Systematic benchmark of reduced-lead configurations for 12-lead ECG reconstruction: multi-model evaluation across all possible subsets

Figures

• 

  Figure 1. Lead information structure (computed on 3,000 randomly sampled PTB-XL training recordings). (a) 12 × 12 pairwise mutual information heatmap [Gaussian approximation: MI ≈ −0.5 log(1−r²)]. Dashed lines separate limb leads (I–AVF, rows/columns 1–6) from precordial leads (V1–V6, rows/columns 7–12). (b) PCA cumulative explained variance curve. Green dashed line at 90% (reached at PC 4); red dashed line at 99% (reached at PC 10).

• 

  Figure 2. Aggregate N–performance curves [averaged over LR, ridge, LightCNN and all C(12,N) combos]. Left panel: mean PCC (withheld leads only) vs. N = 1.11, with ±1 SD shaded band. Right panel: mean macro-F1 vs. N. Upper bound dashed line: F1 = 0.6751. Vertical dashed line at consensus Kneedle knee N = 4.

• 

  Figure 3. Per-model N–performance curves (best combination per N per model). Left panel: PCC vs. N. Right panel: macro-F1 vs. N. Four lines: LR (blue solid), Ridge (green dashed), LightCNN (orange solid), Transformer (purple dotted). Upper bound dashed line: F1 = 0.6751. Kneedle knee marked per model.

• 

  Figure 4. Pareto-front analysis (lead-count efficiency = 1/N). X-axis: acquisition efficiency = 1/N. Y-axis: macro-F1. Point size = PCC. Three panels (LR, Ridge, LightCNN). Pareto-front points: coloured circles labelled with N and best lead subset. Dominated points: grey crosses. Three clinical scenario recommendations annotated.

• 

  Figure 5. CLS sensitivity analysis (electrode-contact efficiency × macro-F1). (a) Composite Lead Score of the per-N optimal combination under five accuracy weights α ∈ {0.20, 0.35, 0.50, 0.65, 0.80}. (b) macro-F1 of the corresponding optimal combinations. Dashed horizontal line: 12-lead upper bound F1 = 0.6751. X-axis: N = 1 to 11.

• 

  Figure 6. Sensitivity heat maps (weight α × lead count N). Left: electrode contact count E of the per-N, per-α optimal combination. Right: macro-F1 of the same combinations. For N = 4–6 under low α, the optimal combination is always an all-limb configuration (E = 4), demonstrating the electrode-sharing benefit.

• 

  Figure 7. F1 benefit per electrode contact (electrode-based efficiency frontier). (a) Best achievable macro-F1 as a function of E (1–10). (b) Marginal ΔF1 per additional electrode contact. The largest single marginal gain occurs at E = 4 (completing the limb electrode circuit, enabling all six limb leads simultaneously); subsequent precordial contacts contribute progressively smaller increments.

• 

  Figure 8. Electrode-contact efficiency (E-based, CLS_E_F1) vs. lead-count efficiency (N-based, CLS_N_F1) comparison at α = 0.50. (a) macro-F1 of the per-N optimal combination under each metric (solid = E-based, dashed = N-based). (b) Electrode contact count E of the E-based optimal combination at each N. (c) CLS scores under both metrics. The E-based model consistently selects lower-F1 but more electrode-efficient configurations at N = 4–6.

• 

  Figure 9. Disease subgroup heatmap (LightCNN). Rows: NORM, MI, STTC, CD, HYP. Columns: top-3 configurations by F1 per N (∼33 columns). Cell colour: per-class F1 (viridis colormap, 0→1). Vertical dividers between N groups. The three deployment-recommended configurations (Tier 1: V6; Tier 2: I + II + AVR + AVF; Tier 3: 7-lead set) are enclosed in bold black bounding boxes with a ★ marker above the column header.

• 

  Figure 10. N-based sensitivity heat maps (CLS_N_F1). Left: lead count N of per-N per-α optimal combination. Right: macro-F1. Compare with Figure 6 (E-based): divergence is largest at N = 4–6, where E-based selects all-limb configurations (E = 4, lower F1) and N-based selects mixed limb-precordial configurations (higher F1).