electuni.
ABNORMAL

AI-powered ECG screening that runs on the device

A portable, low-cost ECG unit that captures a ten-second reading, screens it for abnormal rhythm on board in 0.13 seconds, and shows the result on screen. No internet connection, no subscription, no data leaving the device.

Built by Electuni with the CSE Discipline, Khulna University. The classifier, CAT-Net, is accepted for publication at IEEE SPICSCON 2026.

Talk to us See the evidence
Product visualisation of the current enclosure and lead set.
91%Macro-AUC across rhythm classes, PTB-XL held-out fold
0.13 sClassification time on the device
10 sRecording captured per screening
242 KBModel size in flash on an ESP32-S3

These figures come from the held-out evaluation described under Evidence below. Accuracy on the same fold is 75.95% and macro-F1 is 0.70. Macro-AUC is the headline figure because it measures ranking across every rhythm class rather than performance at a single threshold.

How it works

Three steps, all of them on the device

  1. Record

    Two leads are placed on the chest and the unit captures a ten-second reading through a TI ADS1293 analog front end.

  2. Screen

    CAT-Net, a hybrid 1D CNN and Transformer model quantised to INT8, classifies the reading in 0.13 seconds on the microcontroller itself.

  3. Show the result

    The screen indicates whether the rhythm looks normal or warrants a clinician’s review, so a health worker knows whether to refer.

The device

Built to be handled, not installed

Battery powered, no mains connection, no host computer. The enclosure holds the analog front end, the microcontroller, an OLED readout and a single button.

Enclosure, readout and lead sockets.

Evidence

Measured, then tested on people

CAT-Net was trained and evaluated on PTB-XL, a public database of clinical 12-lead ECG recordings, using a patient-stratified held-out fold so that no patient appears in both the training and test data.

Macro-AUC
0.91
Macro-F1
0.70
Accuracy
75.95%
Inference time
0.13 s, measured on the ESP32-S3
Dataset
PTB-XL, patient-stratified held-out fold
Publication
Accepted at IEEE SPICSCON 2026
Clinical testing in a hospital setting.
Field testing in a rural setting.
A clinician attaching electrode leads to a patient during a hospital test of the device
Prototype under test in a hospital setting alongside standard monitoring equipment.

This is a research prototype. It is not a registered medical device and it does not produce a diagnosis. Clinician review and referral remain essential.

Inside

What the hardware is

Analog front end
TI ADS1293, two-lead acquisition with digital filtering
Compute
ESP32-S3, INT8 inference on device
Model
CAT-Net, hybrid 1D CNN and Transformer
Footprint
242 KB model in flash, 107 KB tensor arena
Interface
On-board OLED, single button, BLE telemetry
Power
Battery, no mains connection required
Board
Custom four-layer PCB, IPC-compliant design

Where we are

What we are looking for

  • Seed funding to build a twenty-unit batch for clinical piloting
  • A cardiology partner for prospective validation against a reference ECG
  • Guidance on the DGDA medical device registration pathway
  • Pilot sites in rural and upazila health facilities
Email Niloy +880 1855 966357