Traditional EEG-based neurological disease classification typically relies on binary Independent Component Analysis (ICA) artifact removal, discarding all non-brain physiological signals under the assumption that they lack diagnostic value. This study proposes Dynamic Weight Learner (DWL) DWL-ICA, an end-to-end trainable framework that learns adaptive ICA component weighting for stroke and traumatic brain injury (TBI) classification. The DWL generates sample-adaptive weights for seven ICA component categories, enabling selective preservation of diagnostically informative physiological signals while suppressing true noise. Evaluated on the TUH-EEG three-class benchmark dataset—the only publicly available multi-class neurological disorder EEG dataset comprising 9,276 labeled segments— DWL-ICA achieved 0.95 macro-average ROC–AUC and 83% accuracy, representing an 11-point improvement over conventional static ICA preprocessing (0.84 baseline). Controlled ablation experiments decomposed this gain: adaptive weighting contributes 4 points (improving from 0.84 to 0.88), TMN features contribute 5 points (from 0.84 to 0.89), and joint end-to-end optimization yields an additional 2-point synergistic gain (from 0.93 to 0.95), demonstrating that the learned preprocessing strategy is specifically tailored to exploit TMN’s spatial-frequency representation. The learned weights assigned moderate importance to Heart (0.60) and Eye (0.47) components without explicit regularization, demonstrating that physiological signals traditionally treated as artifacts contain meaningful diagnostic information. The framework’s computational efficiency and flexible deployment options make it suitable for real-time clinical applications, including portable EEG devices in resource-limited settings.