Repo for MedDRA coding results
BlazingText with Subword Models (BTSM) were built on MedDRA 4-22 & anonymized AE+MH data
-
100 dimension model
-
500 dimension model
-
100 dimention model with artificial augmentation
-
500 dimention model with artificial augmentation
| Methods | Loss | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|---|
| m0 500D BTSM + LSTM+Attension | 0.665 | 0.917 | 0.929 | 0.917 | 0.920 |
| m1 500D BTSM + 2BiLSTM Attension | 0.889 | 0.928 | 0.937 | 0.928 | 0.930 |
| m2 500D BTSM + Attension +LSTM | 0.940 | 0.924 | 0.930 | ||
| m3 100D BTSM + Attension +Conv+biLSTM | 0.610 | 0.935 | 0.944 | 0.935 | 0.935 |
| m0 100D BTSM + 2BiLSTM+Attension | 0.557 | 0.926 | 0.939 | 0.927 | 0.928 |
with MedDRA 4-22
- 500D BTSM + LSTM+Attension
- 500D BTSM + 2BiLSTM Attension
- 500D BTSM + Attension+LSTM
- 100D BTSM + Attension + 3 parallel conv +biLSTM +dense
Earlystop 'val_loss' min_delta=0 patientce 3
- 100D BTSM + BiLSTM+Attension
- 100D BTSM + LSTM+Dense+Attension
- 500D BTSM + LSTM+Attension
- 500D BTSM + Attension+LSTM
- 500D BTSM + 2BiLSTM+Attension
- 500D BTSM + LSTM+Dense+Attension
without MedDRA 4-22