A random forest machine-learning model using ten physiological features discriminated relative cognitive performance in healthy adults with 70.83% accuracy and an AUC of 71.2%.
Cross-Sectional (n=240)
Do machine-learning models using physiological parameters predict cognitive performance in healthy adults?
Machine learning models using cardiovascular and autonomic physiological parameters show moderate ability to predict cognitive performance in healthy adults.
Effect estimate: AUC 71.2%
Introduction: Current cognitive tasks are not suitable for frequent monitoring of cognitive function in healthy adults. Increasing evidence suggests that cardiorespiratory fitness, cardiovascular function, and autonomic regulation are associated with cognitive performance; however, these multidimensional relationships are challenging to interpret using traditional statistical methods. The present study examined the feasibility of using fitness-related physiological and cardiac autonomic indicators, together with interpretable machine-learning approaches, to assess relative cognitive performance in healthy adults. Methods: In a cross-sectional sample of 240 adults, 39 physiological variables were recorded as input features. Trail making test completion time was dichotomized at the median as the outcome variable. Four feature-selection strategies, correlation, mutual information, genetic algorithms, and recursive feature elimination, were combined with grid-tuned classifiers under stratified 5-fold cross-validation and probability calibration. Results: A random forest model with ten RFE-selected features achieved 70.83 % accuracy, 71.38 % F1, and AUC = 71.2%, outperforming an untuned logistic-regression baseline model. SHAP-based interpretation indicated that older age, higher systemic vascular resistance, and higher resting heart rate shifted predictions toward the longer TMT-time group, whereas greater stroke volume, cardiac output, high-frequency power, and respiratory sinus arrhythmia shifted predictions toward the shorter TMT-time group. Conclusions: Physiological parameters related to cardiovascular and autonomic function showed moderate ability to discriminate relative TMT-based performance groups in healthy adults, supporting the feasibility of physiology-based cognitive assessment. Several key features identified by the model are modifiable through exercise and lifestyle interventions, suggesting potential translational value. With further validation and refinement, including evaluation of wearable-accessible physiological features, such models may support lower-burden monitoring and future personalized cognitive-health applications.
Yu et al. (Wed,) conducted a cross-sectional in Healthy adults (n=240). Machine-learning prediction using physiological parameters vs. Untuned logistic-regression baseline model was evaluated on Trail making test completion time dichotomized at the median (AUC 71.2%). A random forest machine-learning model using ten physiological features discriminated relative cognitive performance in healthy adults with 70.83% accuracy and an AUC of 71.2%.