Performance analysis of single and multi-stage metaheuristic optimization on DFFNN for electrocardiogram-based emotion classification

Abstract

Emotion classification based on electrocardiogram (ECG) signals has attracted increasing attention in affective computing and biomedical signal processing. However, training deep feedforward neural networks (DFFNN) using conventional gradient-based learning often suffers from local minima and slow convergence, particularly when dealing with nonlinear and limited datasets. This study presents a comprehensive performance analysis of single-stage and multi-stage metaheuristic optimization strategies applied to DFFNN for ECG-based emotion lassification in elderly participants. Five models were evaluated: Pure DFFNN, DFFNN optimized using genetic algorithm (GA), particle swarm optimization (PSO), grey wolf optimizer (GWO), and a hybrid multi-stage DFFNN+GA+GWO model. Experimental results from six independent trials demonstrate a substantial reduction in mean squared error (MSE) when metaheuristic optimization is applied. Pure DFFNN produced final MSE values in the range of 0.07462–0.08977, whereas DFFNN+GWO reduced MSE to 0.01894–0.02411. The proposed multi-stage DFFNN+GA+GWO achieved the lowest MSE of 0.014286 in the best run and an average MSE of approximately 0.0212 across trials. Training accuracy improved from 57.14%–66.67% (Pure DFFNN) to 80.95%–85.71% using metaheuristic pproaches. Although testing accuracy remained relatively stable at 33.33%–50.00% due to dataset size constraints, convergence behavior analysis shows that multi-stage optimization enhances stability and reduces oscillatory updates. These findings confirm that multi-stage metaheuristic optimization significantly improves training stability and error minimization in DFFNN models, offering a promising strategy for robust ECG-based emotion classification under small-sample conditions.