Table 1.
Baseline characteristics of Gulf-COAST patients.
Table 2.
Cross-validation summary results for GBM, RF and SVM models.
Fig 1.
Important Features that Contribute to the Prediction of Three In-Hospital ACS Related Outcomes, Including (A) mortality, (B) heart failure, and (C) Bleeding. A classification error loss function (“ce”) was used to calculate feature importance. Black dots indicate median “ce,”. CVD: Cardiovascular disease. TIA: Transient ischemic attack.
Fig 2.
Centred Individual Conditional Expectation (ICE) plots for the top three important features that contribute to the prediction of three in-hospital ACS outcomes, including (A, D, G) mortality, (B, E, H) heart failure, and (C, F, I) bleeding. The plots show the relationship between the predicted risk of ACS outcomes and each corresponding feature. The black lines indicate the predicted risk in each patient, while the red line indicates the partial dependence calculated as the average risk across all patients.
Fig 3.
Feature interaction plots calculated using Friedman’s H-statistic.
(A-C) indicate the mortality model; (D-F) indicate the heart failure model; (G-I) indicate the bleeding model. The three plots on the top (A, D, G) showing the overall interaction strength of each feature with the other features. Plots (B & E) demonstrates the overall interaction strength of hemoglobin with the other features, while plot (H) demonstrate the overall interaction strength of chronic renal failure with the other features. Partial dependence plots at the bottom (C, F, I) represent the top interactions that shaped the risk of the three acute coronary syndromes outcomes. (C) interaction between age and initial hemoglobin for the mortality model. The heat matrix corresponds to the risk of death, in which lighter shades of red indicate lower risks of death, and darker shades of reds indicate higher risks of death. The bar on the right indicates (y hat) the relative risk of being dead with all other feature combinations marginalized. (F) interaction between sex and initial hemoglobin for the heart failure model. (I) interaction between chronic renal failure initial hemoglobin for the bleeding model. Red and green partial dependence curves represent different sexes (F) and the status of chronic renal failure (I), while the y-axis indicates the predicted risk of heart failure or bleeding. CVD: Cardiovascular disease. TIA: Transient ischemic attack.
Fig 4.
Value contributions for the respective risk of acute coronary syndromes (ACS) based on Shapley Values (φ) for Six Individual Patients.
(A) a patient discharged dead; (B) a patient with an in-hospital heart failure; (C) a patient with an in-hospital bleeding; (D) a patient discharged alive; (E) a patient discharged alive with no in-hospital heart failure; and (F) a patient discharged alive with no in-hospital bleeding. Red bars indicate positive outcomes, and blue bars indicate negative outcomes. Positive Shapely value indicates that this feature increased the risk of ACS outcome, whereas negative values indicate that this feature decreased the risk of ACS outcome. The values next to each feature name indicate the observed value of that feature for that patient. *CVD: Cardiovascular disease; *TIA: Transient ischemic attack.