Similarity Feature Selection for Survival Analysis

• Research codes about new feature selection method with similarity score for Survival Analysis in UNLV short-term research program
• Yonghwan Lee, Hyojin Yeon, Jihyeon Kwon, Soyeon Bae
• Advisor : Mingon Kang

Task Definition

• Project 4. Deep Learning-Based Survival Analysis Using Genomic Data
• Genomic Data : LGG, GBM Data from TCGA
• Survival_CostFunc_CIndex.py : Referenced to Cox-PASNet
• Datasets : HDLSS -> decreasing dimensions of data(Columns) -> Need Feature Selection
• We proposed new feature selection method with Similarity Score
• We compared ordinary feature selection methods
  • Cox-EN : Selecting by Cox-EN model
  • High Variance : Selecting high variance columns

Model

• model : Cox-nnet(PyTorch Lightning)
  • layer : 2(input_dim -> 128 -> 64 -> 1)
  • dropout : y
  • activation function : y
  • output layer : bias=False, Non-activation
• metric : C-index(Survival_CostFunc_CIndex.py)
• loss function : Negative log-partial likelihood(Survival_CostFunc_CIndex.py)

Settings

• Environments

pip install -r requirements.txt

Preprocess

• input_path : path for input data
• save_path : path for save data
• method : both(Similarity and Cox-EN), sim(Similarity), cox(Cox-EN)
• cox_ratio : Elasticnet's L1 ratio
• sim_threshold : Select Similarity Columns(Genes) Ratio, 0.6 -> if 60% values of column is larger than 0, we select it!

python preprocess.py \
-input_path "./data/result.csv" \
-save_path "./data/" \
-method "both" \
-cox_ratio 0.005 \
-sim_threshold 0.6

Run

• data_path : path for load data
• gpus : Select GPU number, Please input string value like "0" or "1"
• method : similarity(Feature Selected by Similarity Score) or cox_en(Feature Selected by Cox-EN model)
• act_func : Select Activation Function
  • tanh : Hyperbolic tangent
  • sigm : Sigmoid Function
  • relu : ReLU
  • gelu : GELU
  • leak : Leaky ReLU

python run.py \
-data_path "./data/" \
-gpus "0" \
-method "similarity" \
-epoch 2000 \
-batch_size 256 \
-dropout_rate 0.8 \
-act_func "relu"

Result

• Our method(Similarity Score) is the highest C_index Score.
• Learning Curve