Inference Runs FreePlans For Every Builder

Does the Focus Performance Score (FPS) in Pulse Focus really measure selective attention and cognitive inhibition while a player resolves color-word conflict, and is that measurement grounded in the brain processes behind attentional control?

AI that models, predicts or adapts to human attention needs a reliable ground truth. Wearable sensors such as HRV and PPG give continuous signals, but they need behavioral labels to mean anything; self-report is retrospective, and momentary check-ins interrupt the behavior they try to measure.

A real-time, trial-by-trial score from a well-validated cognitive task fills that gap. It can label physiological data, test attention classifiers and act directly as a signal for attention-aware interfaces — but only if the score itself is psychometrically sound.

Each Pulse Focus session has two phases. A 2–3 minute guided breathing session at about 6 breaths per minute sets a personal resting HRV baseline. Then the Stroop game shows color words (RED, BLUE, GREEN, YELLOW) in a conflicting ink color, and the player taps the ink color from four buttons. Three difficulty levels set the response window (3000, 1500 or 800 ms), and catch trials test inhibition. Every trial is timestamped and synchronized with the wearable stream.

FPS runs from 0 to 100. A Base Score for what was achieved (accuracy, speed and stability) is multiplied by an Authenticity Factor for how it was achieved: eight signals of genuine engagement, from the size of the interference effect to slowing down after errors. A session at chance-level accuracy scores zero.

The study tested FPS at three levels:

• Behavioral validity on archival Stroop data: 111,133 trials from 466 adults over three sessions.
• Neural validity on the DMCC55B fMRI dataset: 55 adults.
• Formula validity: the weights checked against four independent sources of evidence, including a confirmatory factor analysis.

• Very large interference effect. FPS captures the Stroop effect with d = 1.339 (a mean gap of 114 ms), replicated in the fMRI sample at d = 1.915.
• Tracks individual differences in attentional control: ρ = 0.785.
• Stable across sessions: test-retest reliability of ICC 0.83–0.93.
• Neurally grounded at the component level: mean incongruent reaction time, the main FPS driver, is significantly associated with activation of the anterior cingulate cortex (ACC), the brain’s conflict monitor. The full composite shows a trend-level association.
• Specific, not generic: no significant link with sustained-vigilance (SART) errors, so FPS measures selective attention rather than general performance.
• A formula check that found a real problem: two components overlapped (r = 0.928), so one became an engagement gate.

Limits the paper states openly: catch trials are not yet validated, the weights stay provisional until they are recalibrated on touchscreen data, and the link between FPS and HRV is still untested.

FPS is the focus score Pulse Focus shows after every session: 0 to 100, with an attention label — Low (below 44), Moderate (44–54) or High (54 and above) — and the session trend. The summary also shows average heart rate and cognitive load labels.

Pulse Focus is live on the App Store and Google Play. It runs on the Synheart human state platform, and every trial is synchronized with the wearable stream, ready for physiological modeling.

A prospective study with at least 50 Pulse Focus players, wearable HRV and a CPT-3 comparison. It will validate the full deployed formula, catch trials included, test how FPS and HRV move together, recalibrate the weights on touchscreen data, and produce the first labeled physiological dataset for attention classification. A planned EEG subset will test the links to brain activity directly.

About

Synheart HSI is a privacy-first, on-device human-state computing infrastructure that enables systems to understand physiological, cognitive, and behavioral states without exposing raw biosignals or personal data. At its core, Synheart introduces the Human State Vector (HSV), a time-scoped, multi-dimensional, pre-interpretive representation of human state, and the Human State Interface (HSI), a standardized contract for exchanging these representations across systems.

“By decoupling sensing, representation, and interpretation, Synheart HSI defines how human state is represented and exchanged—without dictating what it means—ensuring long-term interoperability, reproducibility, and privacy by construction.”

What We Built

Synheart HSI was designed to address foundational failures in today’s human-state and affective computing ecosystems:

• Fragmentation and vendor lock-in caused by incompatible biosignal pipelines and proprietary summary metrics
• Privacy risks stemming from cloud-based inference and transmission of raw physiological and behavioral data
• Lack of a shared abstraction layer between raw biosignals and application-level semantics
• Poor reproducibility across research studies due to inconsistent feature definitions, windowing, and normalization
• Limited real-time usability caused by latency, connectivity dependence, and opaque model outputs

To solve this, we built:

• An on-device HSV runtime that fuses multimodal physiological and behavioral signals into normalized, time-explicit state representations
• A canonical HSI specification that standardizes validation, versioning, confidence reporting, and exchange of human-state data
• A strict privacy boundary that structurally prohibits raw biosignals, semantic content, and PII from leaving the device
• A modular architecture that allows independent evolution of sensors, models, and interpretation layers

The result is a human-state infrastructure that is efficient, interoperable, and deployable on commodity mobile and wearable hardware.

Our Vision

We envision a future where human-state-aware systems are adaptive without being invasive, and intelligent without being opaque:

• Users benefit from context-aware technology without surrendering control over their physiological or behavioral data
• Human-state representations become portable, comparable, and reproducible across devices, studies, and applications
• Developers build adaptive systems using standardized state interfaces, not brittle vendor-specific pipelines
• Researchers share and validate findings using confidence-weighted, pre-interpretive representations instead of irreproducible proxies
• Privacy, consent, and purpose limitation are enforced at the representation layer, not bolted on afterward

Synheart HSI is not an emotion model or a diagnostic system: it is the infrastructure layer that makes responsible human-state computing possible at scale.

Who Can Use Synheart HSI?

Researchers and Scientists
→ Conduct reproducible, multi-site studies on stress, engagement, affect, and behavior using standardized human-state representations.

Wellness Innovators
→ Build state management, and self-tracking tools that operate on-device and respect user privacy by design.

Human–Computer Interaction & AI Developers
→ Create adaptive interfaces, learning systems, and productivity tools that respond to human state without accessing raw biosignal data.

Wearable and Platform Manufacturers
→ Expose human-state capabilities through a stable, interoperable interface without locking users or developers into proprietary formats.

Synheart HSI introduces a pre-interpretive abstraction for human-state computation that separates representation from interpretation. Rather than producing semantic labels such as emotion or focus directly, the system computes a Human State Vector (HSV), a time-scoped, multi-dimensional representation encoding normalized physiological and behavioral state with explicit confidence estimates.

This framework decouples sensing, inference, and application semantics, enabling interoperable and reproducible human-state representations across devices and platforms. By standardizing how human state is represented, without prescribing what it means, Synheart HSI provides a stable foundation for diverse downstream applications while avoiding vendor lock-in and opaque model assumptions.

Key principles

• Separation of concerns: Representation (HSV) is decoupled from interpretation and application logic.
• Pre-interpretive design: Encodes measurable state dimensions without assigning semantic meaning.
• Time-explicit representation: All state estimates are scoped to defined temporal windows.
• Confidence-aware outputs: Each dimension includes a normalized confidence score.

Synheart HSI is implemented as an on-device, layered architecture that enforces strict privacy boundaries. Multimodal signals from wearables and device interactions are processed locally, normalized, and fused into HSV representations over explicit temporal windows. Raw biosignals and semantic content never leave the device.

Derived HSV outputs may be consumed directly by applications or exported through the Human State Interface (HSI), a canonical, versioned specification for exchanging human-state data. Optional interpretation modules (e.g., emotion or focus engines) operate downstream of HSV, ensuring that core representations remain application-agnostic and privacy-preserving.

Core components

• Sensing Layer
  Collects physiological and behavioral signals from wearables and devices.
• Processing Layer
  Performs artifact handling, quality estimation, and per-user normalization.
• HSV Runtime
  Fuses multimodal inputs into fixed-dimensional, time-scoped state vectors.
• HSI Export Layer
  Serializes derived state into a standardized, versioned interface.

Design characteristics

• Privacy by construction — raw biosignals and semantic content never leave the device by default
• On-device execution — low latency and no network dependency
• Modular design — sensors, models, and interpretation modules evolve independently

Synheart HSI establishes a foundational infrastructure for human-state-aware systems by formalizing how human state is computed, represented, and exchanged.

By introducing a standardized, pre-interpretive representation layer, the framework enables:

• Privacy-preserving, real-time human-state computation
• Cross-device and cross-platform interoperability
• Reproducible research and responsible deployment
• Long-term system evolution without vendor lock-in

About

Synheart Behavior is a privacy-preserving framework that models human digital behavior through interaction dynamics rather than content. It transforms low-level events such as taps, scrolling, idle time, and app switching into structured, numerical, and interpretable behavioral representations across event, session, and daily timescales. By focusing on timing, intensity, fragmentation, and interruptions without collecting semantic data or personal identifiers, the framework enables reliable behavioral analysis while maintaining strong privacy guarantees.

What is SynHeart Behavior?

In this paper, Synheart introduces a privacy-preserving framework for modeling digital behavior, defining digital behavior as:

The measurable temporal and structural patterns of a user's interaction with digital systems, independent of content, semantics, or intent.

The Synheart behavior differs fundamentally from traditional psychological or social behavior models. Instead of attempting to infer internal mental states, social relationships, or real-world actions, it focuses exclusively on observable interaction dynamics, such as interaction frequency, session continuity, responsiveness to interruptions, and consistency of use over time. While cognitive and emotional states may indirectly influence these patterns, the framework does not rely on content, semantics, or inferred intentions, ensuring a privacy-preserving approach.

What We Built

• A privacy-first framework that converts raw interaction events into structured numerical behavior metrics
• Behavioral representations across event-level, session-level, and daily timescales
• Interpretable metrics capturing interaction timing, intensity, fragmentation, and interruptions
• Standardized behavioral signals usable directly for downstream modeling and analysis
• An SDK designed to operate without collecting content or personal identifiers

Our Vision

• Enable meaningful understanding of human digital behavior without compromising privacy
• Shift behavior analysis from content-based to interaction-based modeling
• Provide standardized, interpretable behavioral representations across devices and studies
• Bridge raw interaction data and practical, ethical behavioral insights
• Support behavior-aware systems, digital well-being analysis, and adaptive AI models

Scope and assumptions

The framework is restricted to digital interaction metadata, excluding social content, communication meaning, or real-world behavioral observations. Higher-level interpretations emerge solely from temporal and structural patterns of interactions, preserving both user privacy and analytical rigor.

The Synheart Behavior framework is designed around a set of core principles to ensure privacy, interpretability, and scalability while providing actionable insights into digital behavior.

Core Design Principles

• Privacy-First: Collect only non-semantic interaction metadata, ensuring no content, identifiers, or sensitive information is captured. Users retain full control over data collection.
• Behavior-Centric: Focus on interaction patterns rather than content, enabling meaningful behavioral analysis without exposing personal data.
• Temporal Granularity: Data is structured across multiple levels (event, session, and daily) to capture both short-term and long-term behavioral patterns.
• Standardized Representation: Continuous numerical metrics allow consistent aggregation, comparison, and downstream analysis across applications and devices.
• Scalability: Supports on-device computation and modular SDK integration, enabling deployment across multiple applications without centralized data collection.

Conceptual Architecture

1. Data Collection Layer: Captures atomic events (taps, swipes, typing, notifications, calls) and motion context via accelerometer and gyroscope. Event collection occurs only with explicit user permission per application.
2. Data Processing and Representation Layer:} Groups events into sessions, computes inter-event times, and derives session-level behavioral metrics including burstiness, task switch rate, idle ratio, typing behavior, and focus indicators. Metrics are normalized and contextual features (e.g., motion state, orientation) are added. Data is structured in JSON schemas for downstream applications.
3. Aggregation and Habit Layer: Aggregates session-level metrics into daily behavioral representations, computing indicators such as fragmented time ratio, multitasking intensity, behavioral stability, app detachment, and habit strength. Data is structured in JSON schemas for analytics and downstream applications.

High-Level Data Flow

Data Collection and Pre-processing
Our framework transforms raw interaction data into structured behavioral representations across three levels:

• Event level
• Session level
• Daily level

An event is an atomic interaction generated by the user or system at a specific timestamp. Examples include a tap, typing, scroll, swipe, notification events (received, opened, ignored), or call interruption.

Event data is collected only from applications that explicitly receive user permission. The SDK does not track:

• global device usage,
• other applications,
• background system activity.

Each application integrates the SDK independently and receives behavioral analytics only for its own usage context. No text, screen content, audio, images, or identifiers are captured.

This design ensures:

• strict data minimization,
• application-level isolation,
• user transparency and control.

Each event is recorded as:
Event = {event_type, timestamp, session_id, metrics}

Different events expose different measurable properties:

• Tap: duration, long-press flag
• Scroll: velocity, acceleration, direction, direction reversal
• Swipe: direction, distance, duration, velocity
• Typing: duration, speed
• Notification: received, opened, ignored
• Call: answered, rejected, ignored

These metrics form the raw behavioral signals.

Events are grouped into sessions defined as continuous app usage bounded by inactivity gaps (e.g., 30 seconds).

For each session:

• events are ordered by timestamp,
• inter-event times are computed:

These inter-event intervals form the basis for temporal analysis.

Motion State Classification

To enrich behavioral context, motion state is classified every 5 seconds using inertial motion sensor data (accelerometer and gyroscope).

We employ a supervised activity recognition support vector classifier model trained on the UCI-HAR dataset, achieving over 96% accuracy.

The model predicts one of four labels:

• sitting
• standing
• moving
• laying

Motion state is used as contextual metadata rather than a primary behavioral signal, enabling interpretation such as:

• interaction while walking,
• idle behavior while sitting,
• reduced interaction during motion.

A session is a continuous period of interaction within a single application, bounded by:

• app open and close events, or
• a prolonged inactivity gap (e.g., ≥ 30 seconds).

Sessions serve as short behavioral windows (typically 1–5 minutes) for aggregation.

Behavioral Session Data Numerical Representation

• Burstiness
• Notification Load
• Task Switch Rate
• Task Switch Cost
• Idle time
• Fragmentation Idle Ratio
• Scroll Jitter Rate
• Distraction Score
• Focus Hint
• Interaction Intensity
• Deep focus block

Daily Aggregation of Behavior

Session-level metrics are aggregated over a 24-hour period to compute daily behavior indicators.

• Fragmented Time Ratio
• Screen Time Segments
• Recovery-Friendly Minutes
• Multitasking Intensity
• Behavioral Stability Score
• Micro-Tasking Frequency
• Night Activity Ratio
• Average Session Spacing
• App Detachment Score
• Habit Strength Index

These metrics capture long-term behavioral patterns rather than momentary states.

The Synheart Behavior framework is designed as a general-purpose behavioral representation layer rather than a task-specific model. By producing structured, numerical, and privacy-preserving behavioral metrics, the framework can support a wide range of downstream analytical and learning-based applications without relying on content-level data. The following representative application domains illustrate its potential utility and relevance.

Cognitive State and Engagement Modeling

Aggregated interaction patterns such as session duration, activity transitions, inter-event timing, and response latency can serve as input features to models estimating latent cognitive states, including engagement, attentional stability, and mental workload. Synheart Behavior provides a content-independent behavioral signal that can be used alone or combined with other modalities (e.g., physiological or contextual data) for longitudinal analysis. This enables researchers, UX designers, and educational technology developers to monitor digital engagement without accessing sensitive content.

Adaptive and Context-Aware Systems

Numerical behavioral representations derived from interaction dynamics can inform adaptive system behaviors, such as interface adjustments, intelligent notification scheduling, or task pacing. By relying solely on interaction patterns rather than semantic content, these adaptations are privacy-aware while remaining responsive to real-time changes in user behavior. Stakeholders include software developers, HCI researchers, and product teams aiming to enhance usability and user experience ethically.

Personal Analytics and Self-Monitoring

Synheart Behavior metrics can support personal analytics applications that summarize digital habits, usage rhythms, and session fragmentation over daily or longer time horizons. Individuals can gain actionable insights into their own device use without exposing sensitive content or identity-linked data. Example applications include productivity dashboards, self-reflection tools, and digital well-being trackers.

Health and Well-Being–Oriented Applications

Interaction-based behavioral representations may inform exploratory analyses in well-being–oriented contexts, such as detecting prolonged inactivity, irregular interaction patterns, or disrupted daily routines. While the framework itself does not infer clinical or psychological states, it provides standardized behavioral inputs that can be interpreted by downstream models within clearly defined scopes. Potential stakeholders include researchers in occupational health, human factors, and digital mental health, enabling ethically responsible monitoring and intervention design.

Broader Implications and Capabilities

By abstracting digital behavior into interpretable, standardized metrics, Synheart Behavior enables:

• Privacy-preserving behavioral analytics across multiple applications and platforms.
• Consistent, comparable behavioral features for research, product development, and personalized services.
• The development of adaptive, context-aware, and ethically grounded systems that respond to user behavior without accessing sensitive content.
• Empowering individuals with self-monitoring tools that inform habits, productivity, and well-being.

Our framework bridges the gap between raw interaction data and actionable behavioral insights, providing new capabilities for stakeholders across research, product design, and personal analytics while maintaining privacy and ethical standards.

Future directions for extending this work include:

• Longitudinal modeling to capture behavioral changes over extended periods.
• Personalization of normalization parameters to account for individual differences in interaction patterns.
• Learning cross-application behavioral embeddings to enable unified representation across multiple digital environments.
• Integration with additional data modalities, such as physiological or contextual signals, to enrich behavioral insights.
• Exploration of self-supervised learning objectives for deriving more generalizable and robust behavioral embeddings.

By clearly defining the current scope and identifying areas for refinement, this framework provides a transparent foundation for future research and application while maintaining privacy, interpretability, and ethical standards.

This paper presents Synheart Behavior, a structured, privacy-preserving framework for numerically representing digital behavior. By transforming low-level interaction events into interpretable session- and day-level metrics, the framework captures the temporal and structural patterns of user interactions while remaining fully independent of content or personal identifiers.

The proposed representation enables a wide range of downstream applications, from cognitive and engagement modeling to adaptive systems and personal analytics, all while maintaining strong ethical and privacy standards. Key insights include the importance of temporal dynamics, session aggregation, and standardized behavioral metrics as powerful indicators of human behavior in digital environments.

Overall, this work highlights that how people interact rather than what they consume is a rich, underutilized signal for understanding behavior. Future research and applications can build on this foundation to develop longitudinal analyses, personalized metrics, and cross-application representations, further advancing ethical and actionable digital behavior modeling.

About

Synheart Emotion is a privacy-preserving on-device emotion recognition system that utilizes wrist-based photoplethysmography (PPG) biosignals. The system processes Heart Rate Variability (HRV) features to classify emotions while ensuring user data remains on the device, addressing critical privacy concerns in emotion detection applications.

Human–computer interaction increasingly demands systems that recognize not only explicit user inputs but also implicit emotional states. While substantial progress has been made in affective computing, most emotion recognition systems rely on cloud- based inference, introducing privacy vulnerabilities and latency constraints unsuitable for real-time applications.

What We Built

The Synheart Emotion system addresses fundamental challenges in emotion recognition from biosignals:

• Privacy concerns with cloud-based emotion recognition systems that transmit sensitive biosignal data.
• Limited computational resources on wearable devices requiring efficient model optimization.
• Need for real-time emotion classification without network latency or connectivity requirements.
• Balancing model performance with memory footprint and inference speed on edge devices.

Our Vision

We envision a future where emotion-aware wearable technology respects user privacy while delivering accurate insights:

• Users can benefit from emotion recognition without compromising their sensitive biosignal data.
• Wearable devices become more intelligent and responsive to emotional states in real-time.
• Healthcare providers gain valuable tools for mental health monitoring with patient consent.
• Developers can build privacy-preserving emotion-aware applications using standardized on-device models.

Who Can Use Synheart Emotion?

• Healthcare Professionals → Monitor patients' emotional states for mental health assessment and treatment.
• Researchers → Study emotion patterns and develop new affective computing applications.
• Wearable Manufacturers → Integrate privacy-preserving emotion recognition into their devices.
• App Developers → Build emotion-aware applications that respect user privacy and work offline.

Research Methodology

Our research phase involved comprehensive analysis of biosignal processing techniques, feature extraction methods, and machine learning model comparisons to identify the optimal approach for on-device emotion recognition from PPG signals.

1: Dataset & Signal Processing

WESAD Dataset

• Utilized the WESAD (Wearable Stress and Affect Detection) dataset with PPG signals from wrist-worn devices.
• Dataset contains biosignals from 15 subjects performing various activities designed to elicit emotional responses.
• PPG signals sampled at 64 Hz providing high-quality cardiovascular data for HRV analysis.
• Emotions labeled as Baseline, Stress, Amusement, and Meditation states.

HRV Feature Extraction

We extracted five critical HRV metrics using NeuroKit2 library with 60-second sliding windows:

SDNN
Standard deviation of NN intervals

Mean RR
Average interval between heartbeats

PNN50
Percentage of NN intervals differing by more than 50ms

Mean HR
Average heart rate over the window

RMSSD
Root mean square of successive differences

2: Model Evaluation Approach

We systematically compared three categories of machine learning models to identify the optimal approach for on-device emotion recognition:

1. Classical Machine Learning
   Evaluated traditional algorithms including ExtraTrees, Random Forest, AdaBoost, XGBoost, Logistic Regression, and SVM for their efficiency and performance on HRV features.
2. Deep Learning Models
   Tested 1D-CNN and LSTM architectures to assess whether deep learning could extract additional patterns from sequential PPG data beyond handcrafted HRV features.
3. Transformer Architecture
   Explored multi-head attention mechanisms to model complex temporal relationships in HRV feature sequences for emotion classification.

3: Key Research Findings

From our comprehensive analysis, we derived several critical insights:

4 : Model Comparison Analysis

We evaluated three categories of machine learning approaches for emotion recognition:

Classical Machine Learning Models

• ExtraTrees achieved highest F1 score of 0.826 with excellent balance of precision and recall.
• Random Forest and AdaBoost also performed well, demonstrating ensemble methods' effectiveness.
• Logistic Regression and SVM showed moderate performance, suitable for simpler deployments.
• Classical models offer smaller memory footprints ideal for resource-constrained devices.

Implementation Details

This section outlines the technical implementation of the Synheart Emotion system, detailing the software stack, model deployment pipeline, and on-device integration strategies.

Technology Stack

Deployment Pipeline

The process of taking a trained ML model and deploying it onto a target device involves several critical steps to ensure efficiency and compatibility.

1. Model Training & Validation
   Train and validate models using HRV features and WESAD dataset.
2. Conversion to ONNX Format
   Use `skl2onnx` to convert scikit-learn models to ONNX.
3. On-Device Inference Engine
   Integrate ONNX Runtime on the target wearable device.
4. Feature Engineering & Preprocessing
   Implement real-time HR data signal processing and HRV feature extraction.

Performance Metrics

The Synheart Emotion system achieves state-of-the-art performance using optimized classical machine learning models on-device. This section details the key performance indicators and comparative analysis.

Model Comparison