Linear ECG-lead transformations (LELTs) are used to estimate unrecorded target leads from a number of recorded basis leads. Generalized (GEN) LELTs are developed using data from multiple subjects, while patient-specific (PS) LELTs are developed using data of a single subject. The estimation performance of PS LELTs has been reported to be superior to what is achieved using GEN LELTs.
Recent work has demonstrated that GEN LELTs should be developed using ECGs belonging to different diagnostic classes. However, PS LELTs are often developed and tested on ECGs of the same diagnostic class. This potentially leads to overly optimistic performance metrics. In this research, we compare the estimation performance of PS and GEN LELTs in the presence of conduction changes that occur after the development of the PS LELTs. This is performed for LELTs with target leads V1, V3, V4 and V6 and basis leads I, II, V2 and V5.
First, two ECGs per subject were selected from secondary 12-lead ECG data of 25 subjects. More precisely, one normal ECG and one ECG showing signs of incomplete right bundle branch block were selected and used to assemble a training and a testing dataset respectively. Second, 25 different PS LELTs were developed using the training dataset. Third, one GEN LELT was developed using 12-lead ECG data of 171 different subjects. Forth, the estimation performance of the GEN LELT and each PS LELT was quantified using root-mean-squared-error (RMSE) values calculated between the estimated and the recorded target leads of the testing dataset.
Mean RMSE values associated with the PS and the GEN approach were found to be {V1:59.3µV; V3:170.8µV; V4:96.9µV; V6:76.0µV} and {V1:60.5µV; V3:156.0µV; V4:80.7µV; V6:51.5µV} respectively. Our findings indicate that changes in the cardiac conduction can reduce the estimation performance of PS LELTs to a level similar to that of GEN LELTs.