Every year, cardiovascular diseases take millions of lives worldwide. Doctors use ECG tests all the time because they are quick and non-invasive, but reading those signals by hand is still slow and tiring. When you have hundreds of patients, small details can easily get missed. That’s why I wanted to see if modern AI could help catch problems earlier and more reliably.

The everyday challenge in heart diagnostics
Cardiologists look at complex waveforms with P waves, QRS complexes and T waves. It works well, but it takes time and concentration. My thesis looked at whether computers could do a big part of this job automatically.

What I actually did
I used the well-known MIT-BIH Arrhythmia Database and trained three different deep learning models: Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Gated Recurrent Unit (GRU). These are popular techniques that learn patterns directly from the raw ECG signals instead of needing someone to pick out features manually.

The results that really stood out
The numbers speak for themselves. My CNN model reached an accuracy of 98.27%, that was the best performer. The GRU model came very close at 97.80%, while the simpler RNN got to 82.84%.

What pleased me most was how well the top two models handled both normal and abnormal beats. The confusion matrices showed only a few mix-ups, and those usually happened between similar types of abnormalities, which is quite normal in medical data. The training graphs also looked clean and the models learned steadily without major overfitting.

What this actually means in practice
These results tell us that deep learning can classify ECG heartbeats very accurately and extremely fast. CNNs turned out to be especially good at spotting the shapes and patterns in the signals, while GRU models handled the time-based connections between beats nicely.

In a real hospital setting this kind of system could:

• Help doctors review large numbers of ECGs much quicker
• Support continuous monitoring with wearable devices
• Make early warnings possible before symptoms get serious

It’s interesting to note that Apple Watch already uses similar technology to detect atrial fibrillation. My work goes a step further by identifying five different heartbeat classes with very high accuracy using open tools.

The road ahead
Of course there are still challenges, such as class imbalance in the data, but the strong performance of the CNN and GRU models gives me confidence that this approach has real potential. I believe AI and doctors working together can make cardiac care faster, more accurate, and more accessible.

This thesis was a small but practical step toward smarter healthcare, and I’m excited to see where the field goes next.

Link to the full thesis Dhanani, Ridham (2026). Early Detection of Cardiac Abnormalities using ECG. Bachelor’s thesis. Turku University of Applied Sciences. Available in Theseus: https://urn.fi/URN:NBN:fi:amk-2026052215445