QT-MIL-FHR: A Context-Aware Multiple Instance Learning Framework with Quality- and Time-Conditioned Attention for Fetal Acidemia Prediction from Intrapartum CTG

Predicting fetal acidemia from intrapartum fetal heart rate (FHR) signals remains a major challenge in electronic fetal monitoring. Visual interpretation of cardiotocography (CTG) is highly dependent on clinical expertise and is subject to considerable inter-observer variability. During CTG assessment, obstetricians do not assign equal importance to all signal segments; instead, their attention is influenced by contextual factors such as signal quality and proximity to delivery. However, most existing deep learning approaches rely primarily on signal content and lack an explicit mechanism to incorporate such clinical context. This study proposes QT-MIL-FHR (Quality- and Time-Conditioned Multiple Instance Learning for Fetal Heart Rate), a context-aware multiple instance learning (MIL) framework that explicitly embeds clinical contextual information into the attention mechanism. The proposed architecture consists of two complementary branches: a temporal branch based on a CNN-BiGRU for direct FHR analysis and a physiological branch utilizing 11 heart rate variability (HRV) features, including nonlinear indices such as sample entropy and the SD1/SD2 ratio. Attention is conditioned on both signal quality and time-to-delivery, enabling context-guided attention allocation. Information exchange between the two branches is achieved through bidirectional cross-attention, while adaptive integration is performed using a context-aware gating mechanism. The model was evaluated on the CTU-UHB database comprising 547 patients using five-fold patient-level cross-validation. QT-MIL-FHR achieved a mean AUROC of 0.76 ± 0.04 with stable performance across folds. Ablation studies demonstrated that removing contextual conditioning reduced predictive performance and increased variability, whereas using either signal quality or time-to-delivery alone failed to match the complete model. Attention analysis further revealed that the model consistently emphasized clinically relevant signal segments, while calibration assessment demonstrated acceptable agreement between predicted probabilities and observed outcomes. These findings suggest that explicitly incorporating clinical context into attention mechanisms can improve both predictive reliability and alignment with clinical interpretation. More importantly, this work introduces clinical attention allocation as a principled framework for integrating clinical knowledge into deep learning models for intelligent fetal monitoring.