July 14, 2025
Day 35 -Advancing ECG Diagnostics Through Deep Learning
What I Learned
Today at my internship, I focused on the development and evaluation of deep learning models designed to classify ECG signals using the PTB-XL dataset. The goal of this work is to improve the accuracy of cardiovascular disease diagnosis by leveraging various forms of data representation—specifically time-domain, frequency-domain, and image-based spectrogram data. This effort forms a core part of the broader research project I’m contributing to, which seeks to create a multimodal diagnostic system powered by artificial intelligence.In parallel, I also began preparing the data and framework for training a 2D CNN model on spectrograms generated through wavelet transforms (also known as scalograms). These visual representations of ECG signals offer another layer of abstraction and are expected to capture important patterns that may not be easily visible in raw waveform or frequency data. I ensured that the spectrogram images were correctly matched to their corresponding labels and that they were split into training, validation, and test sets accordingly. The actual training of the 2D CNN will begin next.Beyond model development, I spent time reviewing the label mapping process, verifying that each dataset had the correct format and balance across the target classes. These steps are crucial for ensuring fair evaluation and avoiding model bias. In summary, today’s work was focused, technical, and forward-moving. Each model I’m building plays a key role in the larger system we are designing—one that will ultimately merge these modalities into a hybrid architecture for enhanced diagnostic accuracy. As we progress, we will also explore explainable AI techniques like SHAP and LIME to provide transparency into the model’s decision-making process, an essential feature for medical applications.
Blockers
No blockers.
Reflection
Today at my internship, I focused on training and evaluating three separate deep learning models for ECG classification using the PTB-XL dataset. I’ve already completed a 1D CNN on time-domain data, which performs well (~90% accuracy). That model gave me a strong benchmark to work from. Most of my time today went into refining a 1D CNN + Transformer model trained on frequency-domain (FFT) features. It hasn’t matched the benchmark yet, which pushed me to think more critically about feature quality, architecture choices, and tuning strategies. I also prepped wavelet-based spectrograms for a 2D CNN, which will be trained next. This process reminded me how important it is to approach problems from multiple angles. Every model tells a different story about the same data. Though progress can be slow, I’m starting to better understand how deep learning can truly support real-world diagnostics.