B.Tech Final Thesis · BML Munjal University · 2024
Can a machine read a child's brainwaves and detect ADHD? We built a CNN and tested seven ML models against 19-channel EEG from 121 children — reaching 98.53% accuracy.
98.53%
CNN accuracy at 3000 epochs
5-block CNN with batch normalisation, LeakyReLU, max-pooling, and dropout layers. Validated with 10-fold cross-validation on 19-channel EEG data.
98.5%
LightGBM — best traditional ML
Gradient boosting on 13 extracted time-domain features with StandardScaler preprocessing. Outperformed 6 other classifiers including XGBoost (97.8%) and Random Forest (81%).
121
Children in the dataset
61 diagnosed with ADHD, 60 controls. Ages 7–12. EEG recorded at 128 Hz across 19 channels during a visual attention task designed specifically for children.
The CNN's advantage came from automatic hierarchical feature extraction — it learned spatial and temporal patterns directly from raw EEG without hand-engineering features, matching what gradient boosting needed 13 manually extracted features to achieve.
ADHD affects millions of children. Diagnosis today still relies on subjective behavioural checklists.
"Early detection is very important for effective intervention — EEG signals offer a great way to objectively classify ADHD based on characteristic brain wave patterns."
Attention Deficit Hyperactivity Disorder is a prevalent neurodevelopmental condition that primarily impacts children and can persist into adulthood. It affects learning, social interaction, and quality of life — yet accurate early diagnosis remains challenging due to its reliance on clinician observation and parent/teacher reports.
Electroencephalography (EEG) captures brain electrical activity non-invasively. Prior research showed ADHD produces distinct brainwave signatures — the goal of this thesis was to automate that detection with machine learning.
121
Children total
61/60
ADHD / Control
7–12
Age range (years)
19
EEG channels
EEG Setup
Sampling rate: 128 Hz
19 channels: Fz, Cz, Pz, C3, T3, C4, T4, Fp1, Fp2, F3, F4, F7, F8, P3, P4, T5, T6, O1, O2
Standard 10-20 electrode placement system
ADHD diagnoses by psychiatrists via DSM-5 criteria
Recording Task
Children viewed images of cartoon characters and counted them (5–16 per image). After each response, the image was immediately replaced — maintaining constant visual attention stimulation throughout the recording.
Recording duration varied per child based on performance. Gender-balanced: 98 boys (48 ADHD, 50 control), 23 girls (13 ADHD, 10 control).
01 — Signal Preprocessing
Raw EEG files loaded from ADHD and control folders. Unnecessary columns stripped, channel names assigned to 10-20 standard placement. MNE library used to construct RawArray objects at 128 Hz. Continuous recordings segmented into 6-second epochs for analysis.
02 — Feature Extraction
13 time-domain statistical features computed per channel per segment: mean, standard deviation, variance, peak-to-peak, min, max, arg-min, arg-max, RMS, absolute difference, skewness, kurtosis. Each EEG segment became a 1D feature vector for the ML models.
03 — Traditional ML Evaluation
7 classifiers evaluated with 6 preprocessing scalers each: MinMaxScaler, StandardScaler, MaxAbsScaler, PowerTransformer, QuantileTransformer, Normalizer. 80/20 train-test split with 10-fold cross-validation. Models: Extra Trees, Random Forest, XGBoost, LightGBM, Gradient Boosting, AdaBoost, Logistic Regression.
04 — CNN Model
5-block CNN trained directly on raw EEG (no manual features needed). Each block: Conv1D(5 filters, kernel=3) → BatchNorm → LeakyReLU → MaxPool1D(pool=2) → Dropout(0.4). Followed by Dense(48) → Dense(32) → Dense(1, sigmoid). Adam optimizer (lr=0.0003), binary cross-entropy loss. 10-fold CV.
Accuracy by classifier
CNN vs Best ML
CNN (3000 epochs)
98.53%
LightGBM + StandardScaler
98.5%
XGBoost
97.8%
CNN (50 epochs only)
78.76%
CNN Epochs vs Accuracy
50 epochs
78.76%
200 epochs
~88%
500 epochs
~93%
1000 epochs
~96%
3000 epochs
98.53%
A 5-block convolutional architecture designed to extract hierarchical temporal and spatial features from raw EEG signals without manual feature engineering.
Input
19-ch EEG segments
×5 Conv Block
Conv1D → BatchNorm LeakyReLU → MaxPool Dropout(0.4)
Flatten
3D → 1D vector
Dense
48 units 32 units ReLU + Dropout
Output
1 unit Sigmoid ADHD / Control
Training Configuration
Optimizer
Adam
Learning rate
0.0003
Loss
Binary cross-entropy
Validation
10-fold CV
This thesis was done before I pivoted to UX — but the research mindset it built has been inseparable from my design work ever since. Running a full ML pipeline from raw signal data to validated model taught me what it actually means to work with evidence.
The most interesting tension was CNN vs gradient boosting. Both hit ~98.5%, but through completely different routes — one through hand-crafted features, the other through automatic representation learning. Watching a model learn better features than I could engineer manually was a turning point.
Design connection
ADHD affects how children interact with any interface — attention span, impulsivity, response latency. Having studied it at a signal level makes me think differently about accessible and neurodiverse-inclusive design. The quantitative rigor here feeds directly into how I frame research questions in UX studies.
Full ML pipeline: preprocessing → feature extraction → model selection
EEG signal analysis with MNE library
Cross-validation methodology and avoiding data leakage
Interpreting model performance beyond accuracy (overfitting, generalisation)
Full thesis document
EEG-based ADHD Classification using Machine Learning
BML Munjal University · B.Tech Practice School III · July 2024 · 25 pages
Open to UX/UI internships — Finland & Europe
Interested in working together?