jmmlrc

Joint Multi-Modal Longitudinal Regression and Classification (JMMLRC) Estimator

This repository contains the code for performing a classification and a regression task at the same time on longitudinal data. This work has been applied to the ADNI dataset and has shown state-of-the-art performance in predicting patients with and without Alzheimer's Disease.

Getting Started using Conda

In order to set up your environment you must first install conda. Once this is done follow these commands to setup your environment.

cd jmmlrc
conda create --name jmmlrc python=3.7 --file requirements.txt 
conda activate jmmlrc

Once you have activated the jmmlrc conda environment you can run all associated tests with the following:

pytest test_jmmlrc_estimator.py

The test cases in test_jmmlrc_estimator.py show how the JMMLRC estimator can be used in practice (e.g. fit, predict, hyperparameter tuning, etc.)

Experiment Hyperparameters

In our TMI submission titled "Joint Multi-Modal Longitudinal Regression and Classification for Alzheimer’s Disease Prediction" we compare our method to the an array of machine learning algorithms with the following settings:

Regression

linear = LinearRegression(normalize=True, fit_intercept=False)
ridge = Ridge(alpha=1000)
lasso = Lasso(alpha=0.001)
mlp = MLPRegressor(alpha=1, activation='logistic', hidden_layer_sizes=(10,))
elm = ELM(neurons=100, func="rbf_l2")

jmmlrc_l21 = JMMLRC(gamma1 = 100, gamma2 = 0, gamma3 = 0)
jmmlrc_group = JMMLRC(gamma1 = 0, gamma2 = 100, gamma3 = 0)
jmmlrc_trace = JMMLRC(gamma1 = 0, gamma2 = 0, gamma3 = 100)

jmmlrc = JMMLRC(gamma1 = 1e-5, gamma2 = 0.01, gamma3 = 100)

Classification

log = LogisticRegression(C=0.1, penalty='l1')
tree = DecisionTreeClassifier(criterion='entropy')
svc = SVC(C=10, kernel='sigmoid')
knn = KNeighborsClassifier(weights='distance', algorithm='kd_tree', n_neighbors=20, p=1)
mlp = MLPClassifier(alpha=1, hidden_layer_sizes=(10,), activation='logistic')
sgd = SGDClassifier(alpha=0.01, loss='log', penalty='elasticnet', l1_ratio=0.5)
linearsvc = LinearSVC(C=0.001, loss='hinge')
elm = ELM(neurons=1000, func="tanh", classification="c")

jmmlrc_l21 = JMMLRC(gamma1 = 100, gamma2 = 0, gamma3 = 0)
jmmlrc_group = JMMLRC(gamma1 = 0, gamma2 = 1000, gamma3 = 0)
jmmlrc_trace = JMMLRC(gamma1 = 0, gamma2 = 0, gamma3 = 100)

jmmlrc = JMMLRC(gamma1 = 1e-5, gamma2 = .01, gamma3 = 100)

Citation

If you find this code useful in your research, please consider citing:

@article{brand2019joint,
  title={Joint Multi-Modal Longitudinal Regression and Classification for Alzheimer’s Disease Prediction},
  author={Brand, Lodewijk and Nichols, Kai and Wang, Hua and Shen, Li and Huang, Heng},
  journal={IEEE Transactions on Medical Imaging},
  volume={39},
  number={6},
  pages={1845--1855},
  year={2019},
  publisher={IEEE}
}

LifangHe / jmmlrc