⚡ electrical engineering

Decoding Parkinsonian Tremor: An Explainable Framework Integrating Multi-Revolution Spatial and Spectral Dynamics of Spiral Drawings

This study presents an interpretable, low-cost framework for Parkinson's disease screening that transforms spiral drawings into radial signals to extract and fuse spatial and spectral features, achieving high detection accuracy through a Random Forest classifier while identifying the RMS radial derivative of the outer revolution as a critical biomarker.

Original authors: Tharaka Wijethunge, Maheshi Dissanayake, Sajitha Weerasinghe

Published 2026-07-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tharaka Wijethunge, Maheshi Dissanayake, Sajitha Weerasinghe

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints or footprints, you are looking for the invisible "tremors" in a person's handwriting. This is the world of medical technology, where scientists use computers to help doctors spot diseases like Parkinson's. Parkinson's is a condition that makes the body shake and move slowly, often starting with a shaky hand. Doctors have long known that people with this condition draw spirals differently than healthy people; their lines wobble, overlap, or get messy. But until recently, spotting these tiny differences relied on a doctor's eye, which can be subjective and sometimes miss the subtle clues. The big question in this field is: Can we build a computer program that acts like a super-observant detective, spotting these microscopic wobbles automatically, without needing a massive army of computers or a "black box" that no one understands?

This is exactly what the researchers in this paper set out to do. They didn't just throw a giant, complicated AI model at a pile of pictures and hope for the best. Instead, they built a clever, step-by-step framework that breaks a spiral drawing down into its simplest parts, like taking apart a clock to see how the gears move. They turned the 2D picture of a spiral into a one-dimensional "story" of lines, measuring how far the pen is from the center at every tiny angle. Then, they used math to listen to the "rhythm" of the drawing, looking for the specific high-frequency jitters that signal Parkinson's. Their main finding is that a specific type of computer model called a Random Forest, fed with these carefully measured "spatial" and "spectral" clues, can distinguish between healthy people and those with Parkinson's with impressive accuracy (an AUC of 0.832). Crucially, they found that the most important clue wasn't just how big the spiral was, but the "roughness" or "jerkiness" of the outer edge of the drawing.

The paper also explicitly argues against the idea that the latest, most complex deep learning models (like the ones that recognize cats in photos) are the best tool for this specific job. When the researchers tried to train a standard "Vanilla CNN" (a basic deep learning image recognizer) on their spiral pictures, it failed completely, performing no better than flipping a coin (50% accuracy). This suggests that for small, specific medical tasks like this, throwing a giant AI at the problem doesn't work; you need a smaller, smarter, and more explainable approach.

So, how does their method actually work? Imagine the spiral drawing as a race track. The researchers start by finding the exact center of the track. Then, they shoot invisible laser beams (rays) out from the center in all directions, like the spokes of a wheel. Where each laser beam hits the ink line, they measure the distance. They do this for the first, second, and third loops of the spiral, turning the squiggly picture into three simple lines of numbers. These numbers tell a story: "At this angle, the pen was here; at that angle, it was there."

Next, they look at these lines of numbers in two different ways. First, they look at the "shape" (spatial features). They check how much the line wiggles up and down. Is the line smooth like a calm river, or is it jagged like a mountain range? They found that the "RMS radial derivative" of the outer loop is the superstar clue. In plain English, this measures how fast and how wildly the pen jumps around on the very last part of the spiral. If the pen is jittering like a nervous bird, that's a strong sign of Parkinson's.

Second, they look at the "music" (spectral features). They take those wiggly lines of numbers and run them through a mathematical tool called a Fast Fourier Transform (FFT). Think of this like taking a song and separating it into its bass, mid-range, and treble notes. Healthy handwriting is like a smooth, steady drumbeat. Parkinsonian handwriting, however, has a lot of high-pitched, rapid "static" or tremors mixed in. The researchers found that the amount of this high-frequency "static" is a key indicator of the disease.

The team tested their idea on 116 spiral drawings, a mix of pictures taken in a hospital in Sri Lanka and pictures from public databases. They trained several different computer models to act as the detective. While some models were okay, the Random Forest model was the champion. It correctly identified the patients with Parkinson's much better than the others. Even more importantly, because they built the system using these clear, measurable rules (like "how much does the line wiggle?"), doctors can actually understand why the computer made its decision. It's not a magic black box; it's a transparent tool that says, "I think this is Parkinson's because the outer loop of the spiral is too bumpy and has too much high-frequency jitter."

The researchers are careful to say that this isn't a magic cure-all yet. They note that their study was a "proof of concept" with a relatively small number of patients. They also point out that their model is designed to tell the difference between Parkinson's and a healthy person, but it hasn't been tested yet to see if it can tell the difference between Parkinson's and other conditions that also cause shaking, like Essential Tremor. They also didn't test how the model performs as the disease gets worse over time. However, their work suggests that by combining simple geometry with rhythm analysis, we can create a low-cost, easy-to-use tool that helps doctors catch Parkinson's earlier, even in places without fancy equipment. It's a reminder that sometimes, the best way to solve a complex medical mystery isn't with a bigger computer, but with a smarter way of looking at the clues.

Technical Summary: Decoding Parkinsonian Tremor via Explainable Spiral Analysis

Problem Statement

Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by motor impairments such as tremor, rigidity, and bradykinesia. Current clinical assessment relies heavily on semi-quantitative scales like the MDS-UPDRS, which are susceptible to inter-rater variability and subjective interpretation. Furthermore, misdiagnosis rates can reach 20%, particularly among non-experts. While digitized Archimedes spiral drawings offer a low-cost, non-invasive proxy for detecting motor dysfunction, existing automated approaches often suffer from significant limitations. Deep learning models (e.g., CNNs) require large datasets, extensive computational resources, and specialized hardware, often functioning as "black boxes" that lack the interpretability required for clinical trust. Conversely, traditional handcrafted feature methods often fail to capture physiological tremor signatures in the frequency domain or lack robustness across diverse acquisition environments. There is a critical need for an objective, explainable, and resource-efficient framework capable of operating on small-scale datasets to support early PD screening.

Methodology

The study proposes an explainable framework that transforms 2D spiral images into 1D revolution signals to extract spatial and spectral dynamics. The methodology proceeds through the following stages:

1. Data Acquisition and Preprocessing

The dataset comprises 116 high-resolution spiral images, integrating 30 clinically collected samples from a teaching hospital in Sri Lanka (diagnosed by a single neurology team to minimize inter-observer variance) with publicly available benchmark datasets. Images undergo a standardized preprocessing pipeline: grayscale conversion, noise reduction (median/bilateral filtering), adaptive thresholding for binarization, morphological operations to bridge discontinuities, and resizing to uniform dimensions.

2. Geometric Center Detection and Normalization

A novel center detection strategy based on circular statistics is employed. Unlike centroid-based methods, this approach identifies the center ( $\hat{c}$ ) that maximizes the circular variance of the angular distribution of foreground pixels, ensuring robustness against non-uniform stroke distributions. Once the center is identified, coordinates are transformed relative to $\hat{c}$ and normalized by the maximum radius ( $R_{max}$ ) to map the spiral to a unit circle, mitigating scale variations.

3. Radial Sampling and Signal Extraction

The 2D spiral is converted into three distinct 1D "revolution" signals ( $r_1, r_2, r_3$ ) using a systematic ray-sampling technique. Rays are cast from the center at 5-degree intervals. The algorithm identifies the first three valid intersections with the ink trace, recording the mean radial distance for each revolution. This process captures the innermost, middle, and outer turns of the spiral. Missing data due to intermittent pen contact is handled via circular linear interpolation and median filtering to maintain signal continuity for spectral analysis.

4. Feature Engineering (Spatial and Spectral)

A total of 20 features are extracted, categorized into spatial and frequency-domain descriptors:

Spatial Features: These include kinematic roughness (e.g., RMS radial derivative of the outer revolution, $r3\_deriv\_rms$ ), zero-crossing rates, robust roughness (Median Absolute Deviation), inter-revolution spacing statistics (mean, standard deviation, coefficient of variation), and overlap metrics (crossing rates).
Spectral Features: Derived via Fast Fourier Transform (FFT) of the revolution signals, features are categorized into three harmonic bands: Low (2–4), Mid (5–8), and High (9–12). Metrics include band energies, high-to-low spectral ratios, dominant amplitude, and spectral magnitude decay rates.

5. Classification and Validation

The extracted features are used to train and evaluate six machine learning classifiers: Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GB), XGBoost, LightGBM, and a Vanilla CNN. The study employs a stratified 5-fold cross-validation protocol with class weighting to address imbalance. Hyperparameters are optimized via grid search using ROC-AUC as the primary objective.

Key Results

Classifier Performance: The Random Forest (RF) classifier demonstrated superior performance, achieving an accuracy of 0.767 and a ROC-AUC of 0.832. This significantly outperformed other models, including SVM (AUC 0.729), LightGBM (AUC 0.831), and the Vanilla CNN, which failed to transcend random chance (AUC 0.50) on the small dataset.
Feature Importance: Stability analysis identified the RMS radial derivative of the outer revolution ( $r3\_deriv\_rms$ ) as the most critical biomarker. Other high-importance features included dominant amplitude ( $r3\_dom\_amp$ ), standard deviation of the outer revolution ( $r3\_std$ ), and spectral decay ( $r1\_spec\_decay$ ).
Robustness: The proposed center detection method outperformed centroid and distance-transform baselines in feature extraction completeness (mean angular coverage of 0.791 vs. 0.696). Sensitivity analysis confirmed that the frequency-domain representation is stable across different harmonic band definitions.
Explainability: The framework successfully identified that the model's decisions are driven by interpretable biomechanical and frequency-based features (e.g., micro-tremors and spacing variability) rather than opaque patterns.

Significance and Claims

The paper claims to provide a robust, low-cost, "white-box" screening tool for Parkinson's disease that addresses the limitations of deep learning in data-scarce environments. Key contributions include:

Interpretability: By explicitly engineering features based on radial fluctuations and spectral dynamics, the model offers transparent decision-making that aligns with clinical understanding of motor dysfunction, fostering clinician trust.
Generalizability: The integration of heterogeneous data sources (clinical and public) demonstrates that the framework can learn features invariant to hardware and environmental differences, reducing the risk of cohort-specific overfitting.
Clinical Utility: The approach serves as a practical alternative to resource-intensive deep learning, suitable for deployment in routine clinical workflows, telemedicine, and home-based monitoring. It enables the quantification of motor irregularities (such as tremor-induced fluctuations) that are often missed in visual assessments.
Validation of Biomarkers: The study empirically validates that specific spatial and spectral indices (particularly outer revolution roughness and high-frequency energy) are effective discriminators between PD patients and healthy controls.

The authors position this work not as a standalone diagnostic replacement but as a clinical decision-support tool (CDST) capable of facilitating early intervention and longitudinal tracking, particularly in resource-constrained settings.