Navigation

  • Home
  • About Me
  • About My Mentors
  • About My Project
  • My Blog

June 24, 2025

Day 21 - Strategic Shift in Dataset Handling

By Ayomide Jeje

What I Learned

Today, I made a key breakthrough in how I’m approaching the ECG dataset we’ve been working with. Initially, I had been removing the time-domain components to focus on transformed data like spectrograms or frequency features. However, after reviewing the structure and performance implications more closely, I decided to stop removing the time components and instead embrace them as a core part of the training data.This pivot in strategy was based on two realizations: The time-domain signals contain valuable raw information that can enhance the learning capacity of our models. By preserving these components, I open the door to training 1D CNN models more effectively and potentially exploring hybrid architectures that combine time and frequency features later on. The rest of the day was spent revisiting preprocessing steps and adjusting my data pipeline to retain the full waveform structure. It required going back and testing earlier stages of the workflow, but the change sets a better foundation for robust modeling going forward.

Blockers

No Blockers

Reflection

Today felt like a turning point in how I’m approaching the ECG data. Up until now, I’d been removing the time-domain components, thinking they were unnecessary compared to more processed forms like spectrograms or frequency features. But something wasn’t sitting right. I kept wondering if I was throwing away useful information too early in the pipeline. After some thinking and reviewing the data again, I made the decision to stop removing the time components. It wasn’t an easy call — it meant backtracking and reworking part of my process — but it felt like the right one. The raw time signals carry structure and nuance that models like 1D CNNs can pick up on. By keeping them, I’m giving the model a chance to learn more directly from the heartbeat patterns. This shift wasn’t just technical — it reminded me that data science is as much about intuition as it is about code. I don’t have all the answers yet, but I feel more confident that I’m moving in the right direction. The rest of the day I spent reworking my pipeline to reflect this change and planning out the next modeling steps. Sometimes progress isn’t about doing more — it’s about doing better. Today was one of those days.

Tags: Training Ai Models