專題報告11/19
MDD/NC
資料處理
SSS
Resample 256Hz
Filter 2-30Hz
Crop 4 sec Overlap 3 sec
Hanning Window
shape (306,1024)
資料一
| Train | Val | |
| MDD | 4260 | 1000 |
| NC | 3613 | 1000 |
訓練函數
- Normalize
- EEG Net
- Binary Cross Entropy
- Adam
- Learning Rate = 1e-4
- 20 Epochs
EEG Net
- Kernel 0: (1,64)
- Kernel 1:(306,1)
- Pool (4)
- Kernel 2:(1,16)
- Kernel 3:(1,1)
- Pool (8)
結果
- Training: 98%
- Validation: 99%
rTMS預測
資料處理
FFT : (306,512)
Crop : (306,128)
Raw : (306,1024)
| Train | Val | |
| MDD | 4260 | 1000 |
| NC | 3613 | 1000 |
Normalize
訓練函數
- Linear(306*128,1)
- Binary Cross Entropy
- Adam
- Learning Rate = 1e-3
- 5 Epoch
結果
- Training: 100%
- Validation: 97.8%
MDD/NC 2
資料二
| Train/Val | Test | |
| MDD | 19 | 10 |
| NC | 15 | 10 |
| W12+ | W12- | |
| W2+ | 5/3 | 4/2 |
| W2- | 2/1 | 8/4 |
資料處理
SSS
Resample 256Hz
Filter 2-30Hz
Crop 4 sec Overlap 3 sec
Hanning Window
shape (306,1024)
Training: Add Noise * 2
資料二
| Train/Val | Test | |
|---|---|---|
| MDD | 10383 | 1799 |
| NC | 8361 | 1862 |
資料二
| W12+ | W12- | |
| W2+ | 2760 / 548 | 2181 / 360 |
| W2- | 1107 / 117 | 4335 / 714 |
訓練函數
- Normalize
- EEG Net
- Binary Cross Entropy
- Adam
- Learning Rate = 1e-4
- 2 Epochs
結果
Train: ~90%
Test: ~59%
接下來
- 以簡單的模型分類rTMS
- 解釋各模型的參數
- 訓練好MDD/NC
- 解釋訓練出來的模型
問題們
- 殺雞用牛刀之感
- 解釋性後的延伸