TinyML-Based ECG Abnormality Detection for Resource-Constrained Embedded Systems

Q: What is the clinical evidence from this study?

Population: Cardiac arrhythmia (ECG abnormality detection) (n=47). Intervention: Quantized INT8 90-8-1 multilayer perceptron (TinyML model) vs. Floating-point 90-8-1 multilayer perceptron reference model. Primary outcome: Binary ECG beat classification F1-score on the held-out test set (27,054 beats).

Q: What does this research mean for the field?

A compact INT8-quantized 90–8–1 multilayer perceptron achieves 86.28% accuracy and a 53.98% F1-score for binary ECG abnormality detection while requiring only 1,500 bytes of model constants and 130 bytes of working memory. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Mclawrence Okoror

doi:10.31224/8569

October 22, 2026Open Access

TinyML-Based ECG Abnormality Detection for Resource-Constrained Embedded Systems

View Full Paper

Ask AI

Bookmark

Share

Why the study?

This study investigated a lightweight neural-network approach for binary ECG beat classification on resource-constrained embedded systems.

Does an integer-quantized (INT8/INT32) compact 90-8-1 multilayer perceptron maintain binary ECG beat classification accuracy and prediction agreement while reducing memory requirements compared to a floating-point reference model in the MIT-BIH Arrhythmia Database?

Key Points

Investigate a lightweight, integer-based multilayer perceptron pipeline for binary ECG beat abnormality classification on resource-constrained embedded microcontrollers.
Extracted 84,708 annotated beats from the MIT-BIH Arrhythmia Database, partitioned into source-aware training (57,654 beats, balanced to 14,064) and test sets (27,054 beats).
Reduced raw ECG input representations from 180 to 90 samples per beat, halving input storage requirements from 720 bytes to 360 bytes.
Trained a compact 90–8–1 multilayer perceptron and quantized it to INT8 weights and activations with INT32 accumulators, verifying execution in C and on an ESP32-S3 simulation target.
The quantized INT8 model achieved 86.28% accuracy and a 53.98% F1-score across 27,054 test beats, matching the 32-bit floating-point baseline (86.27% accuracy and 53.94% F1-score).
The network architecture required 737 trainable parameters, 728 multiply-accumulate operations per inference, 1,500 bytes for stored model constants, and 130 bytes of inference working memory.
The integer C implementation produced zero mismatches across all 27,054 test vectors compared to the Python reference, and 100 evaluated vectors on the ESP32-S3 simulator showed 100% agreement.

Structured PICO

Does an integer-quantized (INT8/INT32) compact 90-8-1 multilayer perceptron maintain binary ECG beat classification accuracy and prediction agreement while reducing memory requirements compared to a floating-point reference model in the MIT-BIH Arrhythmia Database?

P

Population

47 subjects contributing 48 half-hour ambulatory ECG recordings (84,708 annotated beats: 75,033 normal and 9,675 selected abnormal beats) from the MIT-BIH Arrhythmia Database.

I

Intervention

Compact 90–8–1 multilayer perceptron (MLP) with reduced 90-sample ECG beat input representation, converted to an integer-style inference pipeline using INT8 weights and activations with INT32 accumulators, deployed in C and on an ESP32-S3 target in the Wokwi simulation environment.

C

Comparator

32-bit floating-point Python reference implementation of the 90–8–1 MLP, alongside baseline machine learning models (logistic regression, random forest, decision tree) and alternative MLP architectures (90–16–8–1, 90–8–4–1, 90–4–1, 90–2–1).

O

Outcome

Binary ECG beat classification performance (accuracy, precision, recall, and F1-score) and cross-implementation prediction agreement on 27,054 held-out test vectors.surrogate

Main Result

Absolute Event Rate: 53.98% vs 53.94%

Demonstrates an end-to-end TinyML workflow that compresses a binary ECG beat classifier into 1,500 bytes of model constants and 130 bytes of working memory using integer arithmetic with zero prediction loss compared to floating-point inference.

Limitations

Evaluation was limited to the MIT-BIH Arrhythmia Database and may not generalize to other datasets, patient populations, recording conditions, or hardware.
Binary classification only distinguished normal beats (N) from selected abnormal annotations (A and V) rather than multi-class arrhythmia classification.
ESP32-S3 deployment and timing were evaluated in the Wokwi simulation environment rather than on physical microcontroller hardware.
The same fixed test partition was reused across input-reduction, model-comparison, and architecture-selection experiments rather than using an untouched final holdout.
Source-aware splitting reduced beat-level leakage but was not a fully stratified patient-wise validation protocol.
Precision remained modest (40.12%) under natural class imbalance, indicating a substantial number of false-positive detections.
Reported memory figures (1,500 bytes constants and 130 bytes working memory) exclude total firmware flash and overall microcontroller memory usage.
No clinical validation was performed; the system is a research prototype rather than a diagnostic device.
Evaluation was based solely on the MIT-BIH Arrhythmia Database and may not generalize to other datasets, patient populations, or hardware.
Binary formulation only distinguishes normal beats from selected abnormal annotations (A and V) rather than multi-class arrhythmia classification.
ESP32-S3 performance and timing were evaluated in the Wokwi simulator rather than on physical hardware.
The same fixed test partition was reused across input-reduction, model-comparison, and architecture-selection experiments rather than using an untouched holdout set.
Precision remained modest (40.12%) under natural class imbalance, and no clinical validation was performed.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

May enable on-device ECG classification in embedded systems; leaves open prospective clinical validation before adoption.

Abstract

This work investigates a lightweight neural-network approach for binary ECG beat classification on resource-constrained embedded systems. Using the MIT-BIH Arrhythmia Database, 84,708 annotated beats were extracted and divided into source-aware training and testing groups, with normal beats classified against selected abnormal annotations. Reducing each beat from 180 to 90 raw ECG samples halved input storage while maintaining comparable performance. A compact 90–8–1 multilayer perceptron was selected, containing 737 trainable parameters and requiring 728 multiply-accumulate operations per inference. The floating-point model achieved 86.27% accuracy and 53.94% F1-score, while an integer-style implementation using INT8 weights and activations with INT32 accumulators achieved 86.28% accuracy and 53.98% F1-score. The quantized model required 1,500 bytes for the reported model constants and 130 bytes of inference working memory. The integer implementation reproduced the Python reference predictions for all 27,054 test vectors with zero mismatches. The model was subsequently deployed to an ESP32-S3 in Wokwi, where 100 selected test vectors also produced complete agreement with the reference implementation. These results demonstrate an end-to-end approach for implementing compact ECG classification models using integer-based inference on resource-constrained embedded platforms.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Mclawrence Okoror (2026) studied Cardiac arrhythmia (ECG abnormality detection) (n=47). Quantized INT8 90-8-1 multilayer perceptron (TinyML model) vs. Floating-point 90-8-1 multilayer perceptron reference model was evaluated on Binary ECG beat classification F1-score on the held-out test set (27,054 beats).

TinyML-Based ECG Abnormality Detection for Resource-Constrained Embedded Systems

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

TinyML-Based ECG Abnormality Detection for Resource-Constrained Embedded Systems

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading