Table 4 External validation.

From: A novel method for causal structure discovery from EHR data and its application to type-2 diabetes mellitus

Edge

Discovery %

Reason

MC

FV

HDL → Trigl

0

91.7

There is no clear precedence relationship, the two events often coincide

HTN.dx → CRF

88.5

0.1

 

Trigl → DM.dx

100

0

 

Trigl → HL.tx

100

0

 

LDL → HL.dx

72.1

0

 

FPG.125 → DM.dx

100

0

FV uses A1c, not FPG

Trigl → FPG.125

99.5

0.2

 

DBP → HTN.tx

91.5

0

The criteria for diagnosis and treatment are institution specific

SBP → HL.tx

99.3

1.7

 

SBP → HTN.tx

100

29.1

 

Trigl → HTN.tx

83.7

0

 

CHF → MI

0

67.6

SBP is a common cause for CHF and MI, but at FV, this effect was too weak in 68% of the bootstrap iterations

HL.dx → Trigl

0

87.6

While the main driver of Trigl is BMI, at FV, the diagnosis of HL helps explain the variation in Trigl

HL.tx → CAD

0

74.3

LDL drives both HL treatment and CAD

  1. The ‘.tx’ suffix denotes the treatment, and ‘.dx’ suffix denotes the diagnosis of the disease. The abbreviations of the diseases and lab tests can be found in Table 1. The table describes the edges that were discordant between the Mayo Clinic (MC) and M Health Fairview (FV) data sets. It shows the percentage of the bootstrap iterations in which the edge was discovered at MC and FV and a brief reason for the discrepancy.
Back to article page