. Author manuscript; available in PMC: 2022 Jan 1.

Published in final edited form as: J Biomed Inform. 2020 Dec 15;113:103665. doi: 10.1016/j.jbi.2020.103665

Table 3.

Model Performance on Multi-Labeling Task

Parameter	50-D	100-D	200-D	300-D
Exact Match Accuracy (EMA)
EMA_Radiopaedia	0.358	0.440	0.476	0.487
EMA_WG	0.258	0.380	0.438	0.467
EMA_Radiopaedia - EMA_WG^‡	0.100 ^±0.00546	0.060 ^±0.00745	0.038 ^±0.00740	0.020 ^±0.00538
p-value	< 0.001^*	< 0.001^*	0.003^*	0.013^*
Hamming Loss (HL)
HL_Radiopaedia	0.0495	0.0436	0.0402	0.0393
HL_WG	0.0598	0.0494	0.0442	0.0425
HL_Radiopaedia- HL_WG^‡	−0.0103 ^±0.000579	−0.0058 ^±0.000693	−0.0040 ^±0.000595	−0.0032 ^±0.00110
p-value	< 0.001^*	< 0.001^*	0.001^*	0.032^*

Displayed exact match accuracy and Hamming loss correspond to the first permutation of the first seed.

^‡

WG = Wikipedia 2014 + Gigaword 5 Embeddings

^±

errors are given as one standard deviation derived from the averaged 5×2 cross validation variance

indicates significant difference between Radiopaedia and WG