Interpretable Machine Learning-Based Predictive Modeling of Patient Outcomes Following Cardiac Surgery

Structured Abstract

Objective

The clinical applicability of machine learning predictions of patient outcomes following cardiac surgery remains unclear. We applied machine learning to predict patient outcomes associated with high morbidity and mortality after cardiac surgery and identified the importance of variables to the derived model’s performance.

Methods

We applied machine learning to the Society of Thoracic Surgeons Adult Cardiac Surgery Database to predict post-operative hemorrhage requiring re-operation, venous thromboembolism and stroke. We used permutation feature importance to identify variables important to model performance and a misclassification analysis to study the limitations of the model.

Results

The study dataset included 662,772 subjects who had cardiac surgery between 2015 and 2017 and 240 variables. Hemorrhage requiring re-operation, venous thromboembolism and stroke occurred in 2.9%, 1.2% and 2.0% of subjects respectively. The model performed remarkably well at predicting all three complications (AUC 0.92-0.97). Pre- and intra-operative variables were not important to model performance. Instead, performance for the prediction of all three outcomes was driven primarily by several post-operative variables, including known risk factors for the complications such as mechanical ventilation and new-onset of post-operative arrhythmias. Many of the post-operative variables important to model performance also increased the risk of subject misclassification, indicating internal validity.

Conclusions

A machine learning model accurately and reliably predicts patient outcomes following cardiac surgery. Post-operative, as opposed to pre- or intra-operative variables, are important to model performance. Interventions targeting this period including minimizing the duration of mechanical ventilation and early treatment of new-onset post-operative arrhythmias may help lower the risk of these complications.

Graphical Abstract

Central Picture. Machine Learning-Based Predictive Modeling of Patient Outcomes Following Cardiac Surgery

Introduction

A majority of the over 300,000 cardiac surgeries reported in 2019 to the Society of Thoracic Surgeons (STS) Adult Cardiac Surgery Database (ACSD), were performed using cardiopulmonary bypass (CPB).¹ The extracorporeal bypass circuit activates a systemic inflammatory response which contributes to post-operative coagulopathy and neurocognitive dysfunction, increasing the risk of hemorrhage requiring re-operation, venous thromboembolism (VTE) and stroke.^2–4 Each of these complications is an independent predictor of mortality following cardiac surgery and is associated with higher healthcare utilization and cost.⁵ Despite technological advances and improvements in peri-operative care, the rates of hemorrhage requiring re-operation and stroke after cardiac surgery have not improved.^6–9

Machine learning is able to handle and process large volumes and variety of clinical data to improve the quality and automaticity of pattern recognition, and has been used to predict patient outcomes after cardiac surgery.¹⁰ A recent meta-analysis of 15 studies revealed that machine learning models consistently and significantly outperform traditional models at predicting post-operative mortality.¹¹ The clinical relevance of this superior performance, however, remains unclear.¹² No machine learning models have been prospectively validated in cardiac surgery, raising questions about their generalizability, clinical feasibility and efficacy. We previously applied deep learning, an advanced subset of machine learning based on artificial neural networks to routinely collected electronic health record data to accurately and reliably predict patient outcomes, including hemorrhage requiring re-operation in real-time after cardiac surgery.¹³ The model performed well when externally validated and surpassed standard of care clinical reference tools using traditional biostatistics. In order to target specific factors for intervention, understanding which variables are important to a machine learning model’s predictive capacity is essential.

Modern machine learning models are often characterized as ‘black boxes’ because they are based on a network of billions of parameters not readily explicable or comprehensible to humans.¹⁴ Traditional biostatistical models in comparison can delineate associations simply, often as odds ratios. Permutation feature importance (PFI) dissects machine learning models to elucidate the relative importance of variables to the model, addressing this central interpretability problem which promotes clinician mistrust and is a barrier to adoption.¹⁵ This approach (i.e., “explainable artificial intelligence”) can provide valuable insight into clinical potential and identify risk factors to target for intervention as we tackle the logistical and regulatory challenges to the real-time application of machine learning in cardiac surgery.^12,15,16

Our study aim was to use PFI to identify variables and groups of variables important to the performance of a deep learning model developed to predict post-operative hemorrhage requiring re-operation, VTE and stroke after cardiac surgery requiring CPB (Figure 1). We hypothesized that PFI would identify clinically relevant factors that can be targeted to improve patient outcomes. Our secondary aim was a misclassification analysis to compare characteristics of patients correctly classified by our model to those misclassified to study the limitations of the model and to internally validate the PFI results.

Figure 1.

Study outline summarizing methods and results.

Methods

This study was approved by our Institutional Review Board (#413818 on 8/20/2018) and the STS Participant User File Research Program. We applied for and received data submitted to STS from participating centers using data collection form version 2.81. The STS ACSD is the largest adult cardiac surgery outcomes registry with data for over 600 variables categorized into pre-, intra-, and post-operative groups from over seven million cardiac surgeries performed by surgeons from around the world.¹

We merged the data files we received and performed an exploratory data analysis to inform data pre-processing. We systematically cleaned the resulting dataset to minimize the introduction of bias and maximize model performance.¹⁷ Subjects undergoing off-pump cardiac surgery not requiring CPB or missing any of the three patient outcomes of interest data were dropped. Overlapping variables and variables with values with less than one percent variance were dropped. The STS codebook and simple inference were used to determine whether missing data were truly missing or appropriately missing due to branching logic. Finally, variables missing more than five percent of values were dropped. Testing with higher thresholds (dropping less features) did not improve model performance. For this reason, missing data was not imputed.

We built a deep learning feed forward neural network using TensorFlow, a popular open-source platform for machine learning with an extensive library of machine learning tools. The model consisted of two fully connected layers separated by two hidden layers, a rectified linear activation function unit commonly used to introduce non-linearity and a dropout layer to randomly drop input variables at a rate of 20% during training to prevent overfitting (Figure 2). We randomly split the study dataset into discrete training (80%), testing (10%) and validation (10%) cohorts. Single-class modeling was used to predict the three patient outcomes. The model was sequentially trained and tested on the training and testing cohorts respectively, and the model hyperparameters tuned on the validation cohort to maximize model performance. Synthetic minority oversampling technique (SMOTE) was used to overcome the severe class imbalance created by the low prevalence of the three patient outcomes of interest, creating a balanced training cohort with an equal number of subjects with and without each of the outcomes of interest.¹⁸ The derived model was assessed using common performance metrics including discrimination (area under the receiver operating characteristic [AUC]) and calibration.¹⁹

Figure 2.

Schematic of deep learning feed forward neural network development

We defined our three a priori binary patient outcomes of interest using existing post-operative complication variables within the STS ACSD.¹ Hemorrhage requiring re-operation was defined as mediastinal hemorrhage with or without tamponade requiring re-operation, VTE was defined as deep venous thrombosis or pulmonary embolism, and stroke was defined as an ischemic, hemorrhagic or embolic stroke.

The model was trained using all variables and tested before and after all values for each variable or peri-operative group were randomly permuted in the testing cohort, one variable or peri-operative group at a time.²⁰ Permutation rendered the variable or group of variables uninformative. The values were permuted and tested 30 times to determine the mean and standard deviation change in the AUC before and after the permutation. The decrease in model performance after permutation was listed from high-to-low to create a rank list of variables and peri-operative groups important to model performance for each complication.

The distributions of the values of the variables in subjects always correctly classified were compared to the distributions for subjects always misclassified for the same complication (Figure 3). The distributions within each pair of distributions for each variable were compared and ranked by calculating the quadratic-chi distance and the variable means and standard deviations. A Pearson’s chi-squared test was used to determine statistical significance, which was defined as p <0.05.

Figure 3.

Schematic of misclassification analysis. The heatmaps depict the subsets of 300 subjects always correctly classified (red and green) versus subjects always misclassified (orange and yellow) by the deep learning feed forward neural network over 10 randomization runs. The characteristics (age) of the subjects always correctly classified versus subjects always misclassified were then compared

Results

The study dataset included 662,772 subjects across 1,080 centers who had cardiac surgery requiring CPB between 2015 and 2017, and 240 variables: 126 pre-operative, 67 intra-operative and 47 post-operative variables (Figure 4). Supplement Table A lists the characteristics of the subjects in the study dataset. The mean age of the subjects was 64.9 years ± 11.8, the majority were men (69.4%), white (83.5%) and non-Hispanic/Latino (93.3%). The most common cardiac surgery performed was coronary artery bypass graft surgery (65.9%). Outcomes of interest occurred in 2.9%, 1.2% and 2.0% of subjects for post-operative hemorrhage requiring re-operation, VTE and stroke respectively. New-onset post-operative atrial fibrillation was the most common (26.9%) complication.

Figure 4.

Data cleaning – Subjects | Variables

The deep learning model performed very well at predicting all three complications (Table 1, Figure 5). Backpropagation helped model performance improve over successive epochs and there was no evidence of model overfitting (Supplement Figure A). The model was best able to predict hemorrhage requiring re-operation with an AUC of 0.97 (95% CI 0.97 – 0.97), followed by VTE (AUC 0.92, 95% CI 0.92 – 0.92) and stroke (AUC 0.92, 95% CI 0.92 – 0.93). Model calibration was also highest for the prediction of hemorrhage requiring re-operation, c=0.85, followed by c=0.79 for VTE and c=0.75 for stroke. Notably, the model performed well at predicting all three complications for males and females, and all races and ethnicities (Supplement Table B).

Table 1.

Performance of the deep learning feed forward neural network

Post-operative Complication	Accuracy	Sensitivity	Specificity	PPV	NPV	F1-score	Discrimination AUC (95% CI)	Calibration
Hemorrhage requiring re-operation	0.93	0.87	0.93	0.25	0.99	0.39	0.97 (0.97 −0.97)	0.85
Venous thromboembolism	0.90	0.69	0.90	0.08	0.99	0.14	0.92 (0.92–0.92)	0.79
Stroke	0.87	0.77	0.87	0.11	0.99	0.19	0.92 (0.92–0.93)	0.75

AUC-area under the receiver operating characteristic, CI-confidence interval, NPV-negative predictive value, PPV-positive predictive value

Figure 5.

Performance of the deep learning feed forward neural network. Area under the receiver operating characteristic: Hemorrhage requiring re-operation 0.97 (95% CI 0.97–0.97), venous thromboembolism 0.92 (95% CI 0.92–0.92), stroke 0.92 (95% CI 0.92–0.93)

Permutation of pre- and intra-operative groups independently and collectively resulted in only a minor reduction in model performance for the prediction of all complications (Figure 6; Supplement Table C). In contrast, permutation of the post-operative group dramatically lowered model performance for all complications. This was confirmed by a peri-operative group ablation analysis by training the model only on the post-operative group (dropping pre- and intra-operative groups). Thus, post-operative variables contributed most to model performance, while pre- and intra-operative variables contributed minimally. This was also confirmed by variable permutation – prediction of all three complications was primarily informed by post-operative variables (Figure 7; Supplement Table D).

Figure 6.

Decrease in the performance of deep learning feed forward neural network with peri-operative group permutation feature importance for hemorrhage requiring re-operation

Figure 7.

Decrease in the performance of deep learning feed forward neural network with variable permutation feature importance for hemorrhage requiring re-operation (A), venous thromboembolism (B), and stroke (C)

Model prediction of all three complications was driven by just a few post-operative variables (Figure 7; Supplement Table D). Two of the top four variables most important for hemorrhage requiring re-operation were units of post-operative red blood cells and platelets, ranked first and fourth respectively, with fresh frozen plasma ranked eight. The same two variables were most important to model performance for VTE and stroke, post-operative hospital length of stay and new post-operative arrhythmia requiring implantable cardioverter-defibrillator, while the two variables new post-operative arrhythmia not requiring pacemaker and total duration of mechanical ventilation ranked fourth and sixth for the prediction of VTE, and third and fourth for the prediction of stroke. Some variables were important to model performance for the prediction of all three complications. New post-operative arrhythmia requiring implantable cardioverter-defibrillator was the second most important variable for the prediction of all three complications, and new post-operative arrhythmia not requiring pacemaker ranked fifth for hemorrhage requiring re-operation, fourth for VTE and third for stroke. The variable total duration of mechanical ventilation ranked highly for all three outcomes. Several binary variables were important with both values of the variable ranked high, and the more clinically significant value often ranked higher. For example, post-operative reintubation and not reintubated post-operatively ranked highly for hemorrhage requiring re-operation (9^th and 12^th respectively). No and post-operative hemorrhage/emboli due to anticoagulation were important for the VTE, with no post-operative hemorrhage/emboli due to anticoagulation ranked higher. Similarly, new and no new-onset of post-operative atrial fibrillation were important for VTE and stroke, with the former ranked higher for both complications.

There was little overall difference between subjects correctly classified and misclassified as the majority of variables had very short quadratic-chi distances for all three complications. (Supplement Table E and F). Misclassified subjects differed the most from those correctly classified by many of the post-operative variables important to model performance, particularly for hemorrhage requiring re-operation. Total duration of mechanical ventilation was important to model prediction for all three complications. It was the highest ranked variable for subjects misclassified as having all three complications, and ranked third and fourth for subjects misclassified as not having VTE and stroke respectively. Not surprisingly, many of the variables important to model performance for hemorrhage requiring re-operation were related to the transfusion of blood products. The same variables ranked highly for subjects misclassified as having and not having hemorrhage. There were more discrepancies among predictive model performance and variable quadratic chi-distance for VTE and stroke. New-onset of post-operative atrial fibrillation was important to model performance for the prediction of VTE and stroke but was less important to the model misclassifying subjects as having and not having these two complications.

Discussion

We developed a deep learning model to predict three significant post-operative complications after cardiac surgery requiring CPB. Overall, model performance was excellent. For the first time, we identified the relative importance of variables and groups of variables to model performance, and those that also increased subject misclassification. The post-operative period was the most critical for outcome prediction. Clinically targetable variables such as duration of mechanical ventilation and new onset post-operative arrhythmias emerged as important to model performance and subject misclassification regardless of subject demographics. To-date, machine learning has been successfully used to predict patient outcomes after cardiac surgery, but clinical relevance has remained questionable.^12,16 Our interpretability approach suggests studies should target the post-operative period, and specifically strategies to reduce the duration of mechanical ventilation and risk for arrythmias, for potential intervention.

Our model generalized well with a similar training performance, suggesting no overfitting. The model performed well across all performance metrics except positive predictive values and resulting F₁-scores. This is expected given the very low prevalence of all three complications, and was confirmed by re-training on a balanced cohort resulting in high positive predictive values and F₁-scores (Supplement Table G). While the cohort was predominantly male and white, performance was maintained regardless of subject sex, race and ethnicity. The results of the misclassification analysis revealed minor differences in characteristics between subjects who were correctly classified and those who were misclassified, providing internal validity for the model’s reliability and accuracy. Older subjects had a slightly higher likelihood of being misclassified.

Variable rank lists for the three complications only included a fraction of the 240 variables available. Further, permutation of a majority of the ranked variables resulted in a performance drop of almost zero or true zero – the majority of variables were not important to model performance. Our results suggest a more parsimonious collection of data in large registries could be employed, depending on the outcomes of interest. Model performance for all three outcomes was primarily driven by several post-operative variables. This included both values of some binary variables with the more clinically significant value often ranked higher (more important). This ‘clinical logic’ is promising and supports the notion that the rule-based logic of machine learning can mimic clinical reasoning. The presence of both values also mirrors how a clinician thinks, weighing both the presence and absence of risk factors to gauge cumulative risk and make clinical decisions. This also explains why misclassified subjects differed the most from those correctly classified by many of the post-operative variables important to model performance. The model seems to be relying heavily on these variables to make predictions which a majority of the time resulted in the correct prediction.

Our results suggest that the post-operative period after cardiac surgery is the most critical. This makes intuitive sense – a predictive model should ideally integrate data up to the occurrence of the outcome of interest, and exclude data following the outcome. This may be especially relevant for late complications which can be delayed in presentation after cardiac surgery and can have distinct risk factors.²¹ Delayed stroke after cardiac surgery is associated with new-onset of post-operative atrial fibrillation whereas early stroke is associated with intra-operative aortic manipulation.²² A majority of existing studies and risk calculators only include pre- and intra-operative variables, or even just pre-operative variables. The STS Short-term Risk Calculator estimates a patient’s risk of post-operative complications including stroke, and is based almost entirely on pre-operative variables.^6,23,24 While designed to estimate a patient’s risk of post-operative complications before surgery, our results suggest that the risk calculator potentially excludes important post-operative risk factors preceding the complication and underestimates risk. To this point, a validation study demonstrated that the risk calculator consistently underestimated post-operative mortality in two cohorts.²⁵ Despite advances in care, the risk of hemorrhage and stroke after cardiac surgery have not improved.^6–9 While our results require prospective validation, interventions specifically targeting the post-operative factors we identified may help reduce the risk of these complications.

Pre- and intra-operative variables were less important to model performance for all complications. This was confirmed by performing a peri-operative group ablation analysis, sequentially ablating (dropping) peri-operative groups of variables with similar results. Data for patients not offered cardiac surgery because they are too high risk are not included in the STS ACSD. Thus, the pre-operative group of variables may be less important to a model predicting post-operative complications because of selection bias. Further, patients who undergo cardiac surgery are also, whenever possible, routinely optimized before surgery.²⁶ As our cohort is fairly modern (2015 – 2017), current approaches to pre-operative care based on risk factors identified in historical cardiac surgery cohorts may be adequate to minimize post-operative complications.^23,24 Including data on the elective or emergent nature of the surgery would help ascertain if optimization for surgery impacted the importance of peri-operative factors, however values for this variable (status) were missing too often in our study dataset for this secondary analysis to be performed. Nevertheless, given STS is the largest available registry and the multiple approaches we used to assess validity, we submit our results are generalizable to adults who undergo cardiac surgery and require CPB.

Some of the variables important to model performance are known risk factors for our post-operative complications of interest, adding face validity. Several variables which capture CPB duration and left ventricular function contributed to our VTE and stroke models and have been previously linked to these outcomes in observational studies.^27–29 Cardiac arrhythmias including atrial fibrillation are common after cardiac surgery and increase the risk of post-operative hemorrhage and VTE.³⁰ Atrial fibrillation increases the risk for stroke two-to-four fold including during the post-operative period. New-onset of post-operative atrial fibrillation improved the prediction of VTE and stroke, and the related variables post-operative arrhythmia requiring or not requiring a pacemaker and implantable cardioverter-defibrillator improved the prediction of all three complications in our study. Variables related to mechanical ventilation contributed to model performance for all three complications. Prolonged mechanical ventilation >12 hours is known to increase morbidity and mortality; early extubation within six hours is associated with a lower rate of post-operative complications after cardiac surgery.^31,32 While the risks of mechanical ventilation may be attributable to related factors including immobilization (which increases the risk of VTE) and prolonged exposure to sedatives and analgesics, mechanical ventilation in-and-of-itself may lead to operative complications.^33,34 It is noteworthy that there was a consistent signal between variables related to mechanical ventilation and hemorrhage requiring re-operation, VTE and stroke. In experimental models, mechanical ventilation alone causes systemic coagulopathy and cerebral dysfunction.^35,36 Preventative strategies and interventions targeting new-onset post-operative arrythmias and liberation from mechanical ventilation after cardiac surgery may improve surgical outcomes.

Our study has several strengths. It marks the first time this interpretability approach and misclassification analysis have been used to analyze a deep learning model developed to predict post-operative complications after cardiac surgery in the largest available dataset to-date. We intentionally chose the multicenter, multinational STS ACSD with its large number of variables and outcomes standardized across centers and categorized by STS to facilitate our methods. The resulting deep learning model performed remarkably well across all demographics. In addition to our PFI and ablation analysis, we also performed a Shapley Additive explanations (SHAP) analysis and identified remarkably similar features driving model performance (data not shown), further validating our findings.³⁷

Our study has limitations. First, the study dataset included a large number of derived variables (e.g., units of post-operative red blood cells), some of which were found to be important. This is likely a consequence of or directly related to the complications themselves, thus artificially boosting model performance via reverse causation, although this provides validation for the PFI findings. For example, the prediction of hemorrhage requiring re-operation was most informed by variables related to blood transfusion. Obviously, patients who have bleeding requiring re-operation are more likely to require post-operative blood products. Second, the STS ACSD lacks timed data and outcomes which would have allowed modeling using only data preceding the post-operative complications – including post-operative data preceding the complications – limiting reverse causation. Finally, PFI may artificially lower variable importance. When variables are permuted, the model may infer some or all of the information missing from related variables, potentially altering the rank order of the PFI rank lists.

Our study illustrates the potential of the interpretability approach applied to machine learning to predict important patient outcomes after cardiac surgery requiring CPB. Our results demonstrate that model performance for the prediction of hemorrhage requiring re-operation, VTE and stroke was driven primarily by several post-operative variables, notably post-operative arrythmias and mechanical ventilation. These key clinical factors could be targeted for future study and intervention to improve patient outcomes.

Glossary of Abbreviations

AUC

area under the receiver operating characteristic curve

ACSD

adult cardiac surgery database

CPB

cardiopulmonary bypass

PFI

permutation feature importance

STS

Society of Thoracic Surgeons

VTE

venous thromboembolism