Application of Machine Learning Models for Tracking Participant Skills in Cognitive Training

Abstract

A key need in cognitive training interventions is to personalize task difficulty to each user and to adapt this difficulty to continually apply appropriate challenges as users improve their skill to perform the tasks. Here we examine how Bayesian filtering approaches, such as hidden Markov models and Kalman filters, and deep-learning approaches, such as the long short-term memory (LSTM) model, may be useful methods to estimate user skill level and predict appropriate task challenges. A possible advantage of these models over commonly used adaptive methods, such as staircases or blockwise adjustment methods that are based only upon recent performance, is that Bayesian filtering and deep learning approaches can model the trajectory of user performance across multiple sessions and incorporate data from multiple users to optimize local estimates. As a proof of concept, we fit data from two large cohorts of undergraduate students performing WM training using an N-back task. Results show that all three models predict appropriate challenges for different users. However, the hidden Markov models were most accurate in predicting participants' performances as a function of provided challenges, and thus, they placed participants at appropriate future challenges. These data provide good support for the potential of machine learning approaches as appropriate methods to personalize task performance to users in tasks that require adaptively determined challenges.

Keywords: Bayesian filtering; cognitive memory training; deep-learning; hidden Markov model; n-back training; video games.

Copyright © 2020 Sandeep, Shelton, Pahor, Jaeggi and Seitz.

Figures

Figure 1

(Top) Example for a two-back level in the non-gamified Tapback condition. (Middle) Example for a two-back level in the gamified Recollect condition. (Bottom) Example for a two-back level in the gamified Recall condition (Mohammed et al., 2017).

Figure 2

A visual of how the Recollect game translates to accBlock. It shows the 1-back task in which the player needs to collect items that match those seen 1 trial ago. The top part shows the first item in a 1-back task, a pink diamond (trial 1). The bottom part shows examples of responses in the next trial: the player either collects the target (also a pink diamond; true positive), or misses a target (false negative), or collects a non-target (yellow gem; false positive).

Figure 3

Graphical representation of the system. A block's value corresponds to a single node in the model. This is shown in the top portion of the figure, where yt, the accuracy achieved at time t, and zt, the presented n-level of the task at time t, are for a single block and that translates into a node in the model. The bottom is the graphical representation of the hidden Markov model. The arrows represent direct conditional dependencies among the modeled variables. The state transition model (from xt−1 to xt) is described in section 3.2.3.

Figure 4

Graphical representation of a history-driven HMM. Generation of yt, zt from a block remains the same as in Figure 3.

Figure 5

Test MSE across all participants for (top) model-1 and (bottom) model-2. The participants highlighted in red correspond to the samples picked below for analysis.

Figure 6

Comparing models for participant SP438. (A) Trajectory of estimated WM skill from model-1. (B) Trajectory of estimated WM skill from model-2. (C) Prediction vs. observed block accuracy at each block from model-1. (D) Prediction vs. observed block accuracy at each block from model-2.

Figure 7

Comparing models for participant SP259. (A) Trajectory of estimated WM skill from model-1. (B) Trajectory of estimated WM skill from model-2. (C) Prediction vs. observed block accuracy at each block from model-1. (D) Prediction vs. observed block accuracy at each block from model-2.

Figure 8

Comparing models for participant SP451. (A) Trajectory of estimated WM skill from model-1. (B) Trajectory of estimated WM skill from model-2. (C) Prediction vs. observed block accuracy at each block from model-1. (D) Prediction vs. observed block accuracy at each block from model-2.

Figure 9

Test MSE across all participants for (top) model-1 and (bottom) model-2. The participants highlighted in red correspond to the samples picked below for analysis.

Figure 10

Comparing models for participant RLB102. (A) Trajectory of estimated WM skill from model-1. (B) Trajectory of estimated WM skill from model-2. (C) Prediction vs. observed block accuracy at each block from model-1. (D) Prediction vs. observed block accuracy at each block using model-2.

Figure 11

Comparing models for participant RLB162. (A) Trajectory of estimated WM skill from model-1. (B) Trajectory of estimated WM skill from model-2. (C) Prediction vs. observed block accuracy at each block from model-1. (D) Prediction vs. observed block accuracy at each block from model-2.

Figure 12

Samples from experiment 1. (A,D): SP438, (B,E): SP451, (C,F): SP259. (A–C) Estimated n-level trajectory using UKF model. (D–F) Prediction vs. observed block accuracy at each block using UKF model.

Figure 13

Samples from experiment 2. (A,C): RLB102, (B,D): RLB162. (A,B) Estimated n-level trajectory using UKF model. (C,D) Prediction vs. observed block accuracy at each block using UKF model.

Figure 14

Samples from experiment 1. (A,D): SP438, (B,E): SP451, (C,F): SP259. (A–C): Estimated n-level trajectory using LSTM model. (D–F) Prediction vs. observed block accuracy at each block using LSTM model.

Figure 15

Samples from experiment 2. (A,C): RLB102, (B,D): RLB162. (A,B): Estimated n-level trajectory using LSTM model. (C,D) Prediction vs. observed block accuracy at each block using LSTM model.