Accurate, real-time detection of fetal distress is critical during labour. While cardiotocography (CTG) is a standard for fetal monitoring, its manual interpretation is often subjective and error-prone. This study evaluates how CTG signal segment length influences the performance of machine learning models for timely fetal distress prediction. Using the CTU-UHB dataset of 552 intrapartum CTG recordings, signals were preprocessed and segmented into 90, 20, 10, and 5-minute windows. Features were extracted using eight open-source libraries and evaluated across seven classifiers. Notably, Random Forest combined with TSFEL time-domain features achieved peak accuracy (91.5%) on 10-minute segments. The last 20-minute segment also performed well (91.3%), indicating that shorter windows are sufficient for accurate prediction. Five-minute segments, however, showed reduced robustness due to data constraints. Findings suggest that 10–20 min CTG segments are best for real-time fetal monitoring, with ensemble models offering high accuracy, reliability, and potential for clinical integration in labour management.