Multi-task transfer learning for the prediction of entity modifiers in clinical text: application to opioid use disorder case detection

Table 3 Model Performance on the OUD Corpus Modifiers

Model	Ep	Neg	Sub	DT	IDU	Unc	Avg
macro average F1
MT-SHR	-	0.769	0.863	-	-	0.542	-
MT-OUD	3	0.958	0.920	0.838	0.839	0.833	0.878
MT-OUD.fl	4	0.892	0.855	0.816	0.784	0.632	0.796
MT-BOTH	20	0.951	0.911	0.846	0.867	0.765	0.868
MT-SHR-OUD	6	0.948	0.915	0.877	0.848	0.772	0.873
ST-OUD^a	2,3	0.935	0.911	0.846	0.831	0.768	0.858
MT-OUD-no hint^a	3	0.912	0.880	0.739	0.689	0.721	0.788
micro average F1
MT-SHR	-	0.591	0.732	-	-	0.100	-
MT-OUD	3	0.912	0.842	0.776	0.688	0.546	0.753
MT-OUD.fl	4	0.803	0.714	0.729	0.583	0.267	0.619
MT-BOTH	20	0.912	0.825	0.781	0.743	0.533	0.759
MT-SHR-OUD	6	0.912	0.842	0.816	0.705	0.593	0.774

Bold font means the best performance
Underline means statistically significant p-value relative to the model in the previous line
Ep: epochs, Neg: negation, Sub: subject, Unc: uncertainty, DT: DocTime, IDU: Illicit Drug Use, Avg: average
^aAblation study results

ISSN: 2041-1480