This repository contains the code source of experiments performed in the paper "Interpretable Machine Learning for Precision Aging" submitted to the Cognitive Computation journal and additional results. In this paper, we used the Explainable Boosting Machine (EBM) model to assess the relationship between background factors, socio-economic status, daily lifestyle activities, and the cognitive health of older adults.

MODULES

The modules are described as follows:

• experiments.py: contains the main code to run all experiments.
• src/explain.py: contains the Explain class used to provide the following explanations:
  • Feature importance (Mean absolute logit from the EBM model).
  • Weights associated to each variable bin. These weights are used to plot a linear relationship between the dependent and independent variables.
  • Statistical significance scores: z-scores and p-values.
• src/plotting.py: contains utility functions for plotting.
• src/utils.py: contains addtional utility functions.
• src/type_experiments.py: contains a class used to specify the type of experiments to run.
• src/cross_validation.py: contains a class used to cross-validate selected machine learning models. This module is used to compare the performance of the EBM, Logistic Regression, XGBoost, and Multi-Layer Perceptrons.
• src/config.py: contains experimental settings (data set name, plotting parameters, etc...).

EXPERIMENTS

All experiements in our paper are reproducible. To that end, please follow the instructions below.

1. Dowload the 2012 and 2016 waves of Health and Retirement Study dataset here.
2. Concatenate the two waves so that each individual in the dataset has two observations (2012 and 2016).
3. Drop all missing values and filter the dataset by selecting only individuals with age >= 60.
4. Select the variables used in the experiments:
   • Immadiate and delayed cognitive scores used to create a composite score.
   • Continuous variables: Age and Education.
   • Binary variables: see Table 1 in the paper.
5. Process the selected variables:
   • Cognitive scores: create a composite score (Immadite + Delayed scores) and stratify the obtained scores into 3, 5, and 9 cognitive categories.
   • Binarize the daily lifestyle activities with the process described in the Appendix of the paper.
6. Install package dependencies:

pip install -r requirements.txt

7. Set the experiment configurations (dataset name, saving paths in your computer) in the config.py file.
8. Run the experiments with the following command:

python experiments.py

If you would like to integrate performance comparison into each experiment, specify the --perf_comp flag as follows:

python experiments.py --perf_comp

SUPPLEMENTARY RESULTS

Due to space considerations in the main paper and supplementary material, we reported results corresponding to the 12 most important variables in the EBM model. We provide the following supplementary results:

• patterns of all independent variables.
• P-values associated with all bins for continuous variables and category 1 for all independent variables

Dependent variable with 3 cognitive categories (Cogn3)

patterns

patterns associated with all independent variables when age and education are considered in the experiment.

patterns associated with all independent variables when age and education are not considered in the experiment.

P-values

P-values with Age and Education : P-values associated with independent variables when age and education are considered in the experiment.

P-values without Age and Education: P-values associated to independent variables when age and education are not considered in the experiment.

Dependent variable with 5 cognitive categories (Cogn5)

patterns

patterns associated with all independent variables when age and education are considered in the experiment.

patterns associated with all independent variables when age and education are not considered in the experiment.

P-values

P-values with Age and Education: P-values associated with independent variables when age and education are considered in the experiment.

P-values without Age and Education: P-values associated to independent variables when age and education are not considered in the experiment.

Dependent variable with 9 cognitive categories (Cogn9)

patterns

patterns associated with all independent variables when age and education are considered in the experiment.

patterns associated with all independent variables when age and education are not considered in the experiment.

P-values

P-values with Age and Education: P-values associated to independent variables when age and education are considered in the experiment.

P-values without Age and Education: P-values associated to independent variables when age and education are not considered in the experiment.