Table 3 Comparison of NLP and ICD approaches for PPH subtyping.

Subtype	Prevalence (%)	Method	Sensitivity	PPV	Binary F1	Acc.	P-value
Tone	41.8	NLP	0.931	0.964	0.947	0.957	<0.001
		ICD	0.644	0.667	0.655	0.716
Tissue	38.9	NLP	0.827	0.957	0.887	0.918	<0.001
		ICD	0.370	0.909	0.526	0.740
Trauma	40.9	NLP	0.671	0.826	0.740	0.808	0.018
		ICD	0.576	0.690	0.628	0.721
Thrombin	6.3	NLP	0.385	0.714	0.500	0.952	1.0
		ICD	0.462	0.600	0.522	0.947

The true prevalence of each subtype and the NLP and ICD model performance are shown. Note that a single episode of postpartum hemorrhage can be classified with multiple subtypes. The model evaluation was performed on 285 notes with confirmed PPH via manual annotation. In 27.0% of the notes, the annotators were unable to determine a PPH subtype from the discharge summary alone (not shown). These notes were excluded to avoid inflation of false positives in the ICD model in cases where subtype information was captured elsewhere in the medical record.
NLP natural language processing, PPH postpartum hemorrhage, ICD international classification of diseases.
Bold values indicates the highest value of the NLP and ICD approaches for PPH subtyping.

Quick links

Search