Neuropsychological Predictors of Fatigue in Post-COVID Syndrome

Abstract

Fatigue is one of the most disabling symptoms in several neurological disorders and has an important cognitive component. However, the relationship between self-reported cognitive fatigue and objective cognitive assessment results remains elusive. Patients with post-COVID syndrome often report fatigue and cognitive issues several months after the acute infection. We aimed to develop predictive models of fatigue using neuropsychological assessments to evaluate the relationship between cognitive fatigue and objective neuropsychological assessment results. We conducted a cross-sectional study of 113 patients with post-COVID syndrome, assessing them with the Modified Fatigue Impact Scale (MFIS) and a comprehensive neuropsychological battery including standardized and computerized cognitive tests. Several machine learning algorithms were developed to predict MFIS scores (total score and cognitive fatigue score) based on neuropsychological test scores. MFIS showed moderate correlations only with the Stroop Color-Word Interference Test. Classification models obtained modest F1-scores for classification between fatigue and non-fatigued or between 3 or 4 degrees of fatigue severity. Regression models to estimate the MFIS score did not achieve adequate R² metrics. Our study did not find reliable neuropsychological predictors of cognitive fatigue in the post-COVID syndrome. This has important implications for the interpretation of fatigue and cognitive assessment. Specifically, MFIS cognitive domain could not properly capture actual cognitive fatigue. In addition, our findings suggest different pathophysiological mechanisms of fatigue and cognitive dysfunction in post-COVID syndrome.

Keywords: cognitive; fatigue; machine learning; neuropsychological; post-COVID syndrome.

Figure 1

F1-scores for each Modified Fatigue Impact Scale classification type (binary, 3-classes, and 4-classes) for each model evaluated: random forest (RF), K-nearest neighbors (KNN), support vector machine (SVM), Gaussian naive Bayes (GNB), complement naive Bayes (CNB), and logistic regression (LR).

Figure 2

F1-scores for each Modified Fatigue Impact Scale-cognitive classification type (binary, 3-classes, and 4-classes) on each model evaluated: random forest (RF), K-nearest neighbors (KNN), support vector machine (SVM), Gaussian naive Bayes (GNB), complement naive Bayes (CNB), and logistic regression (LR).

Figure 3

$R^2$ scores for each Modified Fatigue Impact Scale regression model (linear, ridge, lasso, elastic net, ANN 1, and ANN 2) for each feature reduction type (no principal component analysis [PCA], hard PCA, soft PCA). The negative section of the vertical axis is not represented to scale with the positive section to improve the visualization of values.

Figure 4

$R^2$ scores for each Modified Fatigue Impact Scale–cognitive regression model (linear, ridge, lasso, elastic net, ANN 1, and ANN 2) for each feature reduction type (no principal component analysis [PCA], hard PCA, soft PCA). The negative section of the vertical axis is not represented to scale with the positive section to improve the visualization of values. This section may be divided by subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.