Table 1 Performance of the zero-shot Flan-T5 model in identifying postpartum hemorrhage (PPH) concepts in discharge summaries.

From: Zero-shot interpretable phenotyping of postpartum hemorrhage using large language models

Concept (n)

Sensitivity

Specificity

PPV

Binary F1

Accuracy

Laceration (n = 421)

0.976

0.996

0.993

0.984

0.989

Bakri balloon (n = 46)

0.957

1.000

1.000

0.978

0.998

Cesarean delivery (n = 470)

0.983

0.970

0.957

0.970

0.975

Methylergonovine (n = 104)

0.952

0.996

0.961

0.957

0.992

Placenta previa (n = 170)

0.982

0.986

0.923

0.952

0.986

Fresh frozen plasma (n = 79)

0.962

0.995

0.938

0.950

0.993

Carboprost (n = 49)

0.898

0.999

0.978

0.936

0.995

Misoprostol as a uterotonic (n = 112)

0.938

0.991

0.913

0.925

0.986

Hysterectomy (n = 55)

0.982

0.993

0.871

0.923

0.992

O’Leary sutures (n = 12)

0.917

0.999

0.917

0.917

0.998

Uterine atony (n = 107)

0.850

0.999

0.989

0.915

0.986

Retained products of conception (n = 102)

0.892

0.993

0.929

0.910

0.985

Direct mention of PPH (n = 362)

0.881

0.963

0.914

0.897

0.938

Packed red blood cells (n = 165)

0.945

0.971

0.843

0.891

0.968

Abruption of the placenta (n = 47)

0.915

0.994

0.860

0.887

0.991

Coagulation disorders (n = 44)

0.795

0.996

0.897

0.843

0.989

Cryoprecipitate (n = 29)

1.000

0.990

0.725

0.841

0.991

Dilation and curettage (n = 122)

0.762

0.990

0.894

0.823

0.966

Platelets (n = 29)

1.000

0.986

0.644

0.784

0.986

Placenta accreta spectrum (n = 45)

0.822

0.987

0.712

0.763

0.980

Manual extraction of placenta (n = 110)

0.782

0.972

0.741

0.761

0.954

PPH due to surgical causes (n = 27)

0.741

0.983

0.513

0.606

0.978

Uterine rupture (n = 5)

1.000

0.992

0.357

0.526

0.992

Concept (n)

Sensitivity

PPV

Note Acc.

  

Estimated blood loss (n = 212)

0.745

0.905

0.951

  
  1. We assessed model performance on 1175 manually annotated notes with PPH ICD diagnostic codes. We report the prevalence of each concept in the annotated test set. We report model sensitivity, specificity, positive predictive value (PPV), binary F1 score, and accuracy for binary concepts, and we report sensitivity, PPV, and note-level accuracy in extracting the correct estimated blood loss values.
  2. PPH postpartum hemorrhage.