Biomedical Embedded AI

PPG–HRV Cognitive Load

Traces timing integrity from earlobe PPG acquisition through IBI/HRV feature construction and CNN+Transformer classification.

Paper and code results kept separate

⌘ Repository ↗ Project Page

Problem

The project carries PPG sample timing through IBI/HRV feature construction and cognitive-task classification.

Architecture

Earlobe PPG sensor/AFE → STM32 acquisition → IBI/HRV preprocessing → CNN+Transformer evaluation → comparison of fold results and paper metrics under separate conditions.

Hardware and Software

PPG AFE
STM32
Python
HRV
CNN
Transformer

Data Flow

PPG sample timing is preserved through peak/IBI calculation before HRV features and sequences enter the classifier.

Methodology

Preprocessed PPG acquisition timing and peak-derived IBIs
Constructed time- and frequency-domain HRV features
Evaluated CNN+Transformer models and compared results by evaluation condition

Results

Paper-reported average AUC of 0.85
Paper-reported F1-score of 0.82
End-to-end PPG/STM32/HRV/CNN research pipeline

01Problem

The project carries PPG sample timing through IBI/HRV feature construction and cognitive-task classification.

02Architecture

Earlobe PPG sensor/AFE → STM32 acquisition → IBI/HRV preprocessing → CNN+Transformer evaluation → comparison of fold resul...

03Method

Preprocessed PPG acquisition timing and peak-derived IBIs

04Result

Paper-reported average AUC of 0.85

A compact problem-to-result map generated from the project data used on this page.

Architecture and result views

Research architecture from PPG acquisition through HRV preprocessing and CNN evaluation — PPG acquisition, IBI/HRV preprocessing, and model-evaluation architecture

What I worked on

IBI/HRV preprocessing and feature construction
CNN+Transformer evaluation pipeline
Comparison of paper metrics and code-archive metrics under their respective evaluation conditions

Code and results

PPG AFE, STM32 firmware, and Python source
Fold-result CSV files and a public paper PDF
Group-level aggregate plots

Next steps

The paper-summary AUC/F1 and the code archive's five-fold metrics use different evaluation conditions. Next, I plan to rerun them with the same dataset split and metric definitions for a direct comparison.

Project resources

Repository Project Pages

I derived IBI and HRV features from PPG peaks and compared CNN and Transformer models. Metrics from the paper table and saved code runs use different evaluation conditions, so I list them separately.