Cross-modal reconstruction of cardiac signals
Formal thesis title: “Cross-Modal Reconstruction of Cardiac Signals: Decoding ECG-PCG Electromechanical Coupling Dynamics” (NUST, 2026).
Heart sounds capture the mechanical consequences of cardiac activity. Murmur grading, timing and shape depend on where a sound falls within the cardiac cycle, yet stethoscope-only settings have no simultaneous ECG to mark those phases. My thesis tests whether the heart sound’s relationship with the ECG can provide that timing reference.
- Question
- Can timing inferred from heart sounds support murmur characterisation for unseen patients?
- Method
- Conditional signal reconstruction, phase-based features and patient-grouped comparisons with equal-time and acoustic baselines.
- My contribution
- Developed reconstruction and phase-analysis workflows, designed experiments and patient-grouped evaluation, led the first-author manuscripts and a blind cross-device audit of the reconstruction.
Current finding · October 2026: A re-evaluation kept each patient’s recordings in one fold and fitted preprocessing on training data only. Expert-reviewed phase boundaries carried useful information for murmur characterisation. The tested synthetic-ECG timing route did not establish an advantage over equal-time windows in this patient-grouped dataset evaluation. This separates the value of accurate phase information from the still-unestablished benefit of automatically reconstructed timing.
Next question: Can a reliable automatic timing method preserve that phase information across patients and recording systems? How this leads to my doctoral question
- Group
Separate patients
Keep every recording from a patient within the same evaluation group.
- Fit
Use training data only
Fit preprocessing within each training fold, then apply it to the held-out data.
- Compare
Match the experiment
Compare timing constructions and a PCG-only neural baseline on the same patient folds.
- Evaluate
Measure each task
Assess murmur presence, intensity grade, timing and shape separately, with paired uncertainty estimates.
Compared on the same patient folds
Reference construction without cardiac-phase alignment.
Dataset annotations test the value of reliable phase boundaries.
Boundaries from reconstructed-ECG R-peaks, using fixed cycle fractions alone or fused with acoustic heart-sound evidence.
Learns from the heart sound directly, without phase boundaries.
Data roles, evaluation and limits
Paired ECG–PCG recordings support reconstruction development; separate-device paired recordings test reconstruction transfer. The public paediatric CirCor DigiScope heart-sound dataset is used for downstream murmur evaluation, with ten patient-grouped folds, macro-F1 for each task and paired patient-level bootstrap uncertainty. These are distinct data roles, rather than one shared clinical cohort.
The blind cross-device audit covered 338 record pairs from a different recording system. It did not establish successful transfer to that hardware. Expert-reviewed semi-automatic phase annotations are also distinct from an automatic segmenter; no fully automatic direct-PCG segmentation arm was tested in the October evaluation.
Current conclusions use the October 2026 patient-grouped evaluation, which supersedes the earlier thesis and congress analyses; those earlier results should not be cited as evidence for synthetic-ECG timing. Whether an improved or retrained timing method transfers reliably to new devices, and how it would fit a clinical workflow, remains unresolved. An inferred timing reference is not a replacement for measured diagnostic ECG.