DHP(Digital Health Hackathon 2021, Korea - AI Track)

• Awarded Excellence Award(1st prize) by CCEI(서울창조경제혁신센터), Naver Care

Problem

• Reinforcement learning based gene selection that is effective for cancer treatment
• Feature selection problem

How to deal with

Pipeline

Filter Method (Cox regression)

Filter the feature subset that is expected to have a correlation between mutant gene information and survival period with p-value after the cox regression

Wrapper Method (QBSO-FS + Cox regression)

With feature subset, QBSO-FS,one of wrapper methods, is used to suggest top 10 candidate mutant gene.
Compared to the original QBSO-FS, cox regression parameters that influence the positive treatment effect compose the reward function.

• Markov Decision Process (MDP)

State: Feature subset that uses for cox regression in the subset
Action: Flipping whether use the feature or not
Reward:

Analysis of Examples

• The effect of G88 gene mutation

G88 is higher effect in cox_treat, less effect in cox_notreat. --> If G88 is mutant gene, cancer treatment has positive effect.

How to implement

Prerequisite

    pip install lifelines scikit-learn pandas xlsxwriter matplotlib

Implementation code

    python main.py

Reference

[Cox Regression]https://www.jstor.org/stable/pdf/2532940.pdf
[QBSO-FS]https://link.springer.com/chapter/10.1007/978-3-030-20518-8_65

Code Reference

[QBSO-FS]https://github.com/amineremache/qbso-fs

Contributor

Minkoo Kang (Leader)
Minsoo Kang
Dongjin Kim
[KIST-KDST]https://kdst.re.kr