Early Imaging Based Predictive Modeling of Cognitive Performance Following Therapy for Childhood ALL

Abstract

In the United States, Acute Lymphoblastic Leukemia (ALL), the most common child and adolescent malignancy, accounts for roughly 25% of childhood cancers diagnosed annually with a 5-year survival rate as high as 94% [1]. This improved survival rate comes with an increased risk for delayed neurocognitive effects in attention, working memory, and processing speed [2]. Predictive modeling and characterization of neurocognitive effects are critical to inform the family and also to identify patients for interventions targeting. Current state-of-the-art methods mainly use hypothesis-driven statistical testing methods to characterize and model such cognitive events. While these techniques have proven to be useful in understanding cognitive abilities, they are inadequate in explaining causal relationships, as well as individuality and variations. In this study, we developed multivariate data-driven models to measure the late neurocognitive effects of ALL patients using behavioral phenotypes, Diffusion Tensor Magnetic Resonance Imaging (DTI) based tractography data, morphometry statistics, tractography measures, behavioral, and demographic variables. Alongside conventional machine learning and graph mining, we adopted "Stability Selection" to select the most relevant features and choose models that are consistent over a range of parameters. The proposed approach demonstrated substantially improved accuracy (13% - 26%) over existing models and also yielded relevant features that were verified by domain experts.

Keywords: eature Selection; graph mining; neurocognitive late effect; predictive modeling.F; predictive modeling.eature Selection; stability selection and control.

Fig. 8:

Stability path of features matrix (for DSB) with a range of regularization parameter (α = 0.01 ~ 100) as a function of (α/α_max)^1/3. The power 1/3 scales the path and enables to visualize the progression along the path.

Fig. 9:

Learning curve for DSB class using random shuffle-split approach. This method reduces the chance of over or under fitting. Green and red lines indicate that the amount of training example increase cross validation accuracy and decrease the training accuracy.

Fig. 1:

Pie plot shows diversity in the dataset. The dataset has demographic measures of different sex, ethnicity, race and age group. The sample size of the average and below-average group of DSB, PS, and BRIEF-Working Memory measures are relatively unbalanced. Here, ethnicity Group1 represents Non-Hispanic and Group 2 represents not otherwise specified Spanish, Hispanic, Latino group respectively.

Fig. 2:

The t-SNE embedded higher dimensional features are represented by 2-dimensional scatter and kernel density estimation (KDE) plot. The green lines with dots and red lines with ‘+’ sign represents average and below-average group data, respectively.

Fig. 3:

Expected number of falsely selected variable E(V) VS q_Λ graph and stability path for DSB class. Left side of the plot (a) show the variation of E(V) and pi_thr on model accuracy. Red solid lines of plot (b) show the relevant features (63) for best E(V) and pi_thr and black dotted lines represents stability path for irrelevant features (956) over a range of regularization parameter.

Fig. 4:

Effect on section threshold over model performance for BRIEF-Working Memory prediction. Three lines of x-label represent the range of each bin of features score (range: 0 to1), number and percent of feature fall in each bin.

Fig. 5:

Vin diagram of EC and RL selected features for (a): DSB, (b): BRIEF-Working Memory, (c): PS class. Cyan, Brown and blue colored circle represent the number of stable features selected by EC, RL and common features among methods. Prediction accuracy and number of selected features are relatively better for RL method. Here ACC represents accuracy.

Fig. 6:

Schematic diagram of the processing pipeline. Feature matrix is randomly shuffled and split into 80 and 20% as training and testing data. Feature selection methods (EC and RL) are applied on training data to find the stable features. Those selected features were used to tune and estimator learning using shuffle-split grid search approach, and finally, models are evaluated on test data.

Fig. 7:

Circular brain connectivity graph for a): BRIEF-Working Memory, (b): DSB, and (c): PS class using RL method. Left and right side of the circle represents left and right hemisphere. The inner squires, outer squires and green connected lined indicate selected ROIs, cortical thickness of ROIs and connectivity among ROIs, respectively. Shape and the size of the outer square varies with rank (importance) in predicting impairment.