Enhancing Interpretability and Broadening Horizons: Advancing the Single-Lead ECG-FM

This week represented a major leap from validating performance to establishing deep system interpretability and planning for expanded clinical capabilities. While previous milestones focused on adaptation, we are now pulling back the curtain on model behavior—using advanced visualization tools to ensure our AI aligns with clinical intuition while preparing for even more complex diagnostic tasks.

Rather than just reporting high numbers, we are now investigating the “why” behind the predictions and hardening our pipeline for real-time application.

Key Accomplishments This Week

• Integrated Model Interpretability Tools We successfully implemented saliency maps using attention weights to visualize where the model focuses during ECG classification. We observed the model correctly highlighting clinical features, such as deep S-waves and high T-waves, when identifying conditions like left bundle branch block.
• Benchmarked High-Performance Single-Lead Accuracy The team achieved exceptional AUROC scores of approximately 0.98 for PVC and Tachycardia detection using our fine-tuned single-lead data. We also identified a specific performance gap in Infarction detection (.73 vs. the original .92), prompting a new strategy to fine-tune on Lead II data to capture more distinct clinical markers.
• Refined PQRST Delineation and Post-Processing We enhanced our U-Net model by incorporating ResNet and DCN 1D layers and resolved previous stability issues by integrating a missing script for QRS peak derivation. New post-processing techniques—including moving averages and minimum distance constraints—have significantly refined our peak-picking cleanliness.
• Initiated Research and Technical Documentation As we move toward a final product, we have begun drafting a formal research paper. This documentation follows the structure of leading foundation model references, detailing our specific architecture, loss functions, and hyperparameters to ensure our work is scientifically reproducible.

Next Steps: Toward a Unified Clinical Product

In the coming week, we will:

• Validate Data Integrity and Rule Out Leakage: Perform a rigorous patient-level review of our testing sets to verify that our high accuracy scores are not influenced by data leakage.
• Fine-Tune for Ejection Fraction (EF): Utilize a new sponsor-provided dataset of 200–300 patients to train the model on categorical heart-pumping efficiency thresholds.
• Experiment with Multi-Lead Training: Test a hybrid approach using both Lead I and Lead II data to increase sample volume and model robustness.
• Consolidate System Integration: Merge our classification, fiducial detection, and saliency mapping into a single, unified application.
• Prepare for Real-Time Presentation: Develop the backend logic to process binary files from wearable devices instantly, moving us from offline analysis to a live diagnostic demo.

By bridging the gap between raw performance and clinical interpretability, we are ensuring that our single-lead system is not only accurate but also trustworthy for the medical professionals who will use it.

See you next week!

Refining Stability and Clinical Plausibility: Validating the Single-Lead ECG-FM

This week marked a transition from initial adaptation to rigorous validation and system stabilization. Our focus shifted toward evaluating the fine-tuned ECG foundation model on real-world single-lead inputs and enforcing physiological constraints to ensure our outputs meet clinical standards.+2

Rather than just proving feasibility, we are now hardening the system—moving from experimental code to a reproducible, high-performance diagnostic pipeline.+1

Key Accomplishments This Week

• Validated High-Performance Single-Lead Adaptation We evaluated the fine-tuned ECG-FM on duplicated single-lead inputs using real-world samples provided by our sponsor. The model achieved a strong AUROC of approximately 0.96, demonstrating that single-lead adaptation can reach performance levels comparable to dual-lead baselines.
• Optimized Temporal Aggregation for Prediction Stability By comparing 10-second and 30-second inference windows, we observed that longer temporal aggregation significantly improves stability. This strategy reduces inconsistent arrhythmia predictions, providing a more reliable output for clinical review.
• Advanced PQRST Delineation with Physiological Constraints We improved our post-processing by enforcing rules such as temporal ordering and minimum inter-peak intervals. By clustering nearby candidate peaks, we successfully reduced duplicate detections and began addressing over-generation issues in T-wave localization.+1
• Refined Clinical Plausibility and Thresholding In collaboration with sponsor feedback, we analyzed multi-label outputs—including PVC, tachycardia, and bundle branch blocks—to assess their clinical plausibility. We investigated thresholding strategies to suppress low-confidence or physiologically impossible diagnoses in a deployment setting.+1
• Initiated Model Interpretability Research We began exploring saliency and attribution maps to localize PVC-related regions. This work supports the debugging of both classification and fiducial detection, ensuring the model is looking at the correct features for its predictions.

Next Steps: Toward a Deployment-Ready System

In the coming week, we will:

• Benchmark window aggregation (10s vs. 30s) alongside tuned confidence thresholds to quantify the trade-offs between system responsiveness and prediction stability.
• Refine T-wave localization by further adjusting temporal windows and confidence filtering within the post-processing pipeline.
• Finalize patient-level data splits and document preprocessing workflows to support future regulatory-oriented documentation.
• Package the fine-tuning pipeline and inference scripts into a clean, reproducible repository for team sharing and review.
• Draft a structured technical report detailing the single-lead adaptation approach, experimental setup, and key performance results.

With our core performance validated and physiological constraints in place, we are moving closer to a cohesive, deployment-oriented system.

See you next week!

From Validation to Adaptation: Fine-Tuning the ECG-FM for Single-Lead Deployment

This week marked our transition from diagnosing limitations to actively adapting our models for deployment. With last week’s findings confirming that naive channel duplication is insufficient, the team shifted focus toward fine-tuning the ECG foundation model directly on duplicated single-lead inputs and tightening the end-to-end inference pipeline for real-world use.

Rather than treating the foundation model as a fixed black box, we began re-shaping it to better reflect the statistical properties and constraints of wearable ECG data. This week was less about proving what doesn’t work—and more about building what will.

Key Accomplishments This Week

Initiated Fine-Tuning on Duplicated Single-Lead Inputs
We set up the fine-tuning pipeline starting from a pretrained ECG-FM checkpoint that had not yet been biased by 12-lead-specific downstream tasks. Single-lead ECG signals were duplicated across channels to match the model’s expected input format, and supervised fine-tuning was launched for multi-label arrhythmia classification. Early training curves indicate improved stability compared to directly using the frozen multi-lead model, suggesting the model is beginning to adapt to the single-lead distribution shift.

End-to-End Inference Pipeline Prototyping
We implemented a standardized Python-based inference wrapper that takes a 5-second single-lead ECG segment as input and returns calibrated probabilities across all 17 diagnostic classes. This wrapper formalizes preprocessing, normalization, segmentation, and post-processing steps into a single callable interface, laying the groundwork for future cloud and mobile deployment.

PQRST Delineation Integration Progress
We continued refining the PQRST delineation pipeline and began aligning its input/output schema with the classification module. This included unifying windowing strategies, timestamp conventions, and output formats for fiducial points, which is critical for downstream clinical interpretation and UI integration. Initial tests confirm that both pipelines can now operate on consistent 5-second windows without manual intervention.

Calibration and Post-Processing Strategy Design
Building on last week’s observation of rare clinically implausible label combinations, we began drafting rule-based constraints and threshold calibration strategies to reduce invalid multi-label outputs. This included early experiments with probability threshold tuning and label co-occurrence filtering to improve clinical plausibility without over-constraining the model.

Sponsor Checkpoint and Deployment Alignment
We shared our fine-tuning plan and early pipeline design with Aventusoft mentors. Sponsor feedback reinforced the importance of treating the inference interface as a “product surface,” not just a research artifact—emphasizing clear contracts for inputs, outputs, latency expectations, and failure modes. This helped anchor our development priorities around deployability rather than pure model performance.

Next Steps: Toward a Deployment-Ready System

In the coming week, we will:

• Continue fine-tuning the ECG-FM on duplicated single-lead data and conduct controlled comparisons against frozen baselines.
• Run more systematic evaluations on sponsor-provided ECG samples to assess real-world generalization.
• Finalize the unified inference API for both classification and PQRST delineation, including standardized JSON-style outputs.
• Begin profiling latency and memory usage to inform future model compression and mobile deployment planning.

With the core assumptions now validated and the adaptation pipeline in motion, the project is moving from experimental prototyping toward a cohesive, deployment-oriented system. The foundation is set—now it’s about refining and hardening the stack.

See you next week!

From Assumptions to Evidence: Validating Single-Lead Limits and Refocusing for Deployment

This week was a pivotal checkpoint for Team Aventusoft. After establishing our preprocessing and modeling foundations, we stress-tested a key assumption: whether a pretrained multi-lead ECG foundation model could reliably operate on single-lead data through simple channel duplication. Through large-scale evaluation and detailed analysis, we confirmed that this approach introduces significant performance limitations.

Rather than viewing this as a setback, the results gave us clarity. The team aligned with our sponsor on a more principled direction—fine-tuning directly on duplicated single-lead inputs and prioritizing a deployment-ready inference pipeline that reflects real-world constraints.

These findings mark a transition from exploratory validation into targeted adaptation and system integration.

Key Accomplishments This Week

Single-Lead Adaptation Stress Testing:
We evaluated the pretrained ECG-FM classification model using single-lead ECG inputs duplicated across 12 channels. Testing on approximately 80,000 ECG segments revealed clear performance degradation compared to native multi-lead inputs, confirming that naive duplication is insufficient for stable downstream use.

Multi-Label Output Analysis and Clinical Plausibility Checks:
We analyzed predictions across all 17 diagnostic labels and verified that most outputs remained clinically plausible. However, we also identified rare invalid combinations (such as simultaneous LBBB and RBBB), reinforcing the need for improved calibration and post-processing strategies.

Progress on PQRST Delineation Pipeline:
Using LUDB data, we advanced our PQRST delineation model by training on 5-second windows with overlapping inference and Gaussian-encoded labels. Evaluation using tolerance-based metrics (±10 ms) and visual inspection of predicted versus ground-truth fiducial points demonstrated improved localization accuracy.

Sponsor Alignment on Deployment Requirements:
We presented both quantitative results and qualitative visualizations to Aventusoft mentors. Sponsor feedback emphasized the importance of a standardized, deployment-ready inference pipeline with clearly defined input–output interfaces for both disease classification and fiducial detection.

Next Steps: Fine-Tuning and Pipeline Consolidation

With assumptions validated and direction aligned, the focus now shifts to execution. Next week, we will begin fine-tuning the ECG-FM model directly on duplicated single-lead inputs, rather than relying on a pretrained multi-lead checkpoint.

In parallel, we will finalize a standardized inference interface that accepts 5-second single-lead ECG segments and outputs calibrated probabilities for all 17 diagnostic classes. We will also continue strengthening the PQRST delineation pipeline by validating performance on more challenging, real-world ECG signals and consolidating classification and delineation components into a unified Python-based library.

Finally, we will test robustness using additional ECG samples provided by Aventusoft to better assess real-world generalization.

The project remains on schedule, and this week’s results provide a strong technical foundation for the fine-tuning and integration work ahead.

See you next week!

From Preprocessing to Fine-Tuning: Building the Foundation for Team Aventusoft

This was a pivotal week for our team! After spending time establishing our core data strategy, we successfully completed the preprocessing pipelines required to align single-lead data with the ECG-FM foundation model. We spent the week diving deep into the technical architecture of the fairseq framework, ensuring that our setup is robust and ready for the computational demands of model training.

These foundational steps have moved us from the planning phase into active experimentation, setting the stage for our first supervised fine-tuning runs using single-lead duplicated ECG signals.

Key Accomplishments This Week

• Multi-Channel Signal Synthesis and Validation: We completed preprocessing pipelines to extract Lead I ECG signals and duplicate them across 12 channels. This ensures full compatibility with the ECG-FM foundation model architecture while maintaining signal alignment and normalization.
• Architectural Mastery of fairseq-signals: The team conducted an in-depth study of the fairseq and fairseq-signals frameworks. We analyzed model architecture, training pipelines, and the specific mechanisms used to load and freeze pretrained checkpoints for downstream tasks.
• Fine-Tuning Infrastructure Ready: We reviewed prior ECG-FM training scripts and documentation to identify the necessary manifests, inputs, and hyperparameters. This preparation was essential to ensure our upcoming fine-tuning runs are efficient and stable.
• Data Integrity and Label Mapping: To ensure clinically meaningful results, we established a plan for patient-level training, validation, and test splits to avoid data leakage. We also began generating supervised diagnostic labels for our dataset using clinical metadata and established mappings.

Next Steps: Launching Experiments and Benchmarking

With the infrastructure finalized and the project remaining on schedule, next week is all about execution. We will launch the first fine-tuning experiments using the ECG-FM pretrained checkpoint with our single-lead duplicated inputs.

A major priority will be monitoring training stability and loss convergence to verify the integration of our new preprocessing pipeline. We will also begin comparing our fine-tuned results against simpler single-lead baseline models to contextualize performance gains. Finally, we will coordinate with our liaison engineers to confirm preferred evaluation metrics like AUROC and F1-score, ensuring our model meets the high standards required for clinical disease classification.

See you next week!

New Year, New Architecture: Setting the Stage on HiperGator

This week, Team Aventusoft hit the ground running with a clear mission: adapting the ECG-FM foundation model for single-lead application.

Our main focus was solving the distribution mismatch challenge. After evaluating our initial approach, we finalized a revised technical strategy to fine-tune the model by duplicating single-lead ECG signals across 12 channels. To support this, we moved our operations to the HiperGator high-performance computing cluster. We successfully downloaded the complete MIMIC-IV-ECG dataset and installed the “fairseq-signals” framework, ironing out all the environment and dependency wrinkles along the way.

With the infrastructure in place, we implemented preprocessing scripts to generate the necessary dataset manifests. The rig is ready. Next steps include generating diagnostic labels and launching our first full-scale GPU fine-tuning run to see how our new strategy performs.