Skip to main content
. 2012 Sep 4;20(1):77–83. doi: 10.1136/amiajnl-2012-001020

Table 2.

Token-level evaluation results considering exact matches

VHA PHI categories MIT deid One-step CRF BoB rules BoB CRF BoB rules+CRF BoB full
R R R R R R P
Patient Name 0.724 0.956 0.977 0.962 0.994 0.985 0.642*
Relative Name 0.909 0.939 0.970 0.970 0.970 0.939
Healthcare Provider Name 0.747 0.925 0.938 0.916 0.965 0.943
Other Person Name 0.867 0.800 1 0.800 1 0.933
Street City 0.765 0.728 0.878 0.798 0.929 0.887 0.682
State Country 0.656 0.812 0.944 0.744 0.956 0.869 0.839
Deployment 0.177 0.859 0.934 0.934 0.869 0.915
ZIP Code 1 1 1 1 1 1
Healthcare Units 0.080 0.716 0.834 0.748 0.902 0.798 0.779
Other Organizations 0.098 0.503 0.798 0.596 0.880 0.721 0.606
Date 0.617 0.922 0.972 0.938 0.978 0.972 0.935
Age>89 0.250 0.500 1 1 1 0.8
Phone Number 0.565 0.810 0.991 0.991 0.939 1
Electronic Address 1 0.500 1 1 1 1
SSN 1 0.407 1 1 1 0.964
Other ID number 0.094 0.855 0.983 0.983 0.936 0.82
Overall macro-averaged 0.597 0.764 0.951 0.968 0.925 0.845
Overall micro-averaged
Precision 0.734 0.931 0.420 0.392 0.815
Recall 0.489 0.859 0.933 0.957 0.921
F1 measure 0.587 0.893 0.579 0.556 0.864
F2 measure 0.524 0.872 0.749 0.743 0.897

*BoB annotates all person names as one PHI category.

CRF, conditional random fields; P, precision; PHI, protected health information; R, recall; VHA, Veterans Health Administration.