To address opaque decision-making and performance bottlenecks caused by limited samples and physiological heterogeneity in deep learning-based sleep staging, this study proposes a Logic-Driven CNN19 (LD-CNN19) neuro-symbolic framework. LD-CNN19 fuses automated feature representation with medical expertise, extracting multi-scale deep and circadian rhythm features from single-channel electroencephalogram (EEG) signals to generate initial predictions. A neuro-symbolic bridge, implemented via Pyswip, utilizes a dual-knowledge base based on American Academy of Sleep Medicine (AASM) standards to reconstruct sensory biases through physiological semantics and the Principle of Minimal Inconsistency. Furthermore, an Adaptive Knowledge Expansion Mechanism employs an Enhancement Evaluation Strategy to evolve the knowledge base by mining pseudo-knowledge from data. Experiments on three public datasets demonstrate that LD-CNN19 improves macro F1-scores by up to 4.22% compared to the backbone and significantly enhances decision stability, with the standard deviation of F1-scores in the challenging N1 stage narrowing from 11.24% to 5.20%. Further validation underscores the framework's model-agnostic versatility, as its integration with diverse backbone architectures consistently yields performance and generalization robustness gains across varying data scales. Logical Attribution Transfer Analysis quantifies the framework's impact, correcting approximately 14.4% of perception-layer misclassifications with a 8:1 Rescue-to-Risk Ratio. This work enhances small-sample learning robustness and offers a traceable reasoning path for clinical decision support.
