ABCD challenge outcome
Leo Brueggeman
Acknowledgements - "modelers"
Tanner Koomar
Brady Hoskins
Yongchao Huang
Tien Tong
James Kent
Approach
- Kaggle inspired
Approach
- Kaggle inspired
Approach
- Kaggle inspired
Data
N = ~ 3700
>400 volumes
test
N>4000
(unlabeled)
y
age
sex
collection site
SES
total brain volume
Data
N = ~ 3700
>400 volumes
Data
Data
Approach
N = ~ 3700
>400 volumes
N = 3000
N = 700
train
validation
models
ensemble model
test
N>4000
(unlabeled)
Approach
train
R
+
Approach
train
Hyperparameter optimization in CV
- error metric: MSE
Linear models
Decision Trees
Boosting
Approach
train
Approach
validation
Approach
validation
Conclusion
Intelligence prediction from ensemble model in CV: Pearson's R = 0.12
Machine learning competitions are a great opportunity for building skills in a new scientific area (Encode)
Teamwork lessons:
github > in person meetings
starter code
Going forward:
modeling ABCD phenotypes with brain volumes (collaborators)