Background: Uterine contractions shape the intrapartum fetal heart-rate (FHR) response as persistent stressors that transiently interrupt uteroplacental blood flow. In healthy fetus the autonomic nervous system produces an FHR reaction to each uterine-pressure activity (UA), yielding high directed information from UA [PW1.1]to FHR with coupling delay. As fetal condition deteriorates toward metabolic acidosis, this coupling changes: UA→FHR transfer- entropy dynamic changes and the delay lengthen. Objectives: Develop a lag-aware surrogate of UA→FHR transfer entropy, embedded within a forecasting model. Explicitly model the UA→FHR lag that improves the clinical interpretability. Methods: We propose lag-attentive Transfer-Entropy-Bottleneck Variational-Autoencoder that forecasts next two minutes of FHRf, from wavelet-scattering and phase-harmonic features of past FHRp and UAp. The architecture has two causal pathways: target-only conditional prior probability p(z|FHR_p) learned from FHRp, and source-conditioned posterior probability q(z|FHR_p,UA_p) that has access to UAp where z is the latent representation. Between them there is a causal lag cross-attention module: the current target state acts as a query over a bounded memory of past UA states, producing an attention distribution over lags. FHRf is decoded as a baseline forecast from FHRp plus a UA-driven residual from the stochastic latent, forcing the bottleneck to encode UA-specific information beyond FHRp. The KL divergence between posterior and prior gives a per-timestep upper bound on transfer-entropy, measuring the extra contribution of UAp to FHRf conditioned on FHRp. Results: Figure-1 illustrates an HIE case where clinically, longer UA→FHR coupling delays are expected. Two observations emerge directly from the learned attention: 1) The UA history that contributes to the latent bottleneck is concentrated within the most recent two minutes; 2) lag-scaled TE surface is episodic with discrete bursts rather than tonic. Conclusion: The model unifies two key aspects of UA→FHR coupling, how much information flows and at what delay, into single output of Transfer-Entropy-Bottleneck