Skip to main content
. 2022 Sep 19;5(3):e38211. doi: 10.2196/38211

Table 2.

Model performance (N=15).

Model Patients with >33% recall, n (%) Anomalies raised, mean Recall (%), mean Precision, %a
LODAb (w=7; IQR 1.2) 14 (93) 37.8 85.7 6.2
Sum of CMPc scores (w=7; quantile 0.97) 14 (93) 33.1 84.7 7.0
Mean of CMP scores (w=7; quantile 0.97) 14 (93) 33.1 84.7 7.0
Equal-weighted multidimensional CMP (w=7; k=1; robust z=1.65) 15 (100) 32.1 84.3 7.2
COPODd (w=7; quantile 0.95) 13 (87) 36.8 79.1 5.9
ABODe (w=21; quantile 0.95) 13 (87) 30.0 77.7 7.1
Distance-weighted multidimensional CMP (w=14; k=0; robust z=1.65) 14 (93) 33.7 76.7 6.2
ApEnf-weighted CMP scores (w=7; quantile 0.97) 12 (80) 29.1 69.9 6.8
Median of CMP scores (w=7; quantile 0.97) 12 (80) 30.8 68.4 6.1
Fuzzy entropy–weighted CMP scores (w=7; quantile 0.97) 10 (67) 27.7 65.5 6.5
Maximum of CMP scores (w=7; quantile 0.97) 10 (67) 24.8 57.9 6.4

aWe have mentioned previously that it is more meaningful in this context to look at relative precision across methods and not at absolute precision.

bLODA: Lightweight Online Detector of Anomalies.

cCMP: Contextual Matrix Profile.

dCOPOD: Copula-Based Outlier Detection.

eABOD: Angle-Based Outlier Detection.

fApEn: Approximate Entropy.