{"id":1163,"date":"2026-02-07T04:15:05","date_gmt":"2026-02-07T04:15:05","guid":{"rendered":"https:\/\/www.ippd.ufl.edu\/blogs\/ay2526team03\/?p=1163"},"modified":"2026-02-07T04:15:06","modified_gmt":"2026-02-07T04:15:06","slug":"week-4-2","status":"publish","type":"post","link":"https:\/\/www.ippd.ufl.edu\/blogs\/ay2526team03\/2026\/02\/07\/week-4-2\/","title":{"rendered":"Week 4"},"content":{"rendered":"\n
\n
\"\"<\/figure>\n<\/figure>\n\n\n\n

From Validation to Adaptation: Fine-Tuning the ECG-FM for Single-Lead Deployment<\/h2>\n\n\n\n

This week marked our transition from diagnosing limitations<\/em> to actively adapting our models for deployment<\/em>. With last week\u2019s 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.<\/p>\n\n\n\n

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\u2019t work\u2014and more about building what will.<\/p>\n\n\n\n

\n\n\n\n

Key Accomplishments This Week<\/h3>\n\n\n\n

Initiated Fine-Tuning on Duplicated Single-Lead Inputs<\/strong>
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\u2019s 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.<\/p>\n\n\n\n

End-to-End Inference Pipeline Prototyping<\/strong>
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.<\/p>\n\n\n\n

PQRST Delineation Integration Progress<\/strong>
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.<\/p>\n\n\n\n

Calibration and Post-Processing Strategy Design<\/strong>
Building on last week\u2019s 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.<\/p>\n\n\n\n

Sponsor Checkpoint and Deployment Alignment<\/strong>
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 \u201cproduct surface,\u201d not just a research artifact\u2014emphasizing clear contracts for inputs, outputs, latency expectations, and failure modes. This helped anchor our development priorities around deployability rather than pure model performance.<\/p>\n\n\n\n

\n\n\n\n

Next Steps: Toward a Deployment-Ready System<\/h3>\n\n\n\n

In the coming week, we will:<\/p>\n\n\n\n