Applied AI · Published 2025

Low-cost single-lead ECG for vascular-age prediction

Co-authored a peer-reviewed Springer study; contributed data collection, the signal and feature pipeline, and model training and evaluation.

Read the paper

The problem

What was needed, and what made it hard.

The study asks whether a low-cost single-lead ECG module can predict vascular age, and whether it reveals smoking-induced changes in the ECG. The cohort was 42 apparently healthy subjects aged 18 to 30, 20 of them light but habitual smokers. That is a small cohort, and one lead carries less information than a twelve-lead ECG. Others built the ECG module, so my work starts at the captured signal. The engineering problem was to get a model that means something out of 42 people.

My role: Co-author with equal contribution, third of seven authors. My part: data collection, the signal and feature pipeline, model training and evaluation. Not the ECG hardware.

The system

One recording, from capture to evaluation. I did not build the capture hardware. I collected data and built everything after it.

Captured signal 1 Preprocessing 2 Segmentation 3 Features 4 Models 5 Evaluation 6 Captured signal 1 Preprocessing 2 Segmentation 3 Features 4 Models 5 Evaluation 6
One recording, from capture to evaluation. I did not build the capture hardware. I collected data and built everything after it.

Select a component to read what it does and how it fails.

  1. Captured signal. Single-lead ECG from a low-cost module that others built. My work starts at the captured signal; every stage after it inherits its noise.
  2. Preprocessing. Cleans the raw signal before anything is measured. A weak cleaning step corrupts every feature after it.
  3. Segmentation. Overlapping 5-second windows with a 1-second stride turn 42 samples into 6,131. The windows overlap, so they are not independent.
  4. Features. 21 features: 13 from the ECG, such as intervals and QRS duration, and 8 demographic or clinical. Every interval definition can hide an error.
  5. Models. Regression baselines, a decision tree, random forest, 1D-CNN and ResNet-18 transfer learning. Random forest was the paper's best model in all three setups.
  6. Evaluation. Three setups: segmented, unsegmented, and unsegmented with transfer learning from a public PPG dataset. All three are reported, including the weak one.

Decisions

What I chose, what I turned down, and what it cost.

1Segment the recordings

I chose
Overlapping 5-second windows with a 1-second stride, which turned 42 subject-level samples into 6,131.
I turned down
Relying on one sample per subject alone, though the paper reports that setup too.
Why
42 rows is too few to train most models on. Windows give the models thousands of rows to learn from.
What it cost
The windows overlap and come from the same 42 people, so they are not 6,131 independent samples. What broke, below, follows from this.

2Transfer learning for the small set

I chose
Pre-train on a public PPG dataset, then fine-tune on this ECG data, for the 42-sample set.
I turned down
Training on the 42 samples alone.
Why
Alone, the random forest scored R² 0.26. With the pre-training it scored R² 0.87 on the same 42 samples.
What it cost
The result leans on a dataset I did not collect, and 42 people are still 42 people.

3Engineered features and classical models first

I chose
21 engineered features feeding classical models, with the random forest as the headline model.
I turned down
A deep network as the headline model, though I implemented a 1D-CNN and ResNet-18 too.
Why
The random forest was the paper's best model in all three setups, on features I could inspect.
What it cost
Every interval and duration comes from my pipeline, so a mistake there would sit silently under every model.

What broke

What I noticed, what caused it, and how I fixed it.

A weak result on the small set

What I saw
Trained on the 42 unsegmented samples alone, the random forest scored R² 0.26 (MSE 3.56).
The cause
42 rows, one per person, gave the model too little to learn from.
The fix
Transfer learning from a public PPG dataset lifted it to R² 0.87 (MSE 0.99) on the same 42 samples. The weak number stays in the paper's table.

A headline number to read with care

What I saw
The segmented result, R² 0.99 on 6,131 samples, is the figure people quote, and the one to read with most care.
The cause
The 6,131 samples are overlapping windows cut from the same 42 people, so they are not 6,131 independent observations.
The fix
Windowing adds rows, not people, so nothing inside this dataset fixes it. I read R² 0.87 on 42 samples, one per person, as the more conservative figure, and I would want more subjects before trusting either.

Outcome

What came of it, and what it was built with.

  • Across 42 subjects and 21 features, the random forest reached MSE 0.07 on the 6,131 segmented samples and MSE 0.99 on the 42 unsegmented samples with transfer learning.
  • I implemented and evaluated ResNet-18 transfer learning as well. The paper prints no accuracy figure for it, so I quote none.

Built with

  • Python
  • Keras
  • scikit-learn
  • NeuroKit2
  • SciPy
  • random forest
  • ResNet-18 transfer learning

The rule I took from this

The line I keep from this project.

Report a result with its sample, or do not report it.

Rule 9 of 10

Keep reading

Next case study · Ownership

An AI voice front desk for dental clinics

Took an AI receptionist for dental clinics from AI-scaffolded code to production, with three voice providers and per-practice tenant isolation.

Also worth reading

Working on something similar?

Tell me what the system is doing, and what it should be doing.

Email me about this build