. J Clin Anesth. 2023;90 doi: 10.1016/j.jclinane.2023.111194. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Brueckmann B., Villa-Uribe J.L., Bateman B.T., et al. Development and validation of a score for prediction of postoperative respiratory complications. Anesthesiology. 2013;118:1276–1285. doi: 10.1097/ALN.0b013e318293065c. [DOI] [PubMed] [Google Scholar]
  • 45.Han S.S., Azad T.D., Suarez P.A., Ratliff J.K. A machine learning approach for predictive models of adverse events following spine surgery. Spine J. 2019;19:1772–1781. doi: 10.1016/j.spinee.2019.06.018. [DOI] [PubMed] [Google Scholar]
  • 46.Yokoo H., Miyata H., Konno H., et al. Models predicting the risks of six life-threatening morbidities and bile leakage in 14,970 hepatectomy patients registered in the National Clinical Database of Japan. Medicine. 2016;95 doi: 10.1097/MD.0000000000005466. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Xue B., Li D., Lu C., et al. Use of machine learning to develop and evaluate models using preoperative and intraoperative data to identify risks of postoperative complications. JAMA Netw Open. 2021;4 doi: 10.1001/jamanetworkopen.2021.2240. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Kunisaki C., Miyata H., Konno H., et al. Modeling preoperative risk factors for potentially lethal morbidities using a nationwide Japanese web-based database of patients undergoing distal gastrectomy for gastric cancer. Gastric Cancer. 2017;20:496–507. doi: 10.1007/s10120-016-0634-0. [DOI] [PubMed] [Google Scholar]
  • 49.Zhao D., Ma A., Li S., Fan J., Li T., Wang G. Development and validation of a nomogram for predicting pulmonary complications after video-assisted thoracoscopic surgery in elderly patients with lung cancer. Front Oncol. 2023;13 doi: 10.3389/fonc.2023.1265204. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50.Johnson R.G., Arozullah A.M., Neumayer L., Henderson W.G., Hosokawa P., Khuri S.F. Multivariable predictors of postoperative respiratory failure after general and vascular surgery: results from the patient safety in surgery study. J Am Coll Surg. 2007;204:1188–1198. doi: 10.1016/j.jamcollsurg.2007.02.070. [DOI] [PubMed] [Google Scholar]
  • 51.Hua M., Brady J.E., Li G. A scoring system to predict unplanned intubation in patients having undergone major surgical procedures. Anesth Analg. 2012;115:88–94. doi: 10.1213/ANE.0b013e318257012b. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52.Lukannek C., Shaefi S., Platzbecker K., et al. The development and validation of the Score for the Prediction of Postoperative Respiratory Complications (SPORC-2) to predict the requirement for early postoperative tracheal re-intubation: a hospital registry study. Anaesthesia. 2019;74:1165–1174. doi: 10.1111/anae.14742. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53.Canet J., Gallart L., Gomar C., et al. Prediction of postoperative pulmonary complications in a population-based surgical cohort. Anesthesiology. 2010;113:1338–1350. doi: 10.1097/ALN.0b013e3181fc6e0a. [DOI] [PubMed] [Google Scholar]
  • 54.Mazo V., Sabaté S., Canet J., et al. Prospective external validation of a predictive score for postoperative pulmonary complications. Anesthesiology. 2014;121:219–231. doi: 10.1097/ALN.0000000000000334. [DOI] [PubMed] [Google Scholar]
  • 55.Zorrilla-Vaca A., Grant M.C., Rehman M., et al. Performance comparison of pulmonary risk scoring systems in lung resection. J Cardiothorac Vasc Anesth. 2023;37:1734–1743. doi: 10.1053/j.jvca.2023.05.035. [DOI] [PubMed] [Google Scholar]
  • 56.Hokari S., Ohshima Y., Nakayama H., et al. Superiority of respiratory failure risk index in prediction of postoperative pulmonary complications after digestive surgery in Japanese patients. Respir Investig. 2015;53:104–110. doi: 10.1016/j.resinv.2014.12.004. [DOI] [PubMed] [Google Scholar]
  • 57.Aceto P., Perilli V., Luca E., et al. Predictive power of modified frailty index score for pulmonary complications after major abdominal surgery in the elderly: a single centre prospective cohort study. Eur Rev Med Pharmacol Sci. 2021;25:3798–3802. doi: 10.26355/eurrev_202105_25947. [DOI] [PubMed] [Google Scholar]
  • 58.Jin Y., Xie G., Wang H., et al. Incidence and risk factors of postoperative pulmonary complications in noncardiac Chinese patients: a multicenter observational study in university hospitals. Biomed Res Int. 2015;2015 doi: 10.1155/2015/265165. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 59.Sousa Menezes A., Fernandes A., Rocha Rodrigues J., et al. Optimizing classical risk scores to predict complications in head and neck surgery: a new approach. Eur Arch Otorhinolaryngol. 2021;278:191–202. doi: 10.1007/s00405-020-06133-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 60.Neto A.S., Da Costa L.G.V., Hemmes S.N.T., et al. The LAS VEGAS risk score for prediction of postoperative pulmonary complications: an observational study. Eur J Anaesthesiol. 2018;35:691–701. doi: 10.1097/EJA.0000000000000845. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 61.Xue Q., Wen D., Ji M.H., Tong J., Yang J.J., Zhou C.M. Developing machine learning algorithms to predict pulmonary complications after emergency gastrointestinal surgery. Front Med Lausanne. 2021;8 doi: 10.3389/fmed.2021.655686. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 62.Ferguson M.K., Celauro A.D., Prachand V. Prediction of major pulmonary complications after esophagectomy. Ann Thorac Surg. 2011;91:1494–1500. doi: 10.1016/j.athoracsur.2010.12.036. ; discussion 1500–1. [DOI] [PubMed] [Google Scholar]
  • 63.Reinersman J.M., Allen M.S., Deschamps C., et al. External validation of the Ferguson pulmonary risk score for predicting major pulmonary complications after oesophagectomy. Eur J Cardiothorac Surg. 2016;49:333–338. doi: 10.1093/ejcts/ezv021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 64.Xing X.Z., Gao Y., Wang H.J., et al. Assessment of a predictive score for pulmonary complications in cancer patients after esophagectomy. World J Emerg Med. 2016;7:44–49. doi: 10.5847/wjem.j.1920-8642.2016.01.008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 65.Pipanmekaporn T., Bunchungmongkol N., Punjasawadwong Y., Lapisatepun W., Tantraworasin A., Saokaew S. A risk score for predicting respiratory complications after thoracic surgery. Asian Cardiovasc Thorac Ann. 2019;27:278–287. doi: 10.1177/0218492319835994. [DOI] [PubMed] [Google Scholar]
  • 66.Pipanmekaporn T., Kitswat P., Leurcharusmee P., et al. External validation of the CARDOT score for predicting respiratory complications after thoracic surgery. BMC Anesthesiol. 2024;24:301. doi: 10.1186/s12871-024-02685-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 67.Wong K.K.F., Fung J.T.K., Leung S.K. Risk factor analysis and predictive score for pneumonia after major/ultra-major operations: a single-centre, retrospective study from the Surgical Outcome Monitoring and Improvement Program database. Surg Pract. 2017;21:32–38. [Google Scholar]
  • 68.Arozullah A.M., Khuri S.F., Henderson W.G., Daley J. For the Participants in the National Veterans Affairs Surgical Quality Improvement Program. Development and validation of a multifactorial risk index for predicting postoperative pneumonia after major noncardiac surgery. Ann Intern Med. 2001;135:847. doi: 10.7326/0003-4819-135-10-200111200-00005. [DOI] [PubMed] [Google Scholar]
  • 69.Markar S.R., Walsh S.R., Griffin K., Khandanpour N., Tang T.Y., Boyle J.R. Assessment of a multifactorial risk index for predicting postoperative pneumonia after open abdominal aortic aneurysm repair. Vascular. 2009;17:36–39. doi: 10.2310/6670.2008.00052. [DOI] [PubMed] [Google Scholar]
  • 70.Allou N., Bronchard R., Guglielminotti J., et al. Risk factors for postoperative pneumonia after cardiac surgery and development of a preoperative risk score. Crit Care Med. 2014;42:1150–1156. doi: 10.1097/CCM.0000000000000143. [DOI] [PubMed] [Google Scholar]
  • 71.Wang D., Huang X., Wang H., et al. Risk factors for postoperative pneumonia after cardiac surgery: a prediction model. J Thorac Dis. 2021;13:2351–2362. doi: 10.21037/jtd-20-3586. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 72.Kinlin L.M., Kirchner C., Zhang H., Daley J., Fisman D.N. Derivation and validation of a clinical prediction rule for nosocomial pneumonia after coronary artery bypass graft surgery. Clin Infect Dis. 2010;50:493–501. doi: 10.1086/649925. [DOI] [PubMed] [Google Scholar]
  • 73.Sanagou M., Wolfe R., Leder K., Reid C.M. External validation and updating of a prediction model for nosocomial pneumonia after coronary artery bypass graft surgery. Epidemiol Infect. 2013;142:540–544. doi: 10.1017/S0950268813001416. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 74.Jin X., Wang S., Zhang C., et al. Development and external validation of a nomogram for predicting postoperative pneumonia in aneurysmal subarachnoid hemorrhage. Front Neurol. 2023;14 doi: 10.3389/fneur.2023.1251570. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 75.Yoon H.K., Kim H.J., Kim Y.J., et al. Multicentre validation of a machine learning model for predicting respiratory failure after noncardiac surgery. Br J Anaesth. 2024;132:1304–1314. doi: 10.1016/j.bja.2024.01.030. [DOI] [PubMed] [Google Scholar]
  • 76.Nizamuddin S.L., Gupta A., Latif U., et al. A predictive model for pediatric postoperative respiratory failure: a national inpatient sample study. J Intensive Care Med. 2021;36:798–807. doi: 10.1177/0885066620928272. [DOI] [PubMed] [Google Scholar]
  • 77.Freundlich R.E., Li G., Domenico H.J., Moore R.P., Pandharipande P.P., Byrne D.W. A predictive model of reintubation after cardiac surgery using the electronic health record. Ann Thorac Surg. 2022;113:2027–2035. doi: 10.1016/j.athoracsur.2021.06.060. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 78.Li P., Gao S., Wang Y., et al. Utilising intraoperative respiratory dynamic features for developing and validating an explainable machine learning model for postoperative pulmonary complications. Br J Anaesth. 2024;132:1315–1326. doi: 10.1016/j.bja.2024.02.025. [DOI] [PubMed] [Google Scholar]
  • 79.Canet J., Gallart L. Predicting postoperative pulmonary complications in the general population. Curr Opin Anaesthesiol. 2013;26:107–115. doi: 10.1097/ACO.0b013e32835e8acd. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Multimedia component 1
mmc1.docx (7MB, docx)

Data Availability Statement

All data generated or analysed during this study are included in this article and its supplementary information files.


Articles from BJA: British Journal of Anaesthesia are provided here courtesy of Elsevier

RESOURCES