TypeState

Quantifying Cognitive Load through Keystroke Dynamics

---

Abstract

TypeState is a research project investigating the correlation between keystroke dynamics and cognitive load. By capturing and analyzing typing patterns in real-time, we demonstrate that adaptive user interfaces can detect and respond to user stress states before they escalate.

This repository contains the complete research pipeline: data collection interface, feature engineering scripts, trained LSTM model, and real-time inference server.

Download Full Research Report (PDF)

---

Research Overview

Metric	Value	Description
Stress Threshold	0.6	Classification boundary for high cognitive load
Temporal Resolution	15ms	Keystroke timing precision
Sequence Window	20	Keystrokes analyzed per inference
Privacy Model	Local	No personally identifiable information stored

Project Milestones

Phase	Status	Description
Data Collection	Complete	Web-based keystroke capture with dual-mode sessions
Feature Engineering	Complete	Flight time, dwell time, rolling variance extraction
Model Training	Complete	LSTM neural network for real-time classification
Web Integration	Complete	Live stress detection with calming UI intervention
Documentation	Complete	Research report available

---

Methodology

System Architecture

graph LR
    A[User Input] --> B[Keystroke Capture]
    B --> C[Feature Extraction]
    C --> D[LSTM Model]
    D --> E{Score > 0.6}
    E -->|Yes| F[Calming Intervention]
    E -->|No| G[Normal State]
    
    style D fill:#065f46,color:#fff
    style F fill:#10b981,color:#fff

Loading

Feature Set

Feature	Definition	Rationale
Flight Time	Time between consecutive key presses (ms)	Hesitation indicates cognitive load
Flight Time Variance	Rolling variance (window=5)	Rhythm inconsistency correlates with stress
Error Rate	Backspace frequency	Frustration manifests as correction attempts

---

Installation

Prerequisites

• Python 3.9+
• TensorFlow 2.x
• FastAPI

Setup

git clone https://github.com/1mystic/TypeState.git
cd TypeState

# Create virtual environment
python -m venv .venv

# Activate
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate

# Install dependencies
pip install fastapi uvicorn tensorflow pandas numpy joblib scikit-learn

Running the System

1. Start the AI Server

cd typestate_ml
uvicorn server:app --reload

Server available at http://127.0.0.1:8000

2. Launch the Interface

Open site/index.html in browser, or use VS Code Live Server.

---

Repository Structure

TypeState/
├── site/                       # Frontend Web Application
│   ├── index.html              # Landing page
│   ├── collector.html          # Data collection interface
│   ├── about.html              # Project information
│   ├── research.html           # Related literature
│   └── assets/
│       ├── css/style.css       # Styling
│       └── js/collector.js     # Keystroke capture
│
├── typestate_ml/               # AI Backend
│   ├── server.py               # FastAPI inference server
│   ├── typestate_model.h5      # Trained LSTM model
│   └── typestate_scaler.pkl    # Feature normalizer
│
├── data_analysis/              # Research Pipeline
│   ├── feature_extraction.py   # Raw data processing
│   ├── process_data.py         # Dataset preparation
│   ├── visualise.py            # Visualization scripts
│   └── stress_signature.png    # Sample output
│
├── Overview-Report.pdf         # Full Research Report
└── README.md

---

Research Process

Phase 1: Data Collection

Participants completed typing tasks under two conditions:

• Relaxed Mode — Standard typing environment
• Stressed Mode — Time pressure with error-induced frustration

// Captured event structure
{
  "key": "KeyA",
  "type": "down",
  "time": 1702534892123,
  "char": "a"
}

Phase 2: Feature Engineering

# Core feature extraction
df['flight_time'] = df.groupby('session_id')['timestamp'].diff()
df['flight_time_var'] = df['flight_time'].rolling(window=5).var()
df['is_error'] = (df['key'] == 'Backspace').astype(int)

Phase 3: Model Architecture

Input: Sequence of 20 keystroke features
    ↓
LSTM Layer (64 units, return_sequences=True)
    ↓
LSTM Layer (32 units)
    ↓
Dense Layer (16 units, ReLU)
    ↓
Output (1 unit, Sigmoid) → Stress Score [0.0 - 1.0]

Phase 4: Intervention Design

When stress score exceeds 0.6, the system:

• Transitions background to calming sage green
• Displays "High Cognitive Load Detected" notification
• Softens UI elements to reduce visual stress

---

Design System

Token	Value	Application
--primary	#3D6A4B	Primary actions
--surface	#F2F6F3	Background
--highlight	#1A4228	Emphasis text
--accent	#065f46	Interactive elements

---

Related Literature

• Lim et al. (2014) — Statistical correlation between keystroke latency and cognitive load using static classifiers
• Nie et al. (2022) — Random forest approach for stress detection in academic writing sessions

---

Documentation

For complete methodology, experimental design, results, and analysis:

Download Research Report (PDF)

---

License

This is an original work by Atharv Khare. All use, reproduction, or distribution must include proper attribution to the original author.

Released under the MIT License : see LICENSE for details.

© 2025-2026 Atharv Khare. All Rights Reserved.

---

Live Demo  |  Research Report