발표시간: 16:00 - 17:00

제목: A Graph-Based Hybrid Framework for Cardiac Electrophysiology

초록:
This thesis studies activation-time error, controlled ECG generation and prediction with ECG-fitted graph models. Starting from classical shortest-path and Bellman theory, it gives a residual bound that accounts for two path depths and source-boundary error. A finite-source analysis then tracks changes in the winning source and the global time origin. Its source–gauge–observer extension bounds ECG changes for a fixed observation operator and a piecewise linear voltage template. The synthesis study separates activation and recovery controls and applies stated rules to generated signals. Direct local screening accepts 40 of 72 candidates. The sufficient bounds certify none of these points or the 24 separately tested intervals, indicating that tighter bounds are needed for useful continuous guarantees.

The inverse study compares independent source-time shrinkage (M1), a grouped source-neighbor prior (M3), and their combined method, M1 + M3. The combined method uses equal weights and fixed limits on relative group timing. Regularization improves simulated activation recovery and agreement between repeated fits. On a new right-ventricular timing pair, M3 predicts the local activation time (LAT) change with 0.953 ms RMSE and the ECG change with 0.01408 mV RMS error. Its prediction is fixed before the intervention simulation. Simpler priors, compared later, give similar results.

M1 + M3 lowers median baseline LAT error across ten starts from 4.125 to 3.354 ms and reduces the spread of their intervention predictions. However, its new-pair LAT prediction error is 7.5% higher than M3. In 14 clinical records, median relative ECG error (rRMSE) rises from 0.28587 to 0.54638, and no evaluated method meets the 5-ms repeatability criterion. These findings show that baseline accuracy, repeatability and prediction require separate tests. The results support conditional model analysis and prediction on shared anatomy; patient LAT accuracy remains untested.