Anomaly Detection in Cardiotocography Time Series

Cardiotocography (CTG) is a diagnostic technique employed for fetal monitoring, widely used in medical practice, especially in labor. The key role of this technique is to provide real-time information about the fetal condition, which impacts decisions regarding medical interventions. Existing computerized solutions for CTG monitoring do not result in a reduction of negative fetal outcomes. The implication of a deep learning approach may resolve this issue. A semi-supervised deep learning approach for point anomaly detection was applied to streaming CTG signals and compared with a moving average solution. The methodology involved training three neural networks (CNN, TCNN, LSTM) to predict the next value by using a context window of sizes 10, 20 and 30. Anomalies were identified when the prediction error exceeded a predefined threshold. To set a threshold, prediction errors were computed for manually labeled data. Next classification tests were conducted using different thresholds to find the one with the highest F1-score. The highest F1-score of 0.72 was achieved by the TCNN model trained on a context window of size 10. Subsequent work may include predicting sequence anomalies, and the results might be employed to develop a real-time alarm system.