Predicting high-performance liquid chromatography anomalies from partial observations using attention-based temporal learning and uncertainty-aware inference

High-Performance Liquid Chromatography (HPLC) anomaly detection is commonly performed retrospectively using complete pressure traces, limiting its usefulness when only part of an ongoing experiment is available. We present the Early Attention-based Temporal HPLC framework (EAT-HPLC), a deployment-oriented approach for predicting experiment-level anomaly labels from partially observed HPLC pressure traces while providing model-internal temporal localization and predictive uncertainty estimation. Because the available dataset provides labels at the level of the complete experiment but does not include temporal annotations of AirBubble onset, the task is formulated as partial-observation anomaly prediction rather than onset-aware early anomaly detection. The framework was evaluated on 961 expert-labeled HPLC traces at observation ratios ranging from 20% to 100% using repeated stratified cross-validation. At 20% observation, EAT-HPLC achieved an anomaly-class F1-score of 0.866 ± 0.030 with a false-alarm rate of 16.9% ± 10.7% . The best mean performance occurred at 80% observation, with an anomaly-class F1-score of 0.945 ± 0.018 and a false-alarm rate of 6.6% ± 9.9% . EAT-HPLC also achieved the highest mean macro-F1 among the evaluated baselines across all observation ratios. These results demonstrate the potential of partial-observation HPLC monitoring and characterize the trade-off between the amount of trace information available and decision reliability, without implying that predictions necessarily precede the physical onset of an AirBubble event.