71st Florida SSEF · Intelligent Machines, Robotics & Systems Software · Lakeland, FL

Quantum-Inspired
Machine Learning
for Brain-Computer
Interfaces

0%
QSVM Accuracy
0%
Motor Recall
+0pp
vs Classical SVM
0
Subjects · PhysioNet
Scroll

01 · The Problem

BCIs are failing
the patients who need them

Every 40 seconds someone has a stroke. Many lose all voluntary movement. Brain-computer interfaces exist to give those people their voice back — but the decoding layer keeps failing.

FAILURE MODE 01

Low Signal,
High Noise

Scalp EEG captures millions of neurons firing through bone and tissue simultaneously. The signal-to-noise ratio is fundamentally poor — non-stationary, noisy, and highly variable across recording sessions and individuals.

FAILURE MODE 02

Inter-Subject
Variability

No two brains produce the same EEG signature. Classical models trained on one person fail on another. Fixed feature spaces cannot adapt to the biological diversity of human neural architecture.

FAILURE MODE 03

Missing the
Interference

Current models treat each EEG channel independently. But the brain works like an orchestra — 64 channels interfering constructively and destructively. Classical kernels hear one violin. We needed to hear all 64.

Classical SVMs achieve ~64.6% accuracy on motor imagery classification — insufficient for a patient whose wheelchair depends on every correct command. When a BCI fails, a paralyzed patient cannot move. 91.3% recall isn't a metric. It's a lifeline.


02 · Engineering Methodology

The Born Rule,
applied to neurons

No quantum computer required. We applied interference-based probability mathematics from quantum physics as a kernel function — measuring how EEG signals interact, not just how far apart they are.

01

Dataset Acquisition

PhysioNet EEG Motor Movement/Imagery Dataset. 109 subjects, 654 recordings segmented into approximately 9,500 trials. Imagined left vs. right hand movement. 64 channels at 160 Hz. Fully de-identified public data.

PhysioNet · 109 subjects · ~9,500 trials
02

Signal Preprocessing

Band-pass filtering at 8–30 Hz to isolate mu and beta motor rhythms. ICA artifact removal. 5-fold stratified cross-validation applied uniformly. Identical preprocessing pipeline across all 9 benchmark models.

8–30 Hz · MNE-Python · 5-Fold CV
03

Quantum-Inspired Kernel

Classical kernels ask: how far apart are two signals? Our quantum kernel asks: how much do they interfere? Welch's method extracts PSD across α, β, and μ bands. The Born Rule maps these into a quantum-inspired similarity space that captures non-linear neural interference patterns classical kernels cannot represent.

Born Rule · Amplitude-Phase Encoding
04

9-Model Benchmark

QSVM tested against Classical SVM, CNN1D, CNN2D, LSTM, EEGNet, DeepConvNet, and additional architectures on identical data. Statistical validation via Wilcoxon signed-rank test with Bonferroni correction.

Wilcoxon p = 0.032 · Bonferroni-Corrected

Dataset Profile

SourcePhysioNet
Subjects109
Recordings654
Segmented Trials~9,500
EEG Channels64
Sampling Rate160 Hz
Cross-Validation5-Fold
Significancep = 0.032

Engineering Criteria — All Met

Robustness — Smaller accuracy degradation under 50% data reduction (3.8% vs 10.5%)
Functionality — Quantum kernel encodes EEG amplitude-phase interference confirmed via PCA
Specification — Statistically significant improvement at p = 0.032 (Bonferroni)

03 · Results

Matching deep learning.
No GPU required.

QSVM achieved 73.4% accuracy and 81.4% F1-score — outperforming Classical SVM by 8.8pp while matching GPU-dependent deep learning on standard CPU hardware.

Model Accuracy TEST SET · n=99
QSVM
73.4%
CNN2D
74.1%
EEGNet
72.8%
CNN1D
71.5%
LSTM
69.2%
Classical SVM
64.6%
F1 Score WILCOXON p=0.032
QSVM
81.4%
CNN2D
82.1%
EEGNet
80.6%
CNN1D
79.2%
LSTM
76.8%
Classical SVM
74.1%
Confusion Matrix QSVM vs CLASSICAL SVM
Classical SVM
Pred: Low
Pred: High
Actual Low
37
True Neg
24
False Pos
Actual High
18
False Neg
20
True Pos
Recall ≈ 53% — misses motor commands
QSVM (Proposed)
Pred: Low
Pred: High
Actual Low
52
True Neg
13
False Pos
Actual High
3
False Neg
31
True Pos
Recall = 91.3% — critical FN eliminated
Data Robustness DEGRADATION UNDER REDUCTION
QSVM
Classical SVM
At 50% data: QSVM −3.8% · Classical −10.5%
Multi-Metric Performance 5 DIMENSIONS · QSVM vs BASELINES
QSVM
Classical SVM
CNN2D

04 · Real-World Impact

Why not just
use deep learning?

Several deep learning models scored marginally higher accuracy. Here is why that framing misses the point entirely.

No GPU Required

Deep learning requires expensive GPU hardware. The average rural clinic, field hospital, or home caregiver does not have one. QSVM runs on a standard laptop and matches deep learning accuracy — the difference between technology in a research lab and technology that reaches a patient in rural Florida or rural Kenya.

Standard CPU · Accessible

Works With Scarce Data

Deep learning needs thousands of labeled samples. Getting one ALS patient to produce clean EEG data takes months of clinical work. QSVM achieves 73% accuracy under 500 training samples and degrades far less when data is reduced. When labeled neurological data is scarce — which it almost always is clinically — quantum-inspired kernels are often the only viable path.

Data Efficient · Low-Resource

FDA-Ready Interpretability

The FDA will not approve a black-box model to control a prosthetic limb. Clinicians need to understand why a model made a decision. The quantum kernel matrix is fully interpretable — you can see exactly how similar two neural states are. Deep learning cannot offer that transparency at the required clinical resolution.

Interpretable · Regulatory
15M
People needing reliable BCIs
91.3%
Motor imagery recall rate
3.8%
Accuracy drop at 50% data
0
GPU hardware required

Our contribution is not beating deep learning on a benchmark. It is matching deep learning accuracy while solving the three problems that prevent deep learning from ever leaving the lab and reaching the patients who need it most.


05 · Research Team

Creekside High School,
St. Johns County

71st Florida State Science & Engineering Fair · March 31–April 2, 2026 · Lakeland, Florida

ST
Sanvi Tummala
Lead Researcher
Creekside High School · St. Johns, FL
GD
Gurhans Dhillon
Co-Researcher
Creekside High School · St. Johns, FL
SH
Sachith Hulikere
Co-Researcher
Creekside High School · St. Johns, FL
PhysioNet EEG Dataset Python · Scikit-Learn MNE-Python Google Colab PennyLane · Qiskit 5-Fold Cross-Validation Wilcoxon Signed-Rank Born Rule Kernel Bonferroni Correction Power Spectral Density