Work
06 DSP & Neuroscience 2026

EEG Working-Memory Biomarker

Reading working memory out of the brain signal itself.

Final-year project · 4 members
DSPFFTPythonSVMCNNPyTorch
Signal path from raw epochs to a single per-person discriminability score, and the models that test it.
Signal path from raw epochs to a single per-person discriminability score, and the models that test it.
Grand-average response at Pz. Targets (red) evoke a larger P300 than non-targets (blue); in hearing loss at 2-back the two almost overlap.
Grand-average response at Pz. Targets (red) evoke a larger P300 than non-targets (blue); in hearing loss at 2-back the two almost overlap.
The index per recording: on theta, alpha and beta (left), and on the delta band and the evoked response (exploratory).
The index per recording: on theta, alpha and beta (left), and on the delta band and the evoked response (exploratory).
Where the CNN looks: Grad-CAM relevance per channel, highest over the back of the head.
Where the CNN looks: Grad-CAM relevance per channel, highest over the back of the head.
AUC on each of the 43 held-out people for the best model of each type. Chance is 0.5.
AUC on each of the 43 held-out people for the best model of each type. Chance is 0.5.
Pooled ROC curves, leave-one-person-out.
Pooled ROC curves, leave-one-person-out.
People
43
Epochs
22,065
CNN AUC
0.616
Channels
32 @ 500 Hz
The finding

Theta, alpha and beta barely tell a memory match from a non-match. The difference sits below 4 Hz, in the slow P300 response, and the CNN found it there on its own.

Overview

A twelve-stage Python pipeline that tests whether working memory can be read from EEG. The data: 32-channel EEG at 500 Hz from 43 older adults, 23 with age-related hearing loss and 20 with normal hearing, recorded during 1-back and 2-back tasks in an unpublished PhD study at MAHE, Manipal. The data stay private; the code and group-level results are public.

Problem

Working memory is one of the first things to slip with age, and paper tests miss the early changes. The question: can a person's EEG alone show how distinctly their brain answers a memory match?

System

In an n-back task the brain should treat a target (a memory match) differently from a non-target. 384 FFT band features per epoch feed a Neural Discriminability Index: Fisher's ratio between one person's target and non-target epochs, at equal sample sizes and corrected for chance. The same features train SVM and KNN classifiers, and 32 × 74 spectrograms train a 2D-CNN, all tested on people the models never saw.

spectrogramsEpochs32 ch, 500 HzQuality control±150 µVFFT features384 per epochNDIFisher's ratioRankingMI, χ², RFESVM, KNNLeave one person out2D-CNNGrad-CAM

How it works

  1. Export: 22,141 epochs read from the Brainstorm protocol, channels put in one order by name, and checked for gaps, flat channels and counts. Quality control dropped 33 epochs and one recording.
  2. Features: a Hann-windowed FFT of each 800 ms epoch gives band energy, mean power, spread and variance in theta, alpha and beta on all 32 channels, 384 features, log-transformed. The same FFT gives a 32 × 74 spectrogram from 1 to 30 Hz.
  3. The index: Fisher's ratio between target and non-target epochs, on 33 epochs per class drawn 50 times, minus its average over 200 label shuffles.
  4. Channel ranking: mutual information, chi-square and recursive feature elimination, with scalp maps.
  5. Classifiers: linear and RBF SVM and KNN behind scaling and SMOTE, tuned by grid search, over 43 leave-one-person-out folds. Then a 2D-CNN in PyTorch, explained with Grad-CAM.

Software

  • A small package (I/O, features, the index, selection, models, the CNN, statistics) and one script per stage, each rerunnable on its own.
  • Every analysis choice is fixed in one config file before any result; anything added after seeing results is labelled exploratory.
  • No leakage: scaling, SMOTE, feature selection, tuning and early stopping only ever see the training people, and a person's two recordings stay together.
  • Permutation tests, Mann-Whitney and Wilcoxon tests, mixed models, bootstrap intervals and FDR correction.
  • 12 unit tests, and the reports are built straight from the result files.

My part

Team of four under Prof. Shrishail Bhat. I own the signal-processing path, from raw epochs to the feature set.

Challenges

A biased index

Fisher's ratio is above zero even on pure noise, by about 1/n. Every recording is scored on the same 33 epochs per class and corrected against shuffled labels. SMOTE stays out of it: synthetic epochs shrink the variance and inflate the score.

A CNN that gave up

At the specified learning rate of 0.001, several folds collapsed to a constant output (loss ln 2, AUC 0.5). The training curves showed it; 0.0001 fixed it.

Short epochs

800 ms gives 1.25 Hz resolution, about three independent bins across theta. Zero-padding smooths the spectrum but adds no information, which limits what theta can show here.

Result

On 22,065 clean epochs from 43 older adults, the band-power index is above chance across recordings (p < 0.001) but small, and it doesn't separate hearing loss. In an exploratory follow-up, the same index on the evoked response is much lower in hearing loss at 2-back (p = 0.002, a large effect) and falls with memory load, matching the source study. For target against non-target in unseen people, the CNN reached AUC 0.616 against 0.569 for the best SVM (chance 0.5), and Grad-CAM shows it relies on 1-3 Hz power over the back of the head.

What's next

  • Confirm the evoked-response index on new people and against individual memory scores.
  • Re-cut the recordings into longer epochs to study theta and alpha properly, with per-trial reaction times.
  • Use the study's 6-month follow-up to see whether the index tracks change over time.
All work