Fine motor decline serves as a critical early biomarker for neurodegenerative diseases like Parkinson’s disease, making its accurate assessment essential for early detection and intervention. While functional near-infrared spectroscopy (fNIRS) offers a portable, non-invasive neuroimaging solution, the precise cortical dynamics underlying varying levels of motor complexity remain underexplored. This study aims to investigate how fine motor task complexity modulates cortical activation and functional network topology. A secondary objective is to develop and validate a high-performance deep learning model to classify motor complexity levels from fNIRS signals. fNIRS data were recorded from healthy participants performing five fine-motor tasks of increasing complexity, and activation analyses were combined with graph-theoretical metrics to characterize neurophysiological responses. To classify the complexity of fine motor tasks from fNIRS signals, this study developed a bidirectional long short-term memory (Bi-LSTM) model. Performance evaluation used leave-one-out cross-validation, supplemented by multi-seed training to improve robustness. The model achieved an average classification accuracy of 90.67% ± 7.07% (95% CI: ± 2.68%) and an AUC of 0.9720 ± 0.0431, outperforming traditional support vector machine (by 21.3%) and Bi-LSTM (by 10.97%). These results demonstrate the model’s strong generalization across subjects and its ability to capture temporal-spatial patterns of cortical activation associated with increasing task complexity, providing a promising foundation for fine motor decoding and adaptive neurorehabilitation.