Electronic medical record-based causal network modeling for acute myocardial infarction diagnosis in the emergency department

Figures

• 

  Figure 1. The flowchart to include the studied cohort EMR, electronic medical record; AMI, acute myocardial infarction.

• 

  Figure 2. Framework of the proposed method and detailed procedure of causal network identification (A) The framework of the proposed method, which can be mainly divided as causal network modeling, model validation, causal effect quantification, and sensitivity analysis on unobserved confounding bias. EMR, electronic medical record; AMI, acute myocardial infarction; FGES, fast greedy equivalence search; DANM, discrete nonlinear additive noise model. (B) The procedure for causal network identification with structural equation model (SEM) and network coarse-graining. Under the premise that no new V-structure or cycle generates, SEM is adopted to identify directions. If the direction cannot be identified or there must be new V-structure or cycle, network coarse-graining is performed to cover the unclear edges. MEC, Markov equivalence class. (C) An instance of network coarse-graining, in which: (a) Y to X introduces a new V-structure, and assume that structural equation model (SEM) cannot identify X to Y; (b) Integrate X and Y as a macro variable, VE. The states of VE are the combinations of the states of X and Y. (c) Rebuild the causality between Z and VE.

• 

  Figure 3. MEC for the onset of AMI in the emergency department Each node represents a variable. Arrows between nodes are the discovered causalities, and the red edges are the unclear associations. MEC, Markov equivalence class; AMI, acute myocardial infarction. Full name of variables in the network can be referred to in Table 2.

• 

  Figure 4. Causal network for the onset of AMI in the emergency department The variables in the Markov boundary of acute myocardial infarction (AMI) are highlighted, as well as the coarse-grained macro variable lifestyle. The causality identified by the discrete nonlinear additive noise model (DANM) is remarked. Full name of variables in the network can be referred to in Table 2.

• 

  Figure 5. Schematic of the variables included in the testing set for each validation (A) The testing set includes all variables except the onset of acute myocardial infarction (AMI). (B) The testing set only includes the Markov boundary of AMI. (C) The testing set only includes the complementary set of the Markov boundary, also excluding AMI.

• 

  Figure 6. Sensitivity analysis for the variables with EI(X→AMI) > 0 Each subfigure is the Austen plot to test the potential unobserved confounding effect on one specific causal relationship marked in the subfigure’s title. The dots in an Austen plot present the effects of observed covariates. The curve indicates the effect of potential unobserved confounding bias. If the dots are closer to zero than the curve, it suggests that the causal relationship is robust. AMI, acute myocardial infarction. Full name of variables in subfigures can be referred to in Table 2.

• 

  Figure 7. Sensitivity analysis for the variables with EI(AMI→X) > 0 Each subfigure is the Austen plot to test the potential unobserved confounding effect on one specific causal relationship marked in the subfigure’s title. The dots in an Austen plot present the effects of observed covariates. The curve indicates the effect of potential unobserved confounding bias. If the dots are closer to zero than the curve, it suggests that the causal relationship is robust. AMI, acute myocardial infarction. Full name of variables in subfigures can be referred to in Table 2.

• 

  Figure 8. The flowchart to include the studied cohort of MIMIC-IV-ED EMR, electronic medical record; AMI, acute myocardial infarction.

• 

  Figure 9. Causal network based on the cohort in MIMIC-IV-ED The causal network contains 5 variables, with all edges determined. The causalities related to numbness and systolic blood pressure (SBP) cannot be identified. ResRhy, respiratory rhythm.