Data Challenge : Leukemia Risk Prediction

Selected Challenge

Overall Survival Prediction for patients diagnosed with Myeloid Leukemia by QRT in collaboration with the Gustave Roussy institute. This challenge was carried out as part of the course "Génération de Données en IA par Transport et Débruitage" by Prof. Stéphane Mallat from the Collège de France.

Team: Nour RIZK, Vivien BRANDT

Code Setup

root dir
  │
  ├── data
  │     ├── target_train.csv
  │     ├── X_train
  │     │     ├── clinical_train.csv
  │     │     └── molecular_train.csv
  │     └── X_test
  │           ├── clinical_test.csv
  │           └── molecular_test.csv
  │
  ├── code
  │     ├── notebook_rizk_brandt.ipynb
  │     └── output
  │           ├── ... submission.csv
  │           └── spearman_corr_btw_submissions.ipynb
  │
  └── ...

Strategy :

1. Feature Engineering

• Ratios (interactions) of the different cell counts (features collected from the clinical data csv) in the blood stream
• Cytogenetic data :
  • 1st approach : construct a one-hot encoding of the risk level of the data (High/Low) --> categorical data
  • 2nd approach : construct a high risk abnormalities dictionary and one-hot encode the presence of those abnormalitites in a patient
  • ** next up** : implement it as a numerical data approach based on thresholds with percentage of risk
• Molecular data :
  • One-hot encode the Gene Mutation data based on VAF information
  • One-hot encode the Mutation Effect data based on the Effect column
  • next up : work on feature selection to reduce correlation and improve the model ipcw c-index
  • next up : train a model (such as a word2vec or VAE) on the molecular data before feeding it to the models$
• Merging the processed Clinical and Molecular data
• Filtering highly correlated data
• Imputing the data with the median value
• Standardizing the data
• Selecting a strategy : either train/test split, or cross validation (both are implemented)

2. Trained models :

• COX Proportional Hazard model with Ridge regularization $0.72$
• Random Survival Forest model - without Znormalizing the data : $0.745$ - with Z-normalizing the data : $0.749$
• DeepSurv with Regularization and Dropout

Implemented a grid search for all models.

3. Defining 2 pairs of models' prediction score for applying ensemble methods :

• (COX-PH, RSF)
• (DeepSurv, RSF)

Z-normalizing the data before stacking it and feeding it to the Meta Learners

4. Ensemble Methods :

• Averaging prediction scores of a given pair
• Linear Regression as a Meta Learner
• XGBoost as a Meta Learner

Results

The table below displays the results of the various implemented methods, given by the Public Leaderboard of the submission platform.

We can see that the optimal methods (equivalently) are RSF, or the Meta Learner trained on COX-PH, RSF pair.