Table 5.

Performance of the LSTM models with max pooling and word embedding trained on MIMIC-III corpus. The segment LSTM max (pooling) best performance was attained with 100 hidden units and pad size 7 for TrP relations, with 200 hidden units and pad size 5 for TeP relations, with 100 hidden units and pad size 5 for PP relations. The sentence LSTM max (pooling) best performance was attained with 150 hidden units for all relations.

System	Problem-Treatment (TrP) Relations			Problem-Test (TeP) Relations			Problem-Problem (PP) Relations
	R	P	F	R	P	F	R	P	F
Segment LSTM max	0.636	0.674	0.655	0.765	0.853	0.806	0.729	0.669	0.698
Sentence LSTM max	0.632	0.650	0.641	0.757	0.793	0.775	0.776	0.666	0.717