Attention-Based Anomaly Detection in Hospital Process Event Data

The increasing complexity of processes makes it difficult to monitor them efficiently. This also includes the care paths in hospitals, whose process reliability is highly relevant, since unnoticed process deviations can have major consequences. For this purpose, a classification model based on self-attention is presented, which derives process logics from the underlying event data and allows to detect event-based anomalies in process flows without prior process knowledge. With an F1 score of 0.93, the model achieves a very good detection accuracy compared to other approaches which achieve F1 scores of 0.65 - 0.75 for classification tasks in similar domains. This can support process mining activities and makes the monitoring of hospital processes more reliable.