Explainable machine-learning predictions for complications after pediatric congenital heart surgery

Figures

• 

  Figure 1. Schematic of data processing and workflow of our proposed model. We used soft-DTW to measure similarities of time-series hemodynamic data, and k-means clustering was applied to capture different dynamic patterns of hemodynamic data. Blood pressure variability was also used to measure blood pressure fluctuations within a period of surgery. All these extracted time-series features combined with the patient- and surgery-specific static features were used to build a predictive model of postoperative complications. An explanation was then built for each prediction. Pink features revealed that increased value is associated with an increased risk on the final prediction, whereas blue features decreased risk. The base value is the mean of the model output over the training dataset. The final prediction risk is the sum of the impacts of all features and base value, and then transformed into probability space; in this case, this patient has a 21% chance of having postoperative complications.

• 

  Figure 2. Clustering of intraoperative blood pressure and comparison of other characteristics between clusters. (a) Similarity matrix of blood pressure where rows and columns were ordered by a partition detected through k-means clustering. (b) Composition of each cluster in terms of whether the patients experienced postoperative complications, risk categories of operation, or primary diagnoses. (c) Distribution of surgical time, CPB time, or aortic cross-clamping time between each cluster. In the box plots, box edges represent the 25th and 75th percentiles, the centerline shows the median and whiskers extend from the box edges to the 1.5× interquartile range and *means P < 0.001 (Kruskal–Wallis H-test).

• 

  Figure 3. Receiver operating characteristic curve of the predictive models in the test data set. (a) Binary classification; (b–f) Multi-label classification. ABC Aristotle basic complexity, RACHS-1 risk adjustment for congenital heart surgery, STS Society of Thoracic Surgeons.

• 

  Figure 4. Feature inspection. (a,b) Importance estimates assigned by our proposed method to the top 15 features in both the binary (a) and multi-label (b) classification. The importance of features in multi-label classification was measured as the averaged feature importance estimated by five categories of complication prediction, and the feature importance of each category prediction was represented by different colored circles. (c,d) Scatter plots showing the relationship between the varying feature and SHAP value for CPB time (c) and patient weight (d). The gray histogram shows the distribution of values for that feature in the training set. The orange line and shade represent the mean and the 95% confidence interval of the regression line. (e) Visualization of two individual examples for the explanation risk of postoperative complications. Each feature contributes to pushing the model output from the base value (the average model output over the training dataset we passed) to the model output. Red features mean pushing the prediction higher risk and blue features pushing the prediction lower risk.