Automated localization and severity period prediction of myocardial infarction with clinical interpretability based on deep learning and knowledge graph

This paper presented an interpretable method for myocardial infarction (MI) localization and severity period prediction using 12-leads electrocardiograms (ECG) based on deep learning and knowledge graph. Firstly, the ontology structure of knowledge graph for MI intelligent diagnosis was established based upon the diagnosis logic and strategy of doctors, and ontology attributes and relationships between attributes were extracted. Then, the entity's attribute values including the beat morphology of QRS waves, ST segments and T waves were extracted along with the method based on DenseNet network and diagnostic rules. Once again, attribute values were linked to the ontology structure of domain knowledge graph. Furthermore, production rules were employed to reason MI diagnosis results. Finally, all the related experiments were conducted and verified with a high-quality ECG database. For the severity period prediction of MI patients, the average accuracy, sensitivity, specificity and F1 value were 93.65%, 94.86%, 97.76% and 94.27%. For MI localization, the F1 value of IMI, ASMI, AMI, EAMI, LMI, APMI and HC with single period and single infarction areas were 97.56%.93.83%. 79.65%.80.81%.87.18% and 70.59%, and the average F1 was 86.88%. Notedly, the overall accuracy was 100.00% for MI patients with the single period and multiple infarction areas and 95.16% for multiple periods and multiple infarction areas. These results all displayed the superiority of the proposed method compared with other deep learning methods, and the clinical interpretability with the knowledge graph of the patient was used to explain how the diagnostic results were given.

Automated summary

This paper proposes an interpretable method for myocardial infarction (MI) localization and severity period prediction using deep learning and knowledge graph. The method extracts attribute values based on ontology structure and diagnostic rules, and employs production rules for diagnosis reasoning. Experimental results demonstrate the superior performance of the proposed method in accuracy and F1 value.