Embedded Biomedical AI · End-to-End Case Study

Sense the pulse.
Classify cognitive load.

A traceable engineering pipeline connecting an earlobe PPG analog front end, STM32F411 real-time processing, IBI/HRV feature generation, and CNN–Transformer evaluation.

Explore the design Five-fold CSV Research paper

5-foldStored code-archive evaluation

0.9988Mean ROC-AUC

0.9775Mean F1

STM32F411Acquisition and DSP MCU

01 · System

From sensor to decision

Every boundary that converts an analog pulse into a digital feature is documented. Circuit, firmware, Python pipeline, and stored results link directly to repository evidence.

Overall system architecture from PPG sensor through STM32 and HRV features to CNN Transformer — PPG analog front end → STM32 peak detection → UART IBI stream → HRV features → cognitive-load classifier

Hardware

Earlobe PPG

The analog amplification/filtering path and STM32 ADC connection are supported by circuit and system-architecture evidence.

Firmware

Real-time peak detection

Moving average, IIR band-pass, adaptive thresholding, and a derivative FSM produce peak and IBI events for UART transmission.

Time/frequency HRV

Windowed features are standardized; CNN local patterns and Transformer long-range relationships support High/Low classification.

02 · Signal Pipeline

Inspectable processing stages

The firmware path from ADC samples to heartbeat intervals is separated from the data path that turns those intervals into model-ready HRV tensors.

STM32 PPG signal-processing pipeline — STM32 firmware signal path

Time and frequency HRV feature formulas — HRV definitions in the Python pipeline

Generic N-back task interface without participant data

03 · Model

CNN + Transformer

The CNN compresses local patterns in HRV features; patch embedding and self-attention model dependencies across windows. All headline metrics on this page are calculated from the stored CSV.

04 · Evidence

Results with boundaries

The code archive's five-fold outputs and the paper summary use different evaluation conditions and are therefore reported separately.

AUC accuracy and F1 across five folds — Five-fold archive summary

High Low holdout confusion matrix — Stored holdout confusion matrix

Source audit · Aggregate only

Additional reviewed aggregate evidence

Only group- and fold-level results from the supplied research archive are public. Raw participant PPG, identifiable subject screens, and individual N-back captures remain withheld.

Aggregate RMSSD SDNN HF and LF HF distributions by performance group — Group-level HRV distributions

Aggregate mean HRV feature heat map by performance group — Group-level mean HRV features

Aggregate real versus synthetic HRV feature distribution comparison — Real-versus-synthetic distribution audit

Synthetic low-performance subgroup HRV time and frequency features — Synthetic subgroup feature profile

Aggregate model training accuracy and loss history — Training accuracy and loss

Aggregate K-fold low and high performance confusion matrix — K-fold aggregate confusion matrix

Evidence	ROC-AUC	Accuracy	F1	Interpretation
Code archive, five-fold mean	0.9988	0.9814	0.9775	Arithmetic mean of the five stored fold rows
Paper summary	0.85	—	0.82	Separately reported paper conditions; not directly compared