From Prototype to Impact: Closing the BeatNet Journey

At the start of this project, we thought we were building a better ECG model.

By the end, we realized we were building something much bigger—a system designed to work in the real world.

---

The Question That Shaped Everything

Our journey began with a simple but ambitious question:

How can we make ECG analysis more accessible, reliable, and practical for everyday use?

On the surface, the problem seemed straightforward. With the rise of deep learning and foundation models, high-performance ECG analysis already exists.

But very quickly, we ran into a fundamental gap.

Most state-of-the-art models are built for 12-lead clinical ECGs, collected in controlled hospital environments.
In contrast, real-world wearable devices rely on single-lead signals, often noisy, inconsistent, and far from ideal.

That gap changed everything.

It wasn’t just a modeling problem anymore—it became a real-world system challenge.

---

When Models Failed

Our early approach followed a natural path: adapt existing multi-lead models to single-lead data.

It didn’t work.

Performance dropped. Predictions became unstable. Outputs were inconsistent.

What looked powerful in a research setting quickly broke down under real-world conditions.

That moment forced a shift in thinking.

Instead of asking,
“How do we improve the model?”

we started asking,
“What does it take to make this actually work in practice?”

---

What We Built Instead

That shift led us to rethink the entire pipeline.

By the end of the project, we developed a complete end-to-end ECG analysis system:

• A single-lead, multi-label classifier capable of detecting 17 cardiac conditions
• PQRST delineation, providing waveform-level insights instead of just predictions
• Confidence-based filtering, improving reliability in uncertain scenarios
• A unified pipeline from signal preprocessing → model inference → visualization

But more importantly, we stopped optimizing only for accuracy.

We started optimizing for:

• Robustness in noisy, real-world signals
• Interpretability that clinicians can actually trust
• Scalability for deployment beyond controlled environments

What emerged was no longer just a model—but a deployable system.

---

The Turning Point: Thinking Like Engineers, Not Just Model Builders

One of the most important lessons we learned is simple:

A strong model is not enough. A strong system is what creates impact.

Throughout this journey, we had to constantly balance trade-offs:

• Accuracy vs. latency
• Complexity vs. deployability
• Performance vs. reliability

We also realized that interpretability is not optional.

By integrating PQRST detection and visualization, we moved away from opaque predictions and toward outputs that provide real clinical meaning.

That shift—from “prediction” to “understanding”—was critical.

---

From Research to Reality

Looking back, the most defining transformation of this project was not technical—it was conceptual.

We moved:

• From prototype → deployable system
• From offline experiments → real-time thinking
• From black-box outputs → interpretable AI
• From hospital-only assumptions → wearable accessibility

Each step brought us closer to answering our original question—not in theory, but in practice.

---

Sharing the Work

As we reached the Final Design Review, we had the opportunity to present not just our results, but our journey.

Alongside the system itself, we created a project poster, demo video, and final presentation—translating months of technical work into a story that others could understand and engage with.

This process challenged us in a different way.

It forced us to step back and ask:

Can others see the value of what we built?

Learning how to communicate that value—clearly and effectively—became just as important as building the system itself.

---

What Comes Next

While this chapter is complete, the system we built is only a starting point.

Future work will focus on:

• Real-time integration with wearable devices
• Scalable cloud-based inference pipelines
• Further optimization for on-device deployment
• Clinical validation with real patient data

These steps will push the system closer to real-world adoption.

---

Final Reflection

This project was never just about ECG signals or machine learning models.

It was about learning how to take something that works in theory—and make it work in reality.

We learned how to deal with imperfect data, unclear constraints, and evolving requirements.
We learned how to build not just for performance, but for usability and trust.
And most importantly, we learned how to think beyond the model.

We didn’t just build an algorithm.

We built a system designed to exist outside the lab—to support real users, in real conditions.

And in doing so, we took one step closer to a future where accessible, real-time cardiac monitoring is not just possible—but practical.

Final Validation: Bridging Development and Delivery

This week, Team 3-Aventusoft transitioned from system development to the high-stakes phase of final validation and presentation refinement. With our end-to-end pipeline now fully functional, we are focusing on “impact-driven storytelling”—ensuring that our technical achievements translate clearly into clinical value. We have moved beyond the “how” of our models to focus on the “why,” preparing to showcase a complete decision-support pipeline for cardiac diagnosis.

Key Accomplishments This Week

• Full System Presentation Rehearsal: The team delivered a complete end-to-end presentation covering problem framing, system architecture, and our PQRST detection pipeline, receiving critical sponsor feedback to maximize clarity.
• Validated End-to-End Pipeline: We successfully demonstrated our model deployed on Azure Container Apps, achieving real-time inference speeds of approximately 3-4 seconds with live PQRST overlays.
• Finalized Dual-Pipeline Design: We confirmed the integrated workflow that combines 1D U-Net + TCN for heatmap-based PQRST detection with an adapted foundation model for multi-label classification.
• Enhanced Model Stability: By implementing dilated TCN blocks, we significantly improved signal understanding and reduced false positives in the detection of P and T waves.
• Refined Interpretation Strategy: We implemented confidence-threshold filtering (~0.9) and segment-level aggregation to ensure our predictions are both reliable and clinically interpretable.
• Prepared Final Deliverables: The project demo video and final presentation materials are now complete and ready for the upcoming final evaluation.

---

Next Steps: The Final Push

As we enter the concluding week of the project, our focus is on polishing our narrative and finalizing professional documentation.

• Clinical Impact Narratives: We are revising our presentation to emphasize the role of single-lead wearable devices and the accessibility gap in modern cardiac diagnosis.
• Technical Documentation & Research: Drafting is underway for the formal technical report and a potential research paper for conference submission.
• Simulated Real-World Workflow: We are preparing a demonstration that simulates a 30-second ECG acquisition flowing from the cloud to a clinician dashboard.
• Fiducial Analysis Justification: We will further clarify how our PQRST fiducial points enable the interval analysis necessary for detecting complex abnormalities.
• Sponsor Approvals: We are finalizing dataset documentation and obtaining sponsor approval for the external sharing of our project highlights.

The project remains strictly on schedule. We are ready to demonstrate a robust, AI-driven solution that bridges the gap between wearable data and clinical insight.

See you at the finish line!

Final Polish: From Prototype to Presentation

This week, the Team 3-Aventusoft group transitioned from the heavy lifting of system development to the critical phase of refining our demonstration and finalizing our technical documentation. With the core development and model training now complete, our focus has shifted toward storytelling, visual clarity, and ensuring the robustness of our end-to-end pipeline. We are no longer just building a tool; we are preparing to showcase a complete diagnostic solution.

Key Accomplishments This Week

• Refined Presentation Narratives: The team developed and polished the final presentation slides, incorporating sponsor feedback to improve visual quality and overall impact.
• Clinical Visualization Updates: We replaced inaccurate ECG visuals with real signals and enhanced the explanation of our core contributions, including PQRST landmark and conduction abnormality detection.
• System Architecture Clarity: Significant progress was made in communicating the system architecture by simplifying technical details and restructuring slides for better audience understanding.
• Cloud Deployment Success: The web-based application was successfully deployed to Azure, moving us closer to a publicly accessible and demonstration-ready prototype.
• Backend Integration: Initial efforts were made to connect our backend models with the frontend interface, with the goal of showing real-time predictions via web or mobile.
• Codebase & Pipeline Review: We conducted a thorough codebase review with our liaison to identify and address missing dependencies and files required for the full ECG Analyzer pipeline.
• Technical Documentation Progress: All sections of the final technical report have been drafted and are currently being consolidated for submission.

---

Next Steps: The Final Countdown

As we move into the final stretch before our project evaluation, our efforts are centered on delivery, validation, and professional polish.

• Final Technical Report: We are finalizing and submitting the comprehensive report, ensuring all datasets, methodologies, and architectures are fully documented.
• Full Mock Presentations: The team is scheduling and rehearsing a complete mock presentation with the liaison team to perfect our flow and clarity.
• Live Demo Readiness: We are completing the Azure deployment to ensure the application is fully functional with integrated backend models for a flawless live demonstration.
• High-Impact Visual Assets: We are preparing annotated PQRST landmarks, heatmaps, and high-quality device images to supplement our presentation slides.
• Packaging and Portability: The full project deliverable—including model weights and the ECG Analyzer class—is being packaged and tested on separate systems to ensure total reproducibility.

The project remains on schedule and nearing completion, well-positioned to deliver a fully functional and impactful final demonstration.

See you at the finish line!

Final Polish: From Prototype to Presentation

This week, the Hemotag team transitioned from the heavy lifting of system integration to the critical phase of refining our demonstration and finalizing our technical documentation. With the core ECG Analyzer system now fully integrated, our focus has shifted toward storytelling, clinical relevance, and ensuring the robustness of our end-to-end pipeline. We are no longer just building a tool; we are preparing to showcase a complete diagnostic solution.

---

Key Accomplishments This Week

• Integrated System Validation: We finalized the end-to-end ECG Analyzer pipeline, which seamlessly combines classification and PQRST delineation into a single demonstration-ready workflow.
• Inference Pipeline Stability: The team completed rigorous validation of the inference engine to ensure consistent input-output behavior for short ECG segments and stable multi-label predictions.
• Enhanced Visualization Dashboard: We refined our dashboard to improve the visual clarity of ECG signals and PQRST annotations, specifically tailoring the interface for high-impact presentation.
• Codebase Consolidation: All classification and delineation modules have been consolidated into a structured codebase, and our GitHub repositories are now finalized with the necessary training scripts and inference modules.
• Sponsor Feedback Integration: Based on insights from our liaisons, we updated the system and presentation to better emphasize clinical relevance, model interpretability, and real-world deployment.
• Robustness Testing: We continued testing the system against a diverse library of ECG samples, including both healthy and pathological cases, to ensure reliable performance across various cardiac conditions.

---

Next Steps: The Final Countdown

As we move into the final stretch before our capstone presentation, our efforts are centered on delivery and documentation.

• Final Design Report (FDR): We are integrating all system components, validation results, and evaluation metrics into our comprehensive final report.
• Presentation & Storytelling: The team is refining our final slides, focusing on the impact and usability of the Hemotag system to provide a compelling narrative for stakeholders.
• Poster & Demo Media: We are finalizing the project poster and a demonstration video to supplement our live presentation.
• Live Rehearsals: We have scheduled multiple rehearsal sessions to perfect our timing, flow, and clarity before the big day.
• Deployment Confirmation: We are seeking final confirmation on deployment expectations (local vs. cloud) to ensure a flawless live execution.

The Hemotag system is on schedule and nearing completion, well-positioned to deliver a fully functional and impactful final demonstration.

See you at the finish line!

Integrating the Vision: A Unified ECG Analysis Dashboard

This week, the Hemotag team reached a major milestone by moving from isolated model components to a unified, functional interface. Our focus shifted toward system synthesis, ensuring that our classification and delineation models don’t just work—they work together in a way that is clinically intuitive. We are no longer looking at raw data in a vacuum; we are now viewing integrated diagnostic outputs through a cohesive UI dashboard.

---

Key Accomplishments This Week

• Unified ECG Analyzer Pipeline: We successfully integrated the 17-label classification model and PQRST delineation into a single pipeline. This allows the system to process an ECG signal and simultaneously identify landmarks and diagnostic labels.
• Functional UI Dashboard Development: We developed a dashboard that visualizes ECG signals, overlays PQRST landmarks, and displays heart rate computed from R-R intervals. This interface provides a clear, real-time visualization of model predictions.
• Clinical Prediction Refinement: To reduce false positives and ensure reliability, we implemented improved thresholding strategies (e.g., 20.9 confidence) for our classification outputs.
• Enhanced PQRST Delineation: We improved the accuracy of T-wave detection by applying Gaussian label refinement and enhanced post-processing, resulting in much more stable predictions.
• Deployment Readiness with Docker: The entire application has been successfully containerized using Docker. This ensures the system runs correctly in local environments and is ready for broader deployment.

---

Next Steps: Finalizing the Prototype

As we look toward the final presentation in April, our efforts are pivoting from backend engineering to presentation and deployment optimization.

• Finalizing Inference Formats: We are refining the pipeline to ensure it handles short (5-10 second) ECG segments and outputs the 17 diagnostic labels in a deployment-ready format.
• Pathological Case Studies: The team is preparing a variety of healthy and pathological ECG case studies to demonstrate the system’s robustness during the final demonstration.
• Azure Cloud Deployment: Once access is granted, we will transition our Docker-based local deployment to the Azure cloud.
• Final Design Report (FDR): We have begun consolidating our technical work into the Final Design Report, focusing on the clinical relevance and real-world impact of our system.
• Liaison Alignment: We are coordinating with liaison engineers to finalize confidence thresholds and lead configurations (Lead I vs. Lead II) to match specific device constraints.

The Hemotag system is now a fully integrated prototype, ready to bridge the gap between technical data and clinical insight.

See you next week!

Bridging the Gap: From Model Refinement to Containerized Deployment

This week, Team 3 transitioned from core development into the critical phase of system integration and deployment. While our previous efforts focused on the mathematical nuances of ECG signal processing—our current challenges involve ensuring our complex deep learning models are portable, reproducible, and ready for clinical cloud environments.

We have moved beyond perfecting individual scripts to building a cohesive, containerized pipeline that the sponsor team can reliably deploy and verify.

---

Key Accomplishments This Week

• Finalized ECG Classification Pipeline The core development of our ECG classification model is now largely complete. We implemented a robust pipeline and associated Python classes that handle everything from raw signal processing to generating final diagnostic predictions.
• Containerization via Docker To ensure the system remains functional regardless of the local computing environment, we successfully tested the model using Docker. This containerization is a prerequisite for our upcoming move to the cloud, ensuring that “it works on my machine” translates to “it works in the clinic.”
• Lead Performance Optimization Our initial performance analysis yielded a significant finding: models trained on Lead II ECG signals consistently outperformed those trained on Lead I. We are now evaluating how to best incorporate both model outputs into the final system to maximize diagnostic accuracy.
• Reproducible Research Repository We organized our training scripts and fine-tuned foundation models into a structured GitHub repository. This allows the sponsor team to reproduce our results exactly, ensuring transparency and long-term utility for the Hemotag system.
• Fiducial Point Detection Progress Development continued on the fiducial point detection component, focusing on accurately locating specific landmarks within the ECG signal recordings to support more granular cardiac analysis.

---

Next Steps: Deployment and Final Documentation

As we approach the final stretch, our focus shifts from “building” to “sharing.” The upcoming spring break will be a high-intensity period dedicated to finalizing our external infrastructure.

• Azure Cloud Integration Once cloud access is granted, we will deploy the full classification pipeline to Azure. This will enable external testing in a live, hosted environment, moving us one step closer to a production-ready tool.
• Comprehensive Technical Documentation We are preparing deep-dive documentation covering our model architecture, training methodologies, and evaluation results. This ensures that the technical “why” behind our “how” is clearly preserved.
• Prototype Demonstration Preparation With the March 31 prototype presentation fast approaching, the team is busy preparing UI demonstrations and saliency map visualizations. These visualizations will help stakeholders “see” what the model sees when it makes a classification.
• Liaison Coordination We are working closely with liaison engineers to secure demo data access and finalize the training data availability to meet all sponsor expectations.

Our system is now exiting the laboratory and entering the deployment phase, bringing us closer to a fully validated Hemotag ECG Classification System.

See you next week!

From Refinement to Reliability: Mastering Temporal Context and Clinical Precision

This week marked an important step forward for Team 3. We moved beyond general model tuning and focused on improving the temporal understanding of cardiac cycles. By integrating more advanced architectures and physiological constraints, our system is evolving from simply detecting peaks to analyzing their role within a full heartbeat structure.

Rather than only asking whether the model can detect a waveform, we are now ensuring that it interprets each peak within a physiologically consistent cardiac cycle.

---

Key Accomplishments This Week

Expanded Temporal Context with TCN Architecture

We integrated Temporal Convolutional Network (TCN) blocks into our fiducial delineation model. This expanded the receptive field from approximately 50 ms to 500 ms, enabling the model to analyze nearly an entire heartbeat cycle rather than short local segments.

This broader temporal context improves the model’s ability to understand relationships between waves within the PQRST complex.

Improved T-Wave Detection Accuracy

T-wave detection has historically been one of the most challenging components due to its lower amplitude and higher variability. By leveraging the new TCN architecture and refining post-processing logic, we reduced the T-wave detection error to approximately 3 ms, significantly improving delineation precision.

Implemented Physiological Guardrails

To prevent unrealistic model outputs, we introduced several physiological constraints into our post-processing pipeline. A 250 ms minimum interval rule between peaks and an “allow orphans” logic help prevent over-classification and reduce physiologically inconsistent predictions.

Standardized Lead Selection and Data Augmentation

We finalized our protocol for selecting the optimal lead (Lead I vs. Lead II) using the MIMIC-IV test set, ensuring an unbiased evaluation independent of sponsor data.

In addition, we strengthened our training pipeline with time-shifting augmentation applied simultaneously to signals and labels, preventing peak misalignment during training.

---

Next Steps: Validation and Final Demonstration

As we move into the final phase of the project, our focus will shift toward real-world validation and preparing for the final presentation on March 24th.

Sponsor Signal Validation
We will evaluate the updated model using Hemotax recordings provided by the sponsor, allowing us to assess performance under noisier and more realistic signal conditions.

Clinical Cross-Verification
We will work with liaison engineers to verify complex predictions from our 6-patient dataset, particularly cases where the model produces multiple high-confidence outputs such as PVC and Left Bundle Branch Block.

Regulatory Documentation Alignment
We will review and incorporate documentation guidelines provided by the sponsor to ensure our model development and reporting structure align with future FDA-oriented regulatory expectations.

Final Demo Strategy
The team will also finalize the demonstration approach for the Final Design Review (FDR), deciding whether to use simulated files or a direct connection to the Hemotax device.

Our system is now transitioning from model refinement toward validation, documentation, and final demonstration readiness.

See you next week!

From Performance to Production: Finalizing Lead Strategy and System Integration

This week marked a transition from experimentation to refinement. After weeks of tuning, testing, and validating, we are now converging toward final design decisions—solidifying lead selection, strengthening reproducibility, and unifying the entire system into a single deployable prototype.

Rather than asking “Can it work?”, we are now asking “Can it work consistently, cleanly, and in production?”

---

Key Accomplishments This Week

Finalized Lead Selection Through Benchmarking

We completed a detailed comparison between Lead I and Lead II for single-lead classification. Results showed that Lead II improves AUROC for infarction detection, helping close the previous performance gap. This marks an important architectural decision for final deployment.

Integrated Reproducibility Protocols

To ensure scientific rigor, we implemented fixed random seeds across training runs. Our model now consistently achieves 98–99% accuracy across repeated experiments, strengthening confidence in stability and repeatability.

Built a Unified End-to-End Prototype

We developed a fully integrated interface that visualizes raw ECG signals, overlays PQRST fiducial points, and displays saliency maps in one application. This transforms our system from separate scripts into a cohesive diagnostic tool.

Refined Training and Learning Rate Strategy

We experimented with higher learning rates (0.005) versus lower rates. While higher rates accelerated training, lower rates produced smoother convergence and better validation stability—guiding our final optimization strategy.

---

Next Steps: Precision and Deployment Readiness

In the coming week, we will:

• Close the T-Wave Gap: Improve T-wave detection accuracy, our current weakest fiducial component.
• Implement Advanced Physiological Post-Processing: Eliminate impossible outputs (e.g., duplicate T-waves or unrealistic proximities).
• Draft Dual Documentation Tracks: One research-oriented manuscript and one technical report aligned with future FDA pathways.
• Begin Cloud Deployment Research: Explore Azure hosting to mirror sponsor infrastructure.
• Validate Against Sponsor Data: Cross-check Lead II outputs with real-world patient samples.

---

We are no longer just optimizing models—we are shaping a clinically aligned, reproducible, deployment-ready ECG AI system.

See you next week!

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!