Bio-impedance spectroscopy-based classification of mental acuity in university students via machine-learning and deep-learning approaches

Abstract

Mental acuity detection is crucial for identifying cognitive impairments linked to body composition imbalances and ensuring overall mental and physical fitness. This study introduces a deep-learning neural network (NN) model with MobileNetV2 deep learning architecture to classify mental acuity levels of university students—excellent, good, and average—using bioelectrical impedance spectroscopy (BIS)-based body composition and bio-impedance measurements. Body composition and bio-impedance features such as basal metabolic rate (BMR), body cell mass (BCM), total body water (TBW), bioelectrical impedance (BI), and phase angle (PA) were utilized as inputs for a feature-based random forest (RF) machine-learning model, achieving an accuracy of 88.26% and an F1-score of 84.89%. However, an image-based NN model with MobileNetV2 deep learning architecture, leveraging 2D impedance spectrum images, outperformed RF model, achieving exceptional accuracy of 98.39% and an F1-score of 97.83%. Additionally, the Nyquist diagram showed that excellent mental acuity had the smallest semicircle, average mental acuity had the largest, and good mental acuity level was intermediate. Similarly, feature analysis revealed that excellent mental acuity level corresponded to high BMR, BCM, TBW, and PA with low BI, while average mental acuity level had the opposite trend, while good mental acuity levels fell in between. The reliability and performance of the NN model in detecting mental acuity using 2D impedance spectrum analysis highlights its potential for this task. These results emphasize the value of deep-learning approaches in integrating BIS data for accurate mental acuity assessment and their broader implications for monitoring cognitive health.