Interactive Dashboard

The framework includes a comprehensive web-based dashboard built with Streamlit for interactive data exploration, training monitoring, and biomarker discovery.

Launching the Dashboard

To start the dashboard, run the following command from the root of the repository:

fishy dashboard

Once running, the dashboard will be available in your browser (typically at http://localhost:8501).

Key Features

Data Exploration

  • Interactive Spectrum Viewer: Filter and overlay individual samples to compare spectral signatures across classes.

  • Cluster Analysis: Visualize data topology using linear (PCA) and non-linear (t-SNE, UMAP) dimensionality reduction.

  • Statistical Insights: View mean spectral signatures with standard deviation shading and class-wise intensity distributions.

Training & Results

  • Real-time Monitoring: Watch training progress with live loss and accuracy curves.

  • Performance Analysis: Detailed confusion matrices, ROC/AUC curves, and class-specific precision-recall metrics.

  • Error Analysis: Use the “Misclassification Spotlight” to identify and visualize samples where the model was confidently wrong.

Interpretability & Biomarkers

  • Single Instance XAI: Explain individual predictions using Grad-CAM (for deep models) or LIME.

  • Biomarker Correlation Network: Visualize statistical links between top diagnostic peaks (r > 0.8) to identify redundant chemical drivers (isotopes/adducts).

  • Biomarker Stability Histogram: Track the frequency of specific m/z features as top-20 diagnostic markers across multiple samples.

  • Class Comparison: Highlight unique biomarkers in Gold/Silver directly on class-representative spectra to verify their alignment with physical peaks.

Leaderboard & Remote Data

  • SSH Proxy Jump: Seamlessly aggregate results from remote high-performance servers (e.g., VUW ECS) through 2FA-secured jump hosts.

  • Data Persistence: Save snapshots of the global leaderboard locally for offline analysis and thesis reporting.

  • Fold Stability: Visualize the consistency of model performance across cross-validation folds using the stability violin plot.