Quantum-optimized spatio-temporal transformer for multimodal cardiovascular risk prediction from clinical and wearable data

Figures

• 

  Figure 1. Class imbalance analysis of MIMIC-III, CAIR-CVD, and UCI heart datasets. (Distribution of cardiovascular risk classes across the three datasets, highlighting differences in disease prevalence and class skew).

• 

  Figure 2. Distribution of key clinical vitals across multimodal cardiovascular datasets (violin plots illustrating dataset-specific variability in age, systolic blood pressure, and cholesterol levels).

• 

  Figure 3. Feature correlation analysis across cardiovascular datasets (correlation heatmaps illustrating inter-feature relationships and redundancy patterns).

• 

  Figure 4. Missingness and temporal sparsity in the MIMIC-III clinical dataset (visualization of feature-level missing data proportions and irregular temporal sampling).

• 

  Figure 5. Proposed overall architecture of proposed QST-trans framework. The contextual.

• 

  Figure 6. Preliminary sanity-check ROC and precision–recall curves using randomized validation inputs.

• 

  Figure 7. Biomarker selection confidence: quantum-inspired EQI-BGWO vs. Classical feature selection.

• 

  Figure 8. Optimization convergence behavior of EQI-BGWO compared with classical swarm optimization.

• 

  Figure 9. Federated learning training dynamics: accuracy and loss convergence.

• 

  Figure 10. Privacy–utility trade-off under adaptive gradient sparsity in federated learning.

• 

  Figure 11. Normalized confusion matrix of the QST-trans cardiovascular risk prediction model.

• 

  Figure 12. Precision–recall curve for high-risk cardiovascular class using QST-trans.