Biomedical signals such as ECG, EEG, and respiratory waveforms are highly nonlinear and temporally dependent, making accurate forecasting difficult. This work introduces a Hybrid Neural Network(HNN) integrating multi-level signal decomposition, deep temporal modeling, and neuromorphic dynamics for biomedical time-series prediction. The framework combines Empirical Mode Decomposition(EMD) for adaptive denoising, a CNN-LSTM attention module for temporal feature learning, and a fine-tuned Liquid Neural Network(LNN) for time dependent forecasting. Bayesian Optimization and self-supervised cross signal pretraining further enhance model performance. Experiments on MIT-BIH Arrhythmia and MIMIC-IV waveform datasets demonstrate that HNN outperforms CNN, LSTM, and Transformer baselines in RMSE and R², while retaining interpretability via attention maps. Findings highlight HNN as a generalized and efficient solution for biomedical signal forecasting and real-time clinical decision support.