CORDOC: A hybrid AI approach to ECG interpretation
The hypothesis of this dissertation is the addition of the qualitative information about the subject would improve the overall recognition efficiency of a NN ECG classification system as compared to the NN system without utilizing this additional information. This philosophy was implemented in the design and development of CORDOC, an ECG arrhythmia classification system for classifying 15 types of ECG rhythms, including normal sinus rhythm. CORDOC utilized both the non-symbolic ECG time domain data, as well as available symbolic, qualitative information about the subject. CORDOC essentially consists of a neural network module to process the ECG data and a rule-based module to process the qualitative information. The overall recognition efficiency was 70.33 percent and 75 percent for the neural network module alone and the complete of CORDOC respectively. The average Specificity was 97.88 percent and 98.17 percent for the NN module alone and for the complete of CORDOC respectively. The average Positive Predictivity was 72.18 percent and 76.67 percent for the neural network module alone and CORDOC respectively.
The rule-based module overruled the NN module's classification for 18 of the test cases. This means, for 282 of the 300 cases, the rule-based module either agreed with the NN classification or did not have sufficient information to overrule the NN classification. Fourteen of these overrules resulted in a correct classification of the case, while the remaining four overrules still resulted in incorrect classifications. The overall sensitivity improved by 6.64 percent, and the Positive Predictivity improved by 6.22 percent over the NN module's classification. The number of errors were reduced by 15.9 percent for the complete CORDOC over the NN module alone.
When the number of classes were abstracted to 10 categories by grouping together the PVC related rhythms and the tachy-rhythms, the overall recognition efficiency of CORDOC was 82.2 percent. On further abstraction by grouping premature atrial contractions with atrial fibrillation cases, a recognition efficiency of 84.2 percent was seen for nine rhythm categories.
Thus the addition of rule-based system resulted in improved overall efficiency and the average Specificity over the NN module for classifying the 15 rhythms. The approach of integrating neural networks and rule-based system provides a very viable solution for ECG diagnosis in particular and medical diagnostic systems in general.