Background: Accurate assessment of fetal well-being during labor using cardiotography to monitor fetal heart rate (FHR) and uterine pressure (UP) is critical for detecting fetuses at risk of developing hypoxic-ischemic encephalopathy (HIE). Here, we propose a novel deep learning architecture that integrates Transfer Entropy Bottleneck (TEB) framework within SeqVAE to model the directed physiological coupling of UP and FHR. SeqVAE-TEB is designed to improve predictive performance and learn informative latent representations, improving early assessment of risk for HIE and acidosis (non-HIE) outcomes. Methods: We utilize a dataset of 250,000 births, including healthy infants with no cord/infant blood gas data (5 minute Apgar score of at least 7) two groups with and without neurological symptoms, HIE and acidosis respectively, identified by low pH (<7.0) or high base deficit (≥10 mmol/L). Building on the SeqVAE model, we condition the FHR latent space on UP using TEB. SeqVAE-TEB model focuses on the transfer entropy-the directed information from UP to FHR-while discarding irrelevant input variance. To emphasize UP-FHR coupling, we used a prediction task rather than reconstruction. Accordingly, we predict the next two minutes of FHR. Pretraining on the final 12 hours of labor used data from 20,968 healthy vaginal deliveries. We appended an LSTM-based classification head to the pretrained model (with a frozen SeqVAE-TEB model) in a 10-fold cross-validation setup on a distinct dataset of 2,988 healthy and 3,200 acidosis deliveries, while all 391 HIE deliveries were reserved exclusively for testing. Results: specificity decreases from 0.80 (six hours before birth) to 0.61 (one hour before birth), while sensitivity increases from 0.3 to 0.56. Distinguishing HIE from healthy cases shows stable performance with lower variability (attributed to the model with fixed HIE cases for all folds). The results show that as delivery approaches, the model predicts a higher risk for both HIE and acidosis.