<p>Sport-related attention is hard to read from any single signal, and models built for simpler settings tend to falter in real competition. We developed a multimodal framework that profiles neural regulation and recognizes task-defined attention states in relation to athletic performance. Seventy-two athletes took part, with EEG, eye tracking, heart-rate variability, and performance recorded in parallel across five standardized task phases. After time synchronization, artifact suppression, window selection, and multimodal feature extraction, the model was trained and tested across attention states, performance levels, modality contributions, and generalization scenarios. The five recognized states correspond to five standardized task phases rather than independently validated latent states. Under high load, fronto-parietal connectivity (weighted phase-lag index) rose to 0.71 ± 0.05, while relative β and γ power reached 0.314 ± 0.029 and 0.126 ± 0.018. The Neural Regulation Index (NRI), a transparent composite of neural, visual, and autonomic regulation features, was 0.742 ± 0.109 and, together with on-target gaze time (68.4 ± 12.0%) and motor onset latency (312.7 ± 48.0 ms), separated the performance groups. Under subject-independent (leave-subjects-out) validation the model reached 90.8% accuracy; the 93.6% obtained with record-level cross-validation is reported only as a within-subject upper bound. A time-only baseline (52.3%) and a temporally shuffled control (87.9% vs. 90.8%) indicated that recognition was not driven solely by fixed task timing, and behavioural outcomes were predicted from EEG, eye-tracking, and HRV alone. Overall, multimodal neural-regulation profiling offers a practical way to characterize the attention features underlying motor performance, with a basis for performance monitoring, load adjustment, and individualized feedback.</p>