June 26, 2025
Day 23 - Laying the Groundwork for Model Expansion
What I Learned
Today marked an important turning point in my internship journey as I officially began work on my additional models beyond the main ECG classifier. After completing the preprocessing pipelines and solidifying the performance of the earlier models, I decided it was time to extend my research into deeper multimodal territory. I spent the morning revisiting the datasets and setting up clean input pipelines for the new models. This included validating that the training, validation, and test splits were consistent across formats — especially since I’ll be comparing models built from different ECG representations like time-series, frequency-domain, and spectrogram data. Once my environment was set, I began coding the architectures for two new branches: one that fuses a 1D CNN with a BiLSTM for temporal learning, and another that incorporates attention mechanisms to better capture feature importance over time. My goal is to benchmark these against my earlier models to see how well they generalize and which fusion strategies yield the best performance on cardiovascular classification tasks.
Blockers
No Blockers
Reflection
Today felt like a shift — not just in what I worked on, but in how I’m thinking about this whole project. After spending the past few weeks focused on one core model, I finally began working on the additional models I had planned. It was a reminder that progress doesn’t always mean finishing something — sometimes, it’s about starting the next step with purpose. Setting up the foundations for the new models took time. I had to ensure the datasets were correctly aligned and preprocessed for each variation: time-series data for the 1D CNN-BiLSTM, frequency-domain inputs for the hybrid, and preparing to revisit spectrograms soon. There’s a mental shift when you go from refining one thing to coordinating multiple moving parts — and today, I really felt that shift. Starting the new models gave me energy. It reminded me why I care about this work in the first place: not just to build something that works, but to explore how different methods can work together. My goal is to create a system that fuses multiple insights, not just rely on one. Today was about laying the bricks for that bigger vision.