Approaches — dual-path late fusion vs early fusion¶
Approach A — dual-path late fusion (primary)¶
Each input gets its own Features + Mamba SSM; the two models converge at the output.
flowchart LR
SIG["signal\nraw ECG waveform"] --> F1["features A"] --> M1["Model Mamba SSM"]
SCAL["scalogram\nCWT heatmap"] --> F2["features B"] --> M2["Model Mamba SSM"]
M1 --> OUT["output\nN / V / S / F / Q"]
M2 --> OUT
Note: the sketch output reads N / V / S / F / S / Q (S twice, six entries). Assumed typo for N / V / S / F / Q. Flag if the sixth output is intentional.
Approach B — early fusion (alternative)¶
Signal and scalogram encoded separately, concatenated per time step, one Mamba SSM, one 5-logit head. Fewer params, joint representation. Risk: stronger branch dominates, needs aligned time axes.
Comparison¶
- Params/compute: B ~1 model; A ~2 models, roughly double train cost.
- Ablation: A gives per-modality scores (signal-only vs scalogram-only) for free; B needs extra runs to prove the scalogram helps.
- Robustness: A survives a noisy modality (one branch can outvote); B can collapse if fusion is poorly balanced.
- Fusion rule (A, undecided): average logits, learned per-class gate, or majority vote. Default: average logits, then per-class threshold tuning.
Decision — A is primary, built first¶
Build A first: train signal-only and scalogram-only branches, record per-branch macro F1 and V recall, then fuse. Build B second as the ablation. Keep whichever wins on DS2 macro F1 + V recall; report both.
Effect on plan¶
- plan.md is research-phased: literature review first, then baselines, Approach A experiments, Approach B ablation, final write-up.
- Loss/metrics in model.md apply unchanged to each branch.