Handwriting AI for Tremor and Parkinson’s Detection

ISEF Category: Translational Medical Science

Subcategory: Disease Detection and Diagnosis  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A pencil can reveal more than neatness. Tiny changes in pressure, speed, and stroke shape can hint at movement disorders before a person even notices them. That makes handwriting a powerful signal for early screening. You can test that signal with public data and machine learning.

What Is It?

This project uses handwriting data from a tablet or stylus. Instead of looking only at the finished letters, you study how the pen moves while someone writes. Pressure tells you how hard the stylus presses down. Velocity tells you how fast it moves. Together, they act like a motion fingerprint.

A CNN, or convolutional neural network, is a machine learning model that learns patterns from data. In this case, it can learn stroke patterns linked to essential tremor and early Parkinson's disease. You are not diagnosing anyone. You are testing whether a model can tell the difference between groups in public datasets such as HandPD and DraWritePD. Per-stroke interpretability means you also check which strokes or motion features drive the model's decision, instead of treating it like a black box.

Why This Is a Good Topic

This is a strong science fair topic because it mixes a clear human health problem with data you can actually study from home. You can ask whether handwriting kinematics contain enough signal to separate tremor from Parkinson's, then test that claim with classification and interpretation methods. The project teaches data cleaning, feature engineering, model evaluation, and error analysis. Those are real research skills, not just coding tricks.

Research Questions

• How does adding stroke pressure improve classification of essential tremor versus early Parkinson's disease?
• What is the effect of using velocity features alone versus pressure features alone on model accuracy?
• Does a CNN outperform a simpler machine learning model on handwriting kinematics data?
• To what extent do per-stroke interpretations match known motor symptoms in each dataset?
• Which handwriting tasks produce the strongest separation between essential tremor and Parkinson's?
• How does removing noisy or incomplete strokes change model performance?

Basic Materials

• Laptop or desktop computer with enough memory to handle public datasets.
• Python installed with Jupyter Notebook.
• Free access to the HandPD and DraWritePD datasets.
• Spreadsheet software for tracking sample labels and results.
• Internet access for reading dataset documentation and papers.
• External hard drive or cloud storage for versioned backups.
• Headphones or quiet workspace for focused coding sessions.

Advanced Materials

• High-performance laptop or workstation with a dedicated GPU.
• Python environment with TensorFlow or PyTorch.
• Jupyter Notebook or VS Code for analysis.
• Data visualization tools such as Matplotlib and Seaborn.
• A software package for model interpretation, such as SHAP or Captum.
• Statistical software or Python libraries for significance testing.
• Secure storage for dataset files and experiment logs.

Software & Tools

• Python: Runs your data cleaning, feature extraction, model training, and evaluation scripts.
• Jupyter Notebook: Helps you document each analysis step and keep code, plots, and notes together.
• pandas: Organizes stroke data into tables you can clean and compare.
• scikit-learn: Provides baseline models, metrics, and train-test splitting tools.
• TensorFlow or PyTorch: Lets you build and train a CNN on handwriting kinematics data.
• Matplotlib and Seaborn: Create plots that compare classes, features, and model performance.
• SHAP: Estimates which stroke features push the model toward each prediction.

Experiment Steps

1. Define the exact classification task and decide which dataset labels you will compare first.
2. Inspect the raw stroke data and choose the features you can trust, such as pressure, velocity, or stroke duration.
3. Plan a baseline model so you can compare the CNN against a simpler reference.
4. Design your train, validation, and test split to avoid mixing strokes from the same person across sets.
5. Build an interpretation plan that links model predictions back to individual strokes or motion features.
6. Decide how you will test whether any performance gain is real and not just random variation.

Common Pitfalls

• Mixing handwriting samples from the same person across train and test sets, which inflates accuracy.
• Treating class imbalance as a minor detail, which can make the model favor the larger group.
• Comparing datasets without matching their task format, which can turn a disease signal into a dataset artifact.
• Ignoring missing or noisy stroke segments, which can confuse both the CNN and the interpretation step.
• Reporting only overall accuracy, which can hide poor sensitivity for one condition.

What Makes This Competitive

A competitive project would do more than train a model and report accuracy. You would test strong baselines, control for person-level data leakage, and compare multiple handwriting tasks or feature sets. You would also explain the model with per-stroke analysis, then check whether those explanations make medical sense. That kind of careful evaluation looks much closer to real research.

Project Variations

• Compare handwriting kinematics from different prompt types, such as looping, copying, or free writing, to see which separates the groups best.
• Test whether adding derived motion features, such as acceleration or jerk, improves over raw pressure and velocity alone.
• Repeat the analysis with a simpler model, such as random forest or SVM, to see whether deep learning truly adds value.

Learn More

• PubMed: Search for review articles on handwriting analysis, Parkinson's disease, and essential tremor to understand the medical background.
• NIH National Institute of Neurological Disorders and Stroke: Read disease overviews and symptom summaries for Parkinson's disease and essential tremor.
• HandPD Dataset papers: Search scholarly articles that describe the public handwriting dataset and its collection protocol.
• DraWritePD papers: Search the dataset publication and related papers to learn the task design and available signals.
• MIT OpenCourseWare: Find free materials on machine learning, pattern recognition, and neural networks for model-building basics.
• IEEE Xplore and Google Scholar: Search for peer-reviewed papers on handwriting kinematics and disease classification to compare methods and metrics.