Contemporary research on stress highlights its significant impact on both physical and mental health, prompting the pursuit of objective methods for measuring this phenomenon. In response to this challenge, the present article proposes an innovative hybrid stress classification method that combines nonlinear Gompertz weighting with adaptive fuzzy logic within an ensemble learning framework of lightweight convolutional neural networks (lightweight CNNs). The key scientific contributions include: an innovative integration of the Gompertz function with a fuzzy system for assessing prediction confidence, comprehensive validation of the approach on a modified version of the Comprehensive Facial Thermal Dataset (CFTD), where the original emotion classes were mapped to three stress levels: no stress, low stress, and high stress. The study used six lightweight CNN architectures – MobileNetV3-Lite, TinyNet-E, FBNetV3, CondenseNetV2, Nanonet, and ShuffleNetV2 – whose predictions were aggregated through a three-stage process: initial nonlinear weighting by the Gompertz function, fuzzy scaling of weights based on classification confidence, and final fusion using fuzzy rules. Experiments were conducted in two variants – using a single thermal palette and mixed palettes – with 5-fold cross-validation. Results demonstrated that using a single thermal palette achieved significantly higher average accuracy (MobileNetV3-Lite: 80.1%) compared to the mixed palettes variant (78.2%). The hybrid approach, combining the Gompertz function and fuzzy logic, significantly improved classification performance by reducing errors by 19–30% depending on the stress class, particularly for the “high-stress” class and scenarios with marked prediction uncertainty. The best performance was observed with the MobileNetV3-Lite architecture, which, thanks to advanced attention mechanisms (SE blocks), effectively leverages thermal representation. Furthermore, fuzzy logic helped mitigate the negative influence of weaker models, resulting in enhanced stability and reliability of the stress classification system.