Measured,
Not Assumed.
A.R.I.A.'s value is grounded in real data. We found that in field conditions, architecture choice does not separate outcomes — labels do. That finding reframes the entire field.
What did Anecoica discover?
Anecoica's core research finding is that in field wearable conditions, the choice of machine-learning architecture does not separate outcomes — the labels do. On the DAPPER field dataset (N=84 cohort), the three architectures we tested — XGBoost with NeuroKit2 features, logistic regression on frozen NormWear embeddings, and XGBoost on frozen NormWear embeddings — converge between 0.534 and 0.565 balanced accuracy at 30 labeled moments (N=20 at that budget). No pairwise comparison reaches significance after correction, and no large effect size appears across 21 tests; medium effects cannot be ruled out at this sample size. This is consistent with published field results (Smets 2018, F1=0.43, N=568; Google LSM-2, F1=0.683, 40 million hours of pretraining, 60,000+ participants). The bottleneck is not algorithmic capacity but the quality and type of per-user labels collected during calibration. Under lab conditions on the WESAD benchmark (N=15, binary), a short structured calibration session covering both activated and resting states lifts balanced accuracy from 79.4% to 89.8%. This reframes the field from “build better algorithms” to “build better calibration protocols.” Anecoica's 25-page paper, submitted to ACM IMWUT in May 2026, is the first systematic measurement of calibration cost across model families under identical protocols.
Research Results
Lab Validation — WESAD, N=15, Binary
89.8% balanced accuracy
Protocol-induced stress detection with a short structured calibration session covering both activated and resting states. In line with published results for wearable affect detection under controlled conditions. Calibration lifts balanced accuracy from 79.4% to 89.8% over the zero-shot baseline.
Field Study — DAPPER (N=84 Cohort)
~56% balanced accuracy
Self-report labels under daily-life conditions. The three architectures we tested converge between 0.534 and 0.565 at 30 labeled moments (N=20 at that budget). No pairwise comparison reaches significance after correction, and no large effect size appears across 21 tests. Consistent with published field results (Smets 2018: F1=0.43, N=568; Google LSM-2: F1=0.683, 40M hours pretraining, 60,000+ participants).
Field accuracy sits far below lab numbers across published methods. Google needed 40M hours of pretraining data and 60,000+ participants to reach F1=0.683. The gap between lab and field is universal. Medium effects cannot be ruled out at this sample size.
47% of subjects achieve above-chance accuracy; 53% cluster at chance. Hartigan's dip test does not support bimodality in this distribution (D=0.038, p=0.62). Understanding what distinguishes learnable subjects is an open research question.
Consumer Hardware — Galaxy Watch
Validated on Galaxy Watch 5 data
71.8% balanced accuracy on Samsung Galaxy Watch 5 data (GalaxyPPG, N=23; heart rate + motion only). In that experiment, dropping the ECG channel that research devices have costs less than 1 percentage point. Live raw-PPG support today runs on Galaxy Watch 8 — a separate deployment path, not an accuracy claim.
Novel Contribution — Paper 1
“The Cost of Personalizing Wearable Arousal Detection: A Cross-Paradigm Comparison of Label Type and Architecture”
No prior paper compares calibration cost across model families under identical protocols. Our data shows that in field conditions, architecture choice does not separate outcomes at the label budgets we tested — the quality and type of labels from users is what limits accuracy. This reframes the entire approach from “better algorithms” to “better calibration protocols.” 25-page paper submitted to ACM IMWUT, May 2026.
Engineering Validation
437 automated tests with full ML data leakage audit
Production-grade codebase with comprehensive test coverage across the full inference pipeline. Includes automated checks for ML data leakage — ensuring no information from test subjects contaminates training, a common source of inflated accuracy in published wearable studies.
Real-Time Inference
Real-time streaming API with WebSocket support for live inference
Demo API (FastAPI) with POST /v1/stream and WebSocket /v1/ws for live sensor data to prediction pipeline. Supports real-time wearable data ingestion with streaming predictions — the foundation for consumer device integration.
Public Demonstrations
A.R.I.A. has been demonstrated live at major international events under real-world, high-pressure conditions — not controlled lab environments.
Sónar+D 2025
Barcelona
Live installation
Science Week 2025
Berlin
Live installation
Science of Consciousness
2025
Featured presentation
Quantum Basel
Switzerland
Quantum-inspired emotional mapping
Research Infrastructure
Validation Lab
GPU-equipped research environment for pipeline validation across public and proprietary datasets.
Field Data Collection (Planned)
GDPR-compliant data collection planned for field validation, building a proprietary dataset from consumer wearables under ecological conditions.
Research Pipeline
Automated experiment pipeline for reproducible validation across datasets, model architectures, and calibration conditions.
For Researchers
A.R.I.A.'s calibration methodology achieves 89.8% balanced accuracy in controlled conditions (WESAD, N=15, binary, leave-one-subject-out cross-validation) after a short structured calibration session covering both activated and resting states. We're measuring whether this transfers to daily life.
If you run wearable affect studies, we'd like to analyze your data with our calibration pipeline — at no cost, under NDA if needed.
What you get: per-subject accuracy with and without personalization, calibration gain analysis, architecture comparison across your dataset.
Get in TouchOne-Pager Summary
A two-page B2B brief covering the calibration thesis, the per-user personalization layer, current Layer 1 evidence across three datasets (WESAD, DAPPER, GalaxyPPG), and the working prototype API.
Academic Foundations
A.R.I.A.'s approach is grounded in peer-reviewed research across neuroscience, affective computing, and signal processing.
Explainable ML with Feature Attribution
Classification models paired with SHAP-based feature attribution. Every inference traces back to specific physiological features — transparent, auditable, and designed for clinical trust.
Heart Rate Variability Analysis
Advanced HRV-spectral analysis techniques that go beyond standard time-domain metrics. Multi-modal fusion with accelerometry and electrodermal activity for robust physiological state estimation.
Contextual Data Fusion
Designed to integrate physiological signals with environmental context (calendar, location, activity type) for contextual attribution — distinguishing exercise-induced arousal from cognitive stress as the platform matures.