data_README.md

How to read the Experiment Configs and Input/Output CSV Files

Input Dataset

1. Input Variables (Ground-truth)

The ADNI patient dataset (dataset/adni/adni_fold{i}.xls) has the following columns.

Column Name	Description
RID	Patient ID
VISCODE	Baseline (bl) or month of measurement (mXX)
Years	Year of clinical measurement
DX_bl/ DX_bl_num	Diagnosis at baseline (year 0) - Type of cognitive impairment (EMCI, CN, LMCI, SMC)
CurAGE	Patient's age
PTGENDER/ PTGENDER_num	Gender (Male/Female)
PTEDUCAT	Years of education
APOEPOS	Presence of APOE ε4 gene
MMSE_norm, ADAS11_norm, ADAS13_norm	Normalized MMSE, ADAS11, ADAS13 scores
mri_FRONT_norm, mri_HIPPO_norm	$X(t)$ - Normalized Frontal/Hippocampal region size
FRONTAL_SUVR, HIPPOCAMPAL_SUVR	$D(t)$ - Instantaneous amyloid accumulation in Frontal/Hippocampal regions from florbetapir-PET scans
cogsc	$C(t)$ - Cognition score (MMSE/ADAS13 was used in these experiments)

2. Estimated parameters for differential equations

The differential equations' parameters that were estimated based on demographics of ADNI patient dataset (dataset/adni/adni_fold{i}_parameters.xls) has the following columns.

Column Name	Description
beta_estm	$\beta$ parameter for amyloid propagation
tpo_estm	Actual pathology time-period at baseline (CurAGE - 50)
alpha1_estm	$\alpha_1$ for brain degeneration
alpha2_gamma_estm	$\alpha_2 \gamma$ for computing activity Y(t)

Experiment Configurations

Each experiment's config is saved under configs\train_configs or configs\eval_config folder.

Column Name	Description
name	Experiment name
seed	Random seed used in the experiment or data generation.
gamma	The gamma parameter used in modeling the relationship between Y(t), X(t) and I(t)
gamma_type	Type of gamma parameter, which can be 'variable' or 'fixed'.
epochs	Number of training epochs or iterations in an experiment.
batch_size	Size of data batches used in training.
cog_mtl	$I_{HC}(0)$ Initial cognition score (baseline year 0) for Hippocampus (HC) region. $I_{PFC}(0) = 10.0 - I_{HC}(0)$
discount	Discount factor applied to rewards in RL.
max_time_steps	Maximum number of time steps (years in this case) n a training episode.
w_lambda	Trade-off between the mismatch (C_task - C(t)) and the energy cost M(t) in the reward function (see Eq 8 of the original paper)
action_lim	Limit or constraint applied to action values. Set to 2.0 , so $\Delta I(t)$ = [-2, 2]
cog_init	Initial value or setting for cognitive measurements. Set to full (a value of 10.0)
cog_type	Type of cognitive data, e.g., 'variable' or 'fixed'.
energy_model	Type or name of the energy model used in the experiment. inverse or inverse_squared
score	cognition score to use (MMSE, ADAS11, ADAS13).
network	MLP network hidden layer size. defauts to 2-layer MLP with hidden_size = 32, so [32,32].
algo	Name or type of the machine learning or RL algorithm.
category	Fixed to 'APOE' which is the APOE ε4 gene.

Results

1. Variables computed using estimated DE parameters and information allocation by RL model

The results for each experiment run are saved in results/{algo}/{fold}/{seed}/{experiment_name}.xlsx and contains the following RL model's predictions for each timestep (in addition to the Input Variables (Ground Truth)):

reg1: medial temporal lobe (mtl/HC) i.e. Hippocampus or hippocampal region

reg2: frontal temporal lobe (ftl/PFC) i.e. Pre-Frontal Cortex (PFC) or frontal region

Column Name	Description
reg1_info_rl	$I_{v1} (t)$ = Information processed by hippocampal region
reg2_info_rl	$I_{v1} (t)$ = Information processed by frontal region
reg1_fdg_rl	$Y_{v1} (t)$ = Hippocampal activity (fgd:fluorodeoxyglucose). Interchangeably used for energy consumption $M=\sum Y$
reg2_fdg_rl	$Y_{v2} (t)$ = Frontal activity (fgd:fluorodeoxyglucose). Interchangeably used for energy consumption $M=\sum Y$
reg1_mri_rl	$X_{v1} (t)$ = Hippocampal region size
reg2_mri_rl	$X_{v2} (t)$ = Frontal region size
reg1_D_rl	$D_{v1} (t)$ = Hippocampal instantaneous amyloid accumulation
reg2_D_rl	$D_{v2} (t)$ = Frontal instantaneous amyloid accumulation
beta_rl, alpha1_rl, alpha2_rl, gamma_rl	Parameters used by RL model for the DE-based simulator
cogsc_rl	$C(t) = \sum I_v (t) $ Cognition score computed by RL (reg1_info_rl + reg2_info_rl)
cogsc	$C(t)$ Cognition score (MMSE in our case)
cog_diff	Difference between cogsc_rl and cogsc

2. RL Reward $\Delta I(t)$ and Errors (MAE and MSE) for the Experiment

Each experiment's errors between RL predictions and ground truth values (Mean Absolute Error and Mean Square Error) are saved in results/summary_{dataset}.csv with the following data. |

Column Name	Description
train_mae, valid_mae, test_mae	Mean Absolute Error (MAE) on the train, validation and test split.
train_mse, valid_mse, test_mse	Mean Squared Error (MSE) on the train, validation and test split.
train_mae_emci, valid_mae_emci, test_mae_emci	MAE for Early Mild Cognitive Impairment (EMCI) category for the 3 splits
train_mae_cn, valid_mae_cn, test_mae_cn	MAE for Early Cognitive Normal (CN) category for the 3 splits
train_mae_lmci, valid_mae_lmci, test_mae_lmci	MAE for Late Mild Cognitive Impairment (LMCI) category for the 3 splits
train_mae_smc, valid_mae_smc, test_mae_smc	MAE for Signficant Memory Concern (SMC) category for the 3 splits
train_mse_emci, valid_mse_emci, test_mse_emci	MSE for Early Mild Cognitive Impairment (EMCI) category for the 3 splits
train_mse_cn, valid_mse_cn, test_mse_cn	MSE for Early Cognitive Normal (CN) category for the 3 splits
train_mse_lmci, valid_mse_lmci, test_mse_lmci	MSE for Late Mild Cognitive Impairment (LMCI) category for the 3 splits
train_mse_smc, valid_mse_smc, test_mse_smc	MSE for Signficant Memory Concern (SMC) category for the 3 splits
train_reward_rl, valid_reward_rl, test_reward_rl	RL-based reward for the 3 splits
train_reward, valid_reward,test_reward	Reward metric for the 3 splits