DT-CTNet: A clinically interpretable diagnosis model for fetal distress

Clinically, Fetal Heart Rate (FHR)-based intelligent cardiotocography classification to diagnose fetal well-being is of utmost importance for obstetricians and gynecologists. However, the current Fetal Distress Diagnosis (FDD) algorithms based on Artificial Intelligence (AI) methods are designed as black box models that usually result in poor performance on clinical interpretation and lack explicit diagnostic evidence. This paper focuses on two aspects: enhancing the accuracy of FD and providing explanations for the model's decisions. Specifically, we propose a novel clinically interpretable dual-stream AI architecture for FDD. The two streams embed multifeature category representations and the original FHR into the Digital Twin model (DT) and the deep convolutional neural Networks-based Case Tracking model (CTNet), respectively, thus marking as DT-CTNet. And the former makes the input-outcome dependency transparent, while the latter provides the samples that are the most similar to the object being diagnosed. By this design, not only is high-precision FDD achievable, but a transparent explanation with global, local and instantiation interpretation can be obtained. To the best of our knowledge, this work is the first to provide an interpretable diagnosis of fetal distress. The proposed method is comprehensively evaluated from five aspects: classification model accuracy, explanation of the classification model decision, consistency of the evidence explanation, pathological difference in visual analysis, and a comparison with state-of-the-art methods. Experimental results, which are performed on a challenging dataset, show that the proposed fetal distress diagnosis method outperforms the state-of-the-art methods and has a high accuracy (0.963).

Automated summary

This paper proposes a novel and interpretable method for diagnosing fetal distress, which improves accuracy and provides explanations for the model's decisions. Experimental results show that the proposed method outperforms state-of-the-art methods with a high accuracy of 0.963.