I only used tabular features, didn't use CT scans or other meta data. I don't have enough time to extract, and my computer is not good at all, so I just do my best on tabular data.
-
target_week
-
base_Week
-
base_FVC
-
base_Percent
-
Age
-
weeks_passed: target_week - base_week
-
Height: (base_FVC / 933.33 + 0.026 * Age +2.89) / 0.0443
-
FEV: $\begin{cases} 0.84 * FVC - 0.23 \text{, if Sex = Male}\ 0.84 * FVC - 0.36 \text{, if Sex = Female} \end{cases}$
-
percent_reciprocal: 1 / base_Percent
-
FEV_ratio: FEV / base_FVC
-
percent_ratio: base_FVC / base_Percent
Because there are few categories, so I just use one hot encoding.
- Female
- Male
- Currently smokes
- Ex-smoker
- Never smoked
Using quantile LGBM Regressor with alpha = [0.75, 0.25] to subtract.
