TinyStressNet: A Quantization-Ready Model for Subject-Independent Academic Stress Sensing

TinyStressNet is a compact neural classifier for three-class academic stress estimation (low, medium, high) from three physiological features (galvanic skin response, heart rate, and skin temperature) designed for low-cost educational sensing. We evaluate it on 1000 interval-level observations from five engineering students collected over three months, with labels derived from repeated 6-item short-form of the State-Trait Anxiety Inventory (STAI-6) self-reports aligned with contemporaneous physiological readings. For subject-independent evidence, we use leave-one-subject-out (LOSO) evaluation as the primary protocol, complemented by a per-user chronological split for temporal robustness. We study augmentation through the lens of approximate invariance to bounded, label-preserving nuisance transformations, comparing no augmentation, Gaussian jitter, a radial-rescaling family, and a bounded affine (BA) family against strong classical baselines (Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Support Vector Machines (SVMs), logistic and ordered-logit models). Under LOSO, the best TinyStressNet configuration reaches 0.899 accuracy and 0.898 weighted F1; within users, it reaches 0.912 accuracy. The model uses 3139 parameters and 2800 Multiply-Accumulate operations (MACs), and post-training int8 export preserves predictions with almost the same accuracy (0.992 float–int8 agreement). Strong classical baselines remain competitive on this low-dimensional representation, which we read as evidence about the geometry of the three-feature space rather than as a limitation of the model. Repeating the full evaluation under ten seeds shows that the augmentation families are not separable on this representation: subject-independent weighted F1 spans 0.894 to 0.900 across the four conditions with a seed-to-seed standard deviation of about 0.005, and no augmentation is not measurably worse than the best family. The BA family is the most consistent condition, the strongest in five of ten seeds and never the weakest, which we report as a stability observation rather than a performance gain. TinyStressNet is presented as a compact, quantization-ready reference model and a controlled testbed for invariance-aware augmentation in low-dimensional physiological sensing.