Secondary Logo

Journal Logo

Artificial intelligence

Machine learning-based prediction of massive perioperative allogeneic blood transfusion in cardiac surgery

Tschoellitsch, Thomas; Böck, Carl; Mahečić, Tina Tomić; Hofmann, Axel; Meier, Jens

Author Information
European Journal of Anaesthesiology: September 2022 - Volume 39 - Issue 9 - p 766-773
doi: 10.1097/EJA.0000000000001721
  • Free

Abstract

KEY POINTS

  • Machine learning algorithms can reliably exclude a patient as likely to need massive perioperative blood transfusion with a low number of routinely available preoperative patient features.
  • Expert clinicians given the same low number of patient features are not able to produce results of similar predictive power.
  • By excluding patients unlikely to require massive perioperative blood transfusion, it is possible to attend more acutely to those patients that remain at-risk of massive transfusion.
  • It may be possible to adjust perioperative treatment across specialities to prevent massive perioperative blood transfusion and in turn prevent adverse consequences from such transfusion needs.

Introduction

Cardiac surgery bears a high risk of perioperative bleeding, anaemia and blood transfusion.1 Although many patients undergoing cardiac surgery do not require perioperative blood transfusion,2 some encounter massive perioperative blood loss, and massive perioperative allogeneic blood transfusion (MPABT) with associated negative influence on outcome.3

’Patient Blood Management’ (PBM), a multimodal, multidisciplinary treatment bundle includes preserving a patient's own blood via the preoperative correction of anaemia using iron, erythropoietin, folate and vitamin B12, the minimisation of intraoperative blood loss, and optimising the physiological anaemia tolerance by maximising oxygen delivery during mechanical ventilation while reducing metabolic demand.4,5 The combination of these measures helps to decrease the utilisation of allogeneic blood, reducing the likelihood of MPABT.6 However, the implementation of PBM measures can be challenging, and complete utilisation is often not achieved.7 Similar to the Pareto principle, the missed opportunity for PBM might be particularly detrimental in the small percentage of cardiac surgery patients requiring the largest amount of perioperative transfusions. Typically, MPABT is not only an indicator for complications but also a cause for increased morbidity and mortality.8

Most of the relevant publications do not focus on patient level prediction and offer low predictive power for MPABT. Findings regarding treatment from these publications are not easily applied to massive transfusion settings. Furthermore, features used in generating these results are not routinely present in the preoperative phase (e.g. the inclusion of features only available during surgery).9–15 However, with an accurate prediction model, using a small number of preoperatively available features, it should be possible to identify patients prone to MPABT and adapt blood conservation pathways tailored to individual patient needs, thus reducing MPABT events.

Artificial intelligence is a bundle of modern methods for analysis of large amounts of data in settings, such as classification or regression. The potential of artificial intelligence to solve problems previously considered unsolvable and push boundaries in current scientific frontiers has been proven.16

Artificial intelligence makes use of methods, such as neuronal networks that mimic the way the brain functions in that data is used to adjust weighting of certain of the network's neurons. In this training phase, the model becomes more and more adept in deciding, for example, between two classes. A trained network is then able to classify previously unseen (test) data in the same way. One strength of artificial intelligence is the ability to make classification predictions using very specific sets of input features that may be limited in quantity or might not have any humanly discernible connection.

Therefore, we developed prediction models based on machine learning to identify patients at high risk of MPABT.

Methods

Ethics

This was a single centre (Kepler University Hospital, Linz, Austria), retrospective cohort study. After approval and written informed consent was waived by our local ethics committee (Ethics Committee of the Johannes Kepler University, Faculty of Medicine, Krankenhausstraße 5, 4020 Linz, Austria; Chairperson: Univ. Prof. Dr Johannes Fischer, study number 1091/2021), and registration at clinicaltrials.gov (NCT04856618, principal investigator: Thomas Tschoellitsch, 16 April 2021), data of patients with and without transfusions who underwent cardiac surgery of any kind between 01 January 2010 and 31 December 2019 was included. All preoperatively and postoperatively available data on transfusions were obtained from the patients’ electronic health records. All predictions were made from features that were already available before surgery. Exclusion criteria were: re-operation of the same patient, adult patients with congenital heart disease, age less than 18 years. This manuscript adheres to the STROBE guidelines for cohort studies.17

Data was merged from different subsets of our data warehouse and anonymised. The preoperative dataset contained age, weight, height, gender, American Society of Anesthesiologists Physical Status Classification Score (ASA), all parameters obtained from the Austrian Inpatient Quality Indicators (A-IQI),18 all preoperative laboratory values (Table 1), and the calculated value of EuroSCORE II.19

Table 1 - Preoperative laboratory data
All n = 3782 RBCmod n = 3643 RBC10 n = 139 P
Hb (g dl−1) 14 ± 1.7 14 ± 1.7 12 ± 2.4 <0.000
Platelets (nl−1) 221 ± 69 221 ± 68 219 ± 94 0.751
aPTT (s) 26 ± 8 26 ± 7 31 ± 23 0.013
leukocytes (nl−1) 7.7 ± 3 7.6 ± 2.8 10 ± 6.9 <0.000
MCV (fl) 88 ± 4.8 88 ± 4.8 89 ± 5.5 0.065
creatinine (mg dl−1) 1.2 ± 0.92 1.2 ± 0.89 1.7 ± 1.4 <0.000
GFR (ml min−1) 70 ± 23 71 ± 23 57 ± 26 <0.000
CRP (mg dl−1) 0.97 ± 2.6 0.89 ± 2.3 3.2 ± 6.6 <0.000

The primary outcomes of interest were the number of patients who received less than or at least 10 [pRBC units during their hospital stay (<10 pRBC units was classified as moderate transfusion (RBCmod), ≥10 pRBC was classified as RBC10] and the ability of five artificial intelligence algorithms to predict, which patient would be in the RBC10 group. The threshold of 10 pRBC units was chosen in approximation to the definition of massive transfusion in trauma medicine, where the critical administration threshold is seen as 10 pRBC within 24 h.20

The dataset underwent extensive data preprocessing and data cleaning, including detection of typographical errors and out of range values as well as the imputation of missing values:

  • (1) All features with more than 25% of missing values were excluded. The remaining missing values were imputed using Strawman imputation,21 and in line with standard statistical and data reporting guidelines, we also employed an advanced multiimputation method, ‘missForest’. Both methods yielded equally good results, hence we decided to apply the computationally less intensive method, Strawman imputation.22
  • (2) Censored numerical data were truncated (e.g. ‘<0.1’ was replaced by 0.1).
  • (3) Categorical features with more than two values were one-hot encoded. Ordinal features were encoded as positive integers. Binary and numerical features were included as they were.

This resulted in a dataset with 64 variables and 3782 patients for analysis. The dataset was split into training (for model training) and test (for performance evaluation of the trained model) datasets, 80 and 20% of the data, respectively. In the case of patients with multiple admissions, we solely assigned their admissions to either the training or the test set to avoid overestimation of the model's ability to generalise to previously unseen data.

To distinguish between RBCmod and RBC10 at the time of hospital admission, we used five machine learning techniques: logistic regression, Random Forest, neural network, gradient boosting machine (GBM) and adaptive boosting (ADA).

For training, we applied a random search with 25 iterations for our hyperparameter selection (i.e. hyperparameter search23) for all models except neural network where we used a grid search, both of which are provided by the ’caret’ package for R (R 4.0.0, Vienna, Austria).24 For training models using each of the machine learning methods, we used five-fold cross validation on the training set.25 Each of the machine learning methods was trained on the same training set.

Finally, we used the test dataset to assess each method's ability and reliability to generalise to previously unseen cases. Data from the test set was not used for training, thus avoiding data leakage.

Statistical analysis

To evaluate the performance of our models, we used the following quality measures: positive-predictive value and negative-predictive value (PPV and NPV, respectively), area under the receiver operator characteristic curve (AUC) and the F1 score.26

Using the Boruta package of the R software package, we could determine the most important features for prediction of MPABT using a Random Forest model.27

In addition, we asked 100 specialists in anaesthesiology and intensive care at the Kepler University Hospital in Linz, Austria, and the University Hospital in Zagreb, Croatia, to predict MPABT using the same eight most important features as the best model and compared these results with the results of the best model.

Results

An overview over the preoperative patient data is given in Table 2. In total, 6313 pRBC units were administered to the 3782 patients included in the study. Of these, 139 patients (3.7%) received at least 10 pRBC units during their hospital stay, accounting for about 50% of all transfusions given (Fig. 1).

Table 2 - Patient characteristics
All n = 3782 RBCmod n = 3643 pRBC range: 0 to 9 pRBC mean: 0.895 RBC10 n = 139 pRBC range: 10 to 87 pRBC mean: 21.964 P
Female 1124 (30%) 1072 (29%) 52 (37%) 0.043
Male 2658 (70%) 2571 (71%) 87 (63%)
Age (years) 67 ± 11 67 ± 11 66 ± 13 0.203
EuroSCORE II 4.0 ± 5.7 3.8 ± 5.4 9.5 ± 9.1 <0.000
ASA
 I 1 (0%) 1 (0%) 0 (0%) 0.845
 II 102 (3%) 102 (3%) 0 (0%) 0.062
 III 2796 (74%) 2739 (75%) 57 (41%) <0.000
 IV 600 (16%) 541 (15%) 59 (42%) <0.000
 V 45 (1%) 33 (1%) 12 (9%) <0.000
NYHA
 I 304 (8%) 288 (8%) 16 (12%) 0.125
 II 1642 (43%) 1614 (44%) 28 (20%) <0.000
 III 1629 (43%) 1568 (43%) 61 (44%) 0.844
 IV 204 (5%) 170 (5%) 34 (24%) <0.000
Main procedure
 CABG 1333 (35%) 1310 (36%) 23 (17%) <0.000
 AVR 636 (17%) 619 (17%) 17 (12%) 0.141
 CABG + AVR 324 (9%) 305 (8%) 19 (14%) 0.029
 MVR 254 (7%) 258 (7%) 6 (4%) <0.000
 Other 1235 (32%) 1151 (32%) 74 (53%) <0.000
Ejection fraction
 >50% 2293 (61%) 2238 (61%) 55 (40%) <0.000
 31 to 50% 893 (24%) 853 (23%) 40 (29%) 0.144
 21 to 30% 215 (6%) 205 (6%) 10 (7%) 0.434
 ≤20% 67 (2%) 57 (2%) 10 (7%) <0.000
Endocarditis 121 (3%) 103 (3%) 18 (13%) <0.000
Myocardial infarction 387 (10%) 357 (10%) 30 (22%) <0.000
CCS angina 101 (3%) 95 (3%) 6 (4%) 0.220
Chronic pulmonary disease 343 (9%) 330 (9%) 13 (9%) 0.906
Elective 3165 (84%) 3310 (91%) 126 (91%) <0.000
Emergency 20 (1%) 10 (0%) 10 (7%) <0.000
Arteriopathy 796 (21%) 765 (21%) 31 (22%) 0.712
Diabetes 205 (5%) 195 (5%) 10 (7%) 0.347
Previous cardiac surgery 248 (7%) 221 (6%) 27 (19%) <0.000
Critical condition 162 (4%) 138 (4%) 24 (17%) <0.000
Aortic surgery 413 (11%) 387 (11%) 26 (19%) 0.002
Pulmonary hypertension 479 (13%) 458 (13%) 21 (15%) 0.361
Renal impairment
 Normal 1316 (35%) 1286 (35%) 30 (22%) <0.000
 Moderate 1680 (44%) 1628 (45%) 52 (37%) 0.090
 Severe 713 (19%) 665 (18%) 48 (35%) <0.000
 Dialysis 69 (2%) 60 (2%) 9 (6%) <0.000
Descriptive parameters for all patients, patients of the RBCmod (negative class cases), and RBC10 (positive class cases) group. Data are shown as mean ± SD or n (%). All descriptive statistics were computed from nonmissing values before imputation. Significance level alpha was set at 0.05. Category ‘Other’ in main procedures includes AVR/MVR combined, AVR/mitral valve repair combined, transcatheter aortic valve implantation (TAVI), MVR/CABG combined, MVR/tricuspid valve repair combined, transcatheter mitral valve implantation (TMVI), mitral valve repair, VAD surgery. ASA, American Society of Anesthesiologists classification; AVR, aortic valve replacement; CABG, coronary artery bypass grafting; CCS, Canadian Cardiovascular Society grading of angina pectoris; MVR, mitral valve replacement; NYHA, New York Heart Association classification; pRBC, packed red blood cells; RBC10, 10 or more red blood cell packs transfusion group.

F1
Fig. 1:
Relative cumulative sum of blood transfusions vs. relative cumulative sum of patients.

Patients who received 10 or more pRBCs during their hospital stay tended to have a higher EuroSCORE II, a higher New York Heart Association (NYHA) classification, and a higher ASA score. Furthermore, these patients tended to have lower ejection fractions, lower Hb values, higher creatinine values, a higher C-reactive protein (CRP) value, and more preexisting diseases (e.g. endocarditis, myocardial infarction, diabetes, renal impairment, etc.). Furthermore, they underwent urgent or emergency surgery in a critical condition more often, and proportionally more were redo cases. There was an excess mortality in the RBC10 group (39%) versus only 4% in the RBCmod group.

Using all features present before surgery, we could predict patients requiring transfusion of less than 10 pRBC units during their hospital stay with satisfactory AUC using the five different machine learning techniques (Table 3 and Fig. 2). The highest AUCs were achieved with the Random Forest and GBM (AUC 0.810 for both models). The other three models all had lower AUC values, although none of them showed a confidence interval of 0.5 for the AUC, indicating that prediction of MPABT is possible with all models.

Table 3 - Model results
PPV NPV AUC ROC AUC F1 score
All features used
 RF 0.110 (0.09 to 0.13) 0.987 (0.98 to 0.99) 0.810 (0.76 to 0.86) 0.150
 NN 0.065 (0.05 to 0.08) 0.977 (0.97 to 0.98) 0.680 (0.62 to 0.74) 0.070
 GBM 0.126 (0.1 to 0.16) 0.985 (0.98 to 0.99) 0.810 (0.76 to 0.86) 0.180
 ADA 0.099 (0.08 to 0.12) 0.987 (0.98 to 0.99) 0.790 (0.74 to 0.84) 0.100
 LR 0.090 (0.07 to 0.11) 0.979 (0.97 to 0.98) 0.660 (0.6 to 0.72) 0.100
Eight most important features used
 RF 0.077 (0.06 to 0.09) 0.990 (0.98 to 0.99) 0.800 (0.75 to 0.85) 0.170
 NN 0.091 (0.07 to 0.11) 0.987 (0.98 to 0.99) 0.750 (0.69 to 0.81) 0.100
 GBM 0.110 (0.09 to 0.14) 0.984 (0.98 to 0.99) 0.800 (0.75 to 0.85) 0.180
 ADA 0.081 (0.06 to 0.1) 0.982 (0.98 to 0.99) 0.730 (0.68 to 0.78) 0.120
 LR 0.091 (0.07 to 0.11) 0.983 (0.98 to 0.99) 0.750 (0.7 to 0.8) 0.170
Statistical parameters of the classification of RBCmod versus RBC10 by different models.ADA, adaptive boosting; AUC, area under the receiver operating characteristic curve; F1, harmonic mean of precision and recall; LR, logistic regression; NN, artificial neural network, GBM, gradient boosting machine; NPV, negative-predictive value; PPV, positive-predictive value; RBC10, 10 or more red blood cell transfusion group; RBCmod, less than 10 red blood cell transfusion group; RF, Random Forest; ROC, receiver operating curve.

F2
Fig. 2:
Classification of RBCmod and RBC10 using all available features.

The most important preoperative data were the EuroSCORE II, urgency of surgery, the ASA score, the haemoglobin concentration, age, assist device surgery, the creatinine level and the glomerular filtration rate. Using solely these eight features, we trained the five different models again. As before, the best AUCs were yielded by the Random Forest and the GBM methods (AUC 0.800 for both models). Notably these AUCs were comparable to that achieved with all features (Table 3 and Fig. 3). This may indicate that the majority of variance necessary for prediction is contained in these features. In general, the predictions made by the 100 experienced colleagues, using the same features as the machine learning tools, achieved inferior results (Fig. 4).

F3
Fig. 3:
Classification of RBCmod and RBC10 using only the eight most important features (EuroSCORE II, urgency of surgery, the ASA score, the haemoglobin concentration, age, assist device surgery, the creatinine level and the glomerular filtration rate).
F4
Fig. 4:
Receiver operating characteristic curve of the best machine learning model using Random Forests and results of human specialists performing the same classification with the same input data.

For all analyses performed, the F1 score was relatively low, ranging from 0.070 to 0.180. This can be explained by the combination of a low PPV, ranging from 0.065 to 0.110, and a high NPV, ranging from 0.977 to 0.990 for all models.

Discussion

We were able to demonstrate that more than 50% of the allogeneic blood used in cardiac surgery is administered to less than 5% of patients (Fig. 1). Comparing the RBC10 group with the RBCmod group, MPABT was associated with a nearly 10-fold increase in mortality. It was a cornerstone of our study to identify these patients at risk of excess mortality before surgery. It may be that MPABT reflects a specific group in cardiac surgery bleeding classification.28

Timely identification of patients prone to MPABT might provide an opportunity to reduce severe perioperative bleeding by combining a targeted PBM, a meticulous surgical technique, autologous autotransfusion and a dedicated point-of-care coagulation management. This could lead to the reduction of allogeneic blood transfusions. Even so, it is a weakness of our study that blood loss is unknown, and therefore, should be investigated further.

To identify patients not in the RBC10 group before surgery, it is necessary to use features that are available at hospital admission. Those with the highest number of perioperative transfusions had a higher EuroSCORE II, NYHA classification and ASA scores. They also tended to have lower ejection fractions and preoperative Hb values, higher creatinine values and preoperative CRP values, and more preexisting diseases. One might speculate that those with multiple preexisting conditions have the highest risk of perioperative anaemia, bleeding and transfusion. Although this is in line with the published evidence, this alone is insufficient to guide clinicians in avoiding MPABT. This is clearly demonstrated by the inferior accuracy of the 100 specialist anaesthesiologists when compared with the machine learning algorithms. The RBC10 group is heterogeneous and difficult to identify by preexisting diseases.

We used five machine learning techniques (logistic regression, Random Forest, neural network, GBM, ADA) to distinguish between RBCmod and RBC10 at the time of hospital admission with Random Forest and GBM attaining the best AUC of 0.81. Even if all but the eight most important features were discarded, machine learning still enabled correct categorisation with an AUC for both Random Forest and GBM of 0.80. Due to asymmetry of the underlying problems (more than 96% of the patients are in the negative class RBCmod group) an AUC of greater than 0.8 alone does not guarantee that correct categorisation will be possible in clinical practice. Although some methods exist to address the common problem of asymmetry in medical datasets, for example, artificially inflating proportions of under-represented classes via oversampling or undersampling, these methods also introduce imprecision that again may result in skewed results. We, therefore, opted to refrain from including resampling methods. However, another approach, namely ensemble learning was used throughout the model training by using Random Forest and AB. These methods construct multiple weak classifiers that are combined in a single strong classifier.

Rare events are hard to predict – it would be a target to increase prediction confidence for the final model evaluation in a prospective study. That way, the problem of asymmetry may be addressed and alleviated better.29 Our results yielded a high NPV of up to 0.99 but very low PPV of 0.07 to 0.11 resulting in low F1 scores. The best machine learning model outperformed logistic regression, the state-of-the-art statistical method for binary classification with an AUC of 0.81 (Random Forest) versus 0.66 (logistic regression), highlighting the advantage of using machine learning over more traditional statistical techniques.

It is a conceivable limitation that because of changes in medical practice and techniques, the performance of the models may be limited as regards application over a time period greatly differing from the one used to train the model. Similarly, it may be that over the period, during which patient data was obtained, practice and procedures changed with changing guidelines and recommendations. Furthermore, variability in surgeons and anesthesiologists with respect to their blood loss/transfusion behaviour need to be considered as limiting factors that could not be analysed more closely because of workers’ council regulations.

A trivial classifier: “patient will not be in the RBC10 group” would result in a NPV of 0.963, which seems to be similar to what we obtained (0.990) but is actually much lower. In the best case, we were able to reduce the number of misclassifications of a trivial classifier by a third.

For daily clinical practice, this means that if our algorithms exclude a patient from the RBC10 group, no specific precautions will be necessary as the error rate will be only 1%

In a previous study, we predicted MPABT in a large (n = 206 271), mixed (conservative, surgical) cohort admitted to three different hospitals.16 Only a small number of specific features were used at the time of hospital admission. Although we had very high AUCs greater than 0.9 in this model for MPABT, sensitivity was low. MPABT only occurred in 0.5%, and therefore, the relative improvement of the NPV above a trivial classifier (0.997 versus 0.995) was lower than that which was obtained in our study (0.963 versus 0.990), indicating that our model is superior in our specific setting.16

We obtained the best results using Random Forest and GBM using a random hyperparameter search provided by the caret package of R.30 We cannot exclude that a more extensive hyperparameter search, another neural network topology or other machine learning techniques might yield better results. As none of the models outperformed the others by a lot, one might speculate that there is not too much room for improvement using other approaches.

In our study, we could prove that with a manageable number of features it is possible to identify patients at low risk of MPABT in cardiac surgery. We do not know, whether there are specific measures that could guarantee a reduction in MPABT in the perioperative course. Large observational studies and a number of randomised controlled trials across different surgical populations have demonstrated that the combination of anaemia treatment, a meticulous surgical technique, coagulation management and perioperative cell salvage are cornerstones of patient blood management, and further studies have to show whether this is also true in those at risk of MPABT.31 The features that enable prediction of MPABT are not necessarily the ones that can be modified by the interdisciplinary treatment team, rather they are indicators for emerging MPABT, and should therefore, not be used as causal agents. Although it would be a helpful result to identify predictive features that are modifiable, this is not necessary to prevent MPABT. If the features used to predict MPABT were excluded from optimisation, we surmise that there are other parameters that may add to the prevention of MPABT. Even though feature importance explain, which elements are relevant to our models’ predictions, a causal connection between features and outcomes cannot be established from them. It should be the target of further studies to evaluate, which patient parameters to optimise in what way after positive prediction of MPABT.

After external validation, the performances of our models may have the potential to be used in a multidisciplinary setting as a decision support tool.

Our models demonstrate that prediction of massive transfusion not occurring in cardiac surgery is possible. This gives an opportunity for the treatment team to prepare for this situation in those patients not identified as low risk. In those at risk of massive transfusion, timing of the surgical procedure is key. Whereas in daily clinical practice, one is reluctant to postpone surgery for adequate treatment with iron/erythropoietin, it is obvious that this should be a prerequisite in this high-risk patient group. Furthermore, the surgical and anaesthesiological team should include the most experienced physicians and nurses who have specific knowledge of cell salvage, point of care coagulation tests and meticulous coagulopathy management, and so forth. Discussion and shared decision-making, taking patient preferences into account, could also take place as to whether alternative treatments might be a better option in these patients.32

Conclusion

We conclude that machine learning algorithms enable stratification of patients into low and high risk for MPABT by detecting patients at low risk. This could enable initiation of a shared decision-making process for patients not at low risk: whether postponement or cancellation of surgery is an option. Furthermore, it could help to identify, which patients might profit most from an extended preoperative preparation with iron/erythropoietin, from an experienced treatment team with specific domain knowledge of cell salvage, point of care coagulation tests and meticulous coagulopathy management.

Acknowledgements relating to this article

Assistance with the article: none.

Financial support and sponsorship: none.

Conflicts of interest: none.

Presentation: none.

References

1. Geissler RG, Rotering H, Buddendick H, et al. Utilisation of blood components in cardiac surgery: a single-centre retrospective analysis with regard to diagnosis-related procedures. Transfus Med Hemother 2015; 42:75–82.
2. Shander A, Moskowitz D, Rijhwani TS. The safety and efficacy of ‘bloodless’ cardiac surgery. Semin Cardiothorac Vasc Anesth 2005; 9:53–63.
3. Shander A, Bracey AW, Goodnough LT, et al. Patient blood management as standard of care. Anesth Analg 2016; 123:1051–1053.
4. Meybohm P, Westphal S, Ravn HB, et al. Perioperative anemia management as part of PBM in cardiac surgery - a narrative updated review. J Cardiothorac Vasc Anesth 2020; 34:1060–1073.
5. Leahy MF, Hofmann A, Towler S, et al. Improved outcomes and reduced costs associated with a health-system-wide patient blood management program: a retrospective observational study in four major adult tertiary-care hospitals. Transfusion 2017; 57:1347–1358.
6. Boer C, Meesters MI, Milojevic M, et al. 2017 EACTS/EACTA Guidelines on patient blood management for adult cardiac surgery. J Cardiothorac Vasc Anesth 2018; 32:88–120.
7. Baron DM, Metnitz PGH, Fellinger T, et al. Evaluation of clinical practice in perioperative patient blood management. Brit J Anaesth 2016; 117:610–616.
8. Ranucci M, Baryshnikova E, Castelvecchio S, et al. Surgical and Clinical Outcome Research (SCORE) Group. Major bleeding, transfusions, and anemia: the deadly triad of cardiac surgery. Ann Thorac Surg 2013; 96:478–485.
9. Leff J, Romano CA, Gilbert S, et al. Validation Study of the Transfusion Risk and Clinical Knowledge (TRACK) Tool in cardiac surgery patients: a retrospective analysis. J Cardiothorac Vasc Anesth 2019; 33:2669–2675.
10. Tinmouth AT. The value of a clinical prediction rule for allogeneic transfusion in cardiac surgery. Transfusion 2006; 46:1072–1074.
11. Al-Khabori M, Al-Riyami AZ, Mukaddirov M, et al. Transfusion indication predictive score: a proposed risk stratification score for perioperative red blood cell transfusion in cardiac surgery. Vox Sanguinis 2014; 107:269–275.
12. Klein AA, Collier T, Yeates J, et al. Contributors. The ACTA PORT-score for predicting perioperative risk of blood transfusion for adult cardiac surgery. Br J Anaesth 2017; 119:394–401.
13. Madhu Krishna N, Nagaraja P, Singh N, et al. Evaluation of risk scores in predicting perioperative blood transfusions in adult cardiac surgery. Ann Cardiac Anaesth 2019; 22:73–78.
14. Alghamdi AA, Davis A, Brister S, et al. Development and validation of Transfusion Risk Understanding Scoring Tool (TRUST) to stratify cardiac surgery patients according to their blood transfusion needs. Transfusion 2006; 46:1120–1129.
15. Bartoszko J, Wijeysundera DN, Karkouti K, et al. Transfusion Avoidance in Cardiac Surgery Study Investigators. Comparison of two major perioperative bleeding scores for cardiac surgery trials: universal definition of perioperative bleeding in cardiac surgery and European Coronary Artery Bypass Grafting Bleeding Severity Grade. Anesthesiology 2018; 129:1092–1100.
16. Mitterecker A, Hofmann A, Trentino KM, et al. Machine learning-based prediction of transfusion. Transfusion 2020; 60:1977–1986.
17. von Elm E, Altman DG, Egger M, et al. STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol 2008; 61:344–349.
18. Anon. Austrian Inpatient Quality Indicators (A-IQI). Available at: https://www.sozialministerium.at/Themen/Gesundheit/Gesundheitssystem/Gesundheitssystem-und-Qualitaetssicherung/Ergebnisqualitaetsmessung.html. Accessed date: 25.06.2021
19. Nashef SAM, Roques F, Sharples LD, et al. EuroSCORE II. Eur J Cardio-Thorac Surg 2012; 41:734–745.
20. McQuilten ZK, Flint AW, Green L, et al. Epidemiology of massive transfusion - a common intervention in need of a definition. Transfus Med Rev 2021; 35:73–79.
21. Anon. [1701.05305] Random Forest Missing Data Algorithms. Available at: https://arxiv.org/abs/1701.05305. [Accessed 17 October 2020]
22. Stekhoven DJ, Buhlmann P. MissForest--nonparametric missing value imputation for mixed-type data. Bioinformatics 2012; 28:112–118.
23. Claesen M, De Moor B. Hyperparameter Search in Machine Learning. arXiv:1502.02127 [cs, stat] 2015.
24. Core R. Team R. A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing; 2013.
25. Ghojogh B, Crowley M. The theory behind overfitting, cross validation, regularization, bagging, and boosting: Tutorial. arXiv:1905.12787 [cs, stat] 2019.
26. Forman G, Scholz M. Apples-to-apples in cross-validation studies: pitfalls in classifier performance measurement. ACM SIGKDD Explor Newslett 2010; 12:49–57.
27. Kursa MB, Rudnicki WR. Feature selection with the Boruta Package. J Stat Software 2010; 36:1–11.
28. Dyke C, Aronson S, Dietrich W, et al. Universal definition of perioperative bleeding in adult cardiac surgery. J Thorac Cardiovasc Surg 2014; 147:1458.e1–1463.e1.
29. Fujiwara K, Huang Y, Hori K, et al. Over- and under-sampling approach for extremely imbalanced and small minority data problem in health record analysis. Front Public Health 2020; 8:178.
30. Kuhn M. Building predictive models in R using the caret Package. J Stat Software 2008; 28:1–26.
31. Meybohm P, Froessler B, Goodnough LT, et al. ‘Simplified International Recommendations for the Implementation of Patient Blood Management’ (SIR4PBM). Perioper Med (Lond) 2017; 6:5.
32. Mihalj M, Carrel T, Urman RD, et al. Recommendations for preoperative assessment and shared decision-making in cardiac surgery. Curr Anesthesiol Rep 2020; 10:185–195.
Copyright © 2022 European Society of Anaesthesiology and Intensive Care. Unauthorized reproduction of this article is prohibited.