Table 2.

Statistics of Clinical Note Datasets with the Natural Language Processing Pipeline and Manual Annotations

	MIMIC-III Disch	MIMIC-III Rad	Tayside Brain Img
|T|	59,652	522,279	156,618
|D|	127,150	109,096	7,761
$|D_{wea{k}^{+}}|$	15,598	13,907	1,137
$|D_{wea{k}^{-}}|$	74,217	65,171	2,898
$|T_{\mathit{RD}}|$	37,110	73,589	7,321
$|T_{\mathit{RD}}^{\mathit{weak}}|$	10,568	21,102	2,855
$|T^{\mathit{ann}}|$	500	1,000	5,000
$|D^{\mathit{ann}}|$	1,073	198	279+4
$|T_{\mathit{RD}}^{\mathit{ann}}|$	312	145	273

|T|, number of documents; |D|, number of mention-UMLS pairs; $|D_{wea{k}^{+}}|$, $|D_{wea{k}^{-}}|$, number of weakly labelled positive and negative mention-UMLS pairs, respectively; $|T_{\mathit{RD}}|$, $|T_{\mathit{RD}}^{\mathit{weak}}|$, number of documents associated with one or more rare diseases detected by SemEHR and SemEHR+WS (i.e. further with weak supervision), respectively; $|T^{\mathit{ann}}|$, $|D^{\mathit{ann}}|$, $|T_{\mathit{RD}}^{\mathit{ann}}|$, number of documents sampled, number of mention-UMLS pairs sampled, and number of the sampled documents with one or more rare diseases identified by SemEHR, respectively. For Tayside data, 4 new positive mention-UMLS pairs in $|D_{\mathit{ann}}|$ were identified from the reports during the manual annotation