💻 computer science

From Pen Strokes to Sleep States: Detecting Low-Recovery Days Using Sigma-Lognormal Handwriting Features

This study demonstrates that a personalized classification framework using Sigma-Lognormal features from everyday handwriting can effectively detect daily low-recovery states in healthy individuals, as validated by nocturnal cardiac indicators in a 28-day in-the-wild experiment.

Original authors: Chisa Tanaka, Andrew Vargo, Anna Scius-Bertrand, Andreas Fischer, Koichi Kise

Published 2026-03-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chisa Tanaka, Andrew Vargo, Anna Scius-Bertrand, Andreas Fischer, Koichi Kise

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 your handwriting isn't just a way to write a grocery list or sign a document. Imagine it's actually a secret diary written by your nervous system, revealing how well your body recovered from the previous night's sleep.

That is the core discovery of this new study. Here is the breakdown of how they cracked the code, explained in simple terms.

The Big Idea: Your Pen is a Sleep Tracker

Usually, when we think of handwriting analysis, we think of police trying to identify a forger or doctors looking for signs of Parkinson's disease. But this study asked a different question: Can a healthy person's handwriting tell us if they had a "bad sleep" day?

Think of your brain and hand like a musical instrument.

When you are well-rested, your brain is a skilled conductor, and your hand plays the notes (the pen strokes) smoothly and rhythmically.
When you are tired or your body hasn't recovered from stress, the conductor is sluggish. The music (your handwriting) might still look like the same song, but the timing and pressure are slightly off.

The researchers wanted to see if they could hear that "off-beat" rhythm just by looking at the digital ink.

The Experiment: 28 Days of "In-the-Wild" Testing

To test this, they didn't lock people in a lab. They sent 13 university students out into the real world for 28 days.

The Sleep Tracker: Every night, the students wore a high-tech smart ring (like an Oura Ring) on their finger. This ring acted as a truth-teller, measuring exactly how well their bodies recovered by tracking their heart rate and heart rate variability (how much their heart rate changes between beats, a sign of relaxation).
The Handwriting Test: Three times a day (morning, afternoon, night), the students used a special digital tablet to draw shapes (circles, triangles, squares) and write short phrases.
The Magic Lens (Sigma-Lognormal Model): This is the fancy part. Instead of just looking at how fast they wrote, the researchers used a mathematical model called Sigma-Lognormal.
- Analogy: Imagine a car driving down a road. A simple analysis just measures the speedometer. The Sigma-Lognormal model, however, analyzes the engine's vibration, the steering wheel's micro-adjustments, and the suspension's bounce. It breaks down the movement into tiny, invisible components that reveal how the "engine" (the brain) is firing.

The Results: Can We Hear the Fatigue?

They trained a computer (an AI) to look at the handwriting data and guess: "Was this a 'low-recovery' day (bad sleep) or a 'good-recovery' day (good sleep)?"

Here is what they found:

The AI Got It Right: The computer could identify bad sleep days significantly better than random guessing. It was like a detective who could spot a lie just by looking at a person's posture.
Heart Rate was the Key: The handwriting was most sensitive to changes in heart rate. When a person's heart rate was high or their heart rate variability was low (signs of stress/poor recovery), their handwriting showed subtle "glitches" in the motor control.
It Didn't Matter What They Wrote: It didn't matter if the student was drawing a circle or writing a sentence. The fatigue showed up in all movements.
It Didn't Matter When: Whether they wrote it right after waking up or right before bed, the signal was there.

Why This Matters (The "So What?")

Currently, to know if you slept well, you have to wear a gadget on your wrist or finger 24/7. Many people hate this because it's uncomfortable, expensive, or they forget to charge it.

This study suggests a superpower for the future:

No Extra Gadgets: If you already use a tablet for school or work, your device could secretly monitor your health just by watching how you write.
For Kids: Imagine a classroom where the teacher's tablet could gently nudge a student who is chronically sleep-deprived, not by asking them, but by noticing their pen strokes are "tired."
Non-Invasive: It's a way to check your health without sticking sensors to your skin or wearing a watch.

The Catch

The study isn't perfect yet. The "signal" is very subtle. Some people's handwriting changes a lot when they are tired, while others' don't change much. It's like some people get "hangry" (hungry + angry) when they skip a meal, while others just get quiet. The AI needs to learn each person's specific "tired signature."

The Bottom Line

This paper proves that your handwriting is a window into your autonomic nervous system. Even when you feel fine, your pen strokes might be whispering that your body needs more rest. It opens the door to a future where your daily writing tasks double as a silent, invisible health monitor.

1. Problem Statement

While handwriting analysis has traditionally been used for character recognition and the diagnosis of pathological conditions (e.g., ADHD, neurodegenerative diseases), its potential to reflect day-to-day physiological fluctuations in healthy individuals remains unexplored.

The Gap: Existing studies focus on long-term developmental changes or disease states. There is a lack of research on estimating daily variations in physical condition (specifically sleep quality and autonomic recovery) among healthy people using handwriting.
The Motivation: Current sleep monitoring relies on wearable devices (smartwatches, rings), which suffer from high abandonment rates due to cost, the need for continuous wearing, and maintenance. The authors propose a non-invasive, device-independent alternative: estimating sleep recovery states from everyday handwriting activities on tablets.

2. Methodology

A. Data Collection

Participants: 13 university students (after excluding 4 who failed to complete the protocol) from an initial pool of 17.
Duration: 28 days ("in-the-wild" study).
Handwriting Data:
- Device: IoT Paper tablet (Wacom) with a sampling rate of 480 Hz.
- Frequency: 3 times daily (after waking, after lunch, before bedtime).
- Tasks: 5 types per session: drawing a circle, triangle, and square (3 times each), plus writing a positive phrase and a negative phrase.
- Total Samples: 420 samples per participant (28 days × 3 timings × 5 tasks).
Sleep Data (Ground Truth):
- Device: Oura Ring (Gen 3 and Gen 4).
- Metrics: Raw sensor data (optical heart rate) rather than algorithmic sleep stages.
- Target Variables:
  1. Total Sleep Duration: Total time asleep.
  2. Avg HRV: Average Heart Rate Variability (autonomic recovery indicator).
  3. Lowest HR: Minimum heart rate during sleep.
  4. Avg HR: Average heart rate during sleep.

B. Feature Extraction: Sigma-Lognormal Model

The study utilizes the Sigma-Lognormal model, a neuromotor theory describing handwriting velocity profiles as a superposition of lognormal distributions.

Process: The model decomposes the velocity profile of each stroke into parameters representing the motor command from the brain.
Features Extracted: 5 basic features per stroke:
1. $D$ : Amplitude of motor command.
2. $t_0$ : Onset time of motor command.
3. $nblog$: Number of lognormal distributions.
4. $SNR$: Signal-to-noise ratio.
5. $SNR/nblog$: Reconstruction accuracy per distribution.
Aggregation: For each task, the mean and standard deviation of these 5 features are calculated, resulting in 10 features per task.

C. Classification Framework

Task: Binary classification to detect "Low-Recovery Days" (days with poor sleep quality).
Labeling Strategy:
- Personalized Thresholds: To account for inter-individual variability, thresholds are set per user.
- Positive Class Definition:
  - For Total Sleep and HRV: Days below the 25th percentile (Q25).
  - For Heart Rate (Lowest/Avg): Days above the 75th percentile (Q75).
- This results in an imbalanced dataset where ~25% of samples are positive.
Model: Random Forest (RF) classifier, chosen for its ability to handle non-linear interactions and small sample sizes.
Validation: Leave-One-Day-Out Cross-Validation (LODOCV).
- One day's data (15 samples: 3 timings × 5 tasks) is held out for testing; the remaining 27 days are used for training.
- This ensures the model predicts future or unseen daily states based on historical personal data.

3. Key Contributions

Novel Application Domain: First study to estimate daily sleep quality and autonomic recovery in healthy individuals using handwriting features, moving beyond disease classification.
Sigma-Lognormal Application: Demonstrates that fine-grained neuromotor parameters (beyond basic kinematics like speed) capture subtle physiological fluctuations.
Robustness: Proved that the method works across different task types (shapes vs. text) and timings (morning vs. night), suggesting the signal is embedded in general movement dynamics.
Device Independence: Proposes a pathway for sleep monitoring using standard educational tablets, removing the barrier of wearable device adoption.

4. Experimental Results

Performance Metrics

The study used PR-AUC (Precision-Recall Area Under Curve) and Recall@25% due to class imbalance. The baseline for a random classifier is 0.25.

Significance: All four target variables showed PR-AUC scores significantly exceeding the baseline (0.25) after False Discovery Rate (FDR) correction.
Ranking of Performance:
1. Lowest HR: PR-AUC = 0.438 (Best performance).
2. Avg HRV: PR-AUC = 0.391.
3. Avg HR: PR-AUC = 0.365.
4. Total Sleep: PR-AUC = 0.352 (Lowest, but still significant).

Robustness Analysis

Task Effect: Friedman tests revealed no significant difference in performance across the 5 handwriting tasks (circles, triangles, squares, phrases).
Timing Effect: No significant difference was found across the 3 daily timings (post-waking, post-lunch, pre-sleep).
Individual Variability: While the group average was significant, individual performance varied widely (e.g., one user achieved a PR-AUC of 0.729 for Lowest HR, while others were near baseline).

5. Significance and Discussion

Physiological Link: The results support the hypothesis that sleep quality affects the autonomic nervous system, which in turn modulates fine motor control. The Sigma-Lognormal parameters ( $D$ , $t_0$ , etc.) appear sensitive to these subtle autonomic shifts.
Cardiac vs. Sleep Duration: Cardiac indicators (HR, HRV) yielded better classification results than total sleep duration. The authors argue this is because HR/HRV are direct measures of autonomic recovery, whereas sleep duration is a "quantity" metric that does not always correlate with "quality" (e.g., long but poor sleep vs. short but deep sleep).
Feasibility: The lack of dependency on specific tasks or times implies that any natural handwriting activity on a tablet could theoretically be used for monitoring, making it highly practical for educational settings.
Limitations & Future Work:
- Individual Differences: Large variance in performance across users suggests the need for personalized models and further investigation into factors like writing style or lifestyle stability.
- Generalizability: The study was limited to university students. Future work must test diverse age groups and occupations.
- Real-world Application: Validation in naturalistic settings (e.g., actual note-taking in class) rather than controlled shape-drawing tasks is needed.

Conclusion: The paper establishes that subtle, within-person fluctuations in autonomic recovery can be detected from everyday handwriting using the Sigma-Lognormal model, offering a promising, non-invasive avenue for health monitoring.