WATCH SENSORS
ECG
BP / PPG
BIA / SpO₂
Lifestyle
WEARABLE-AI SCIENCE & EVIDENCE
Evidence built for real-world Wearable-AI.
Longitudinal wearables, CGM, clinical biomarkers and biochemical ground truth feed a cardiometabolic knowledge base—then NeuTigers translates that intelligence into compact, privacy-preserving models for on-device deployment.
Request an evidence review ↗Cardiometabolic knowledge base
≈2,000 Participants across US/EU/MENA 
Longitudinal Multimodal 
Digital + Clinical Data 
Rare Disease-State Continuum
CGM
Glucose dynamics
Variability
7Days/24h profiles
Free-living
CMH BIOMARKERS
Cardiovascular
Liver
Lipid profile
Other labs
GLUCOSE SOC
FPG
HbA1c
OGTT 1h
OGTT 2h
METADATA
Demography
Risk Factors
Medications
Outcomes
Disease-State Continuum
AT-RISK NON-DIABETES~40% 
DE NOVO PREDIABETES~10% 
DE NOVO TYPE 2 DIABETES~5% 
KNOWN PDs and T2Ds~25% and ~20%
Unique value: the same participant can connect Wearable physiology → CGM dynamics → Cardiometabolic biology → Biochemical ground truth → Outcomes.
Product swarms system
Data flywheel: Each new cohort expands phenotype coverage and creates a faster starting point for the next CMH algorithm or trial use case, e.g. women’s health with established gestational diabetes.
SweetDeep™ Watch
Non-invasive Type 2 Diabetes case-finding
82.5% accuracy 84.5% with abstention
SweetDeep™ CGM
24h Glucose-dynamics classifier
90.2% accuracy 95.1% AUROC
pre-SweetDeep™ Watch
Prediabetes phenotype
85.5% development 88%–91.8% blind evaluation
Clinical validation evidence
SweetDeep™ – AI Models Performance
SweetDeep-Watch
Non-invasive front-end screening
- Study design
- 285 participants • EU/US/MENA • 7-days free-living DCT • Samsung Galaxy Watch
- Inputs
- Daily - ECG + PPG/BP + BIA + age / family history + time-of-day
- Model
- <3k parameters • edge-optimized • patient-level aggregation • abstention for low confidence
82.5% Accuracy 82.1% Macro-F1 79.7% Sensitivity 84.6% Specificity 5.5% ECE
Clinical use: Low-burden screening signal to route at-risk adults to confirmatory biochemical testing.
SweetDeep-CGM
24-hour glucose dynamics classifier
- Study design
- 551 participants • 304k overlapping 24h windows • 7-days free-living DCT • Abbott CGM Freestyle Pro
- Inputs
- Raw 24h CGM traces • no standardized meal or medication restrictions
- Model
- 13k-parameter CNN • calibrated probabilities • external validation in EU/MENA + U.S.
90.2% Accuracy 95.1% AUROC 87.6% Macro-F1 82.3% Sensitivity 3.5% ECE
Clinical use: Short-window metabolic phenotyping to support early case-finding and longitudinal response tracking.
What this enables
for pharma GLP-1 / Obesity-Overweight Companion Stratification Trial Enrichment + Digital Endpoints Payer-Facing Evidence & Care Routing Decentralized Validation Platform The same evidence architecture can support population screening, treatment-response monitoring, RWE, and market-access value demonstrations.
for pharma GLP-1 / Obesity-Overweight Companion Stratification Trial Enrichment + Digital Endpoints Payer-Facing Evidence & Care Routing Decentralized Validation Platform The same evidence architecture can support population screening, treatment-response monitoring, RWE, and market-access value demonstrations.
Validation against standard of care
CLINICAL PERFORMANCE (PATIENT-LEVEL)
| METRIC | SWEETDEEP™ (Wearable-AI) | FINDRISC (Standard Questionnaire) | IMPROVEMENT (SweetDeep vs FINDRISC) |
|---|---|---|---|
| Accuracy | 84.5% | 68.8% | +15.7 pts |
| Sensitivity | 81.3% | 78.1% | +3.2 pts |
| Specificity | 85.2% | 66.9% | +18.3 pts |
| Negative Predictive Value (NPV) | 95.8% | 93.9% | +1.9 pts |
| AUROC | 89.6% | 84.7% | +4.9 pts |
KEY CLINICAL FINDINGS
~20–25% higher accuracy than the widely used FINDRISC questionnaire
Better balance of sensitivity and specificity for more efficient screening
High NPV (95.8%) supports reliable identification of low-risk individuals
Scalable & Non-invasive population screening with consumer wearable devicesWHY IT MATTERS
SweetDeep™ combines wearable physiology with AI to identify non-invasively diabetes risk more accurately than questionnaire-based screening, enabling earlier identification of at-risk individuals and more efficient referral for confirmatory laboratory testing.
FDA wellness guidance
Key Points Relevant to SweetDeep™
Positioning Discipline : Wellness insights only, Validated sensor values, No diagnostic or treatment claims
WHAT FDA CLARIFIED
General-wellness products may include: “may help reduce risk” / “may help living well with” chronic conditions when generally accepted (incl. type 2 diabetes)
FDA may consider certain non-invasive sensing outputs (e.g., blood pressure, blood glucose, HRV) as general-wellness when intended solely for wellness uses and not for diagnosis/treatment.
WHAT THIS ENABLES FOR MARKET
Faster U.S. launch via a wellness positioning (information + coaching + confirmatory-test routing)
Reduced regulatory friction when claims, labeling, and UX avoid medical/clinical context
Competitive advantage — first-mover consumer distribution while maintaining optional regulated roadmap for clinical claims
FDA GUIDANCE ALIGNMENT
Consistent with: FDA Wellness Policy (Jan 2026) • Breakthrough Devices Program Guidance (Feb 2022)
“The program is intended to expedite the development and review of medical devices that provide for more effective treatment or diagnostic of life-threatening or irreversibly debilitating diseases or conditions” — U.S. Food and Drug Administration
SweetDeep™ publications
MFDeep™
Edge-AI™ advantage
1,000x more energy efficient 1,000x Smaller model size 2x faster latency 1.2x more accuracy Deployable on any wearable HIPAA/GDPR compliant
Technical Edge: Wearable-AI™ Architecture
- Wearable-AI™ = AI that runs locally on sensor-enabled device (no cloud), enabling ultra-fast, private, real-time decisions
- Uses NeuTigers’ Grow-and-Prune™ deep learning to compress models by 1,000x without sacrificing accuracy
- Trained on decentralized real-world data (wearables, clinics) → highly generalizable and robust
- No new hardware required — models deploy seamlessly on smartwatches, rings, phones, and embedded IoT devices
- HIPAA/GDPR compliant by design — no transmission of raw biometric data.
Proprietary Technology & IP
- US Patent No. 17,613,284: G&P: Edge-AI Cross-and-Prune Learning Architecture for patient-side real-time biometric monitoring system for wearables
- International/US Patent PCT/US No. 19/41531: SCANN: Syn Compact and Accurate Neural Networks and Method for Synthesis of Compact and Accurate Neural Networks
- US Patent No. 63/647,141: DLN: Differentiable Logical Networks
Multiple Diseases Applications Filed
- US Patent No. 17,619,449: DiabDeep: Pervasive diabetes diagnosis based on wearable medical sensors and efficient neural networks
- US Patent No. 18/235,422: MHDeep: System and method for Mental Health Detection System Based on Wearable Sensors and Artificial Neural Networks
Differentiation
DEFENSIBLE TECHNOLOGY. CLINICAL PROOF. STRATEGIC DISTRIBUTION. Building the Metabolic Intelligence Layer for Every Wearable.
1. PROPRIETARY EDGE-AI
Patented Edge-AI Algorithms and Apps Protect our Technology, Data and Disease Models.
3. ON-DEVICE DEPLOYMENT
Ultra-Efficient On-Device AI that Runs Anywhere with Privacy by Design.
2. CLINICAL KNOWLEDGEBASE
Unique Knowhow in Decentralized Clinical Trials, Large-Scale Sensors and Clinical Data Sets, Deliver High Bar Clinical Proof.
4. STRATEGIC ECOSYSTEM
Integrations with Global Wearable OEM, Top Ivy League University, International Clinical Partners across US/EU/MENA.
Figures and claims are reproduced from the supplied 2026 project source and remain subject to source verification, publication linkage, permissions, and medical, regulatory and legal review before public launch.
RESEARCH FAQ
Questions the evidence answers.
Direct answers drawn from the studies on this page. Investigational research; not a cleared medical device.
How accurate is smartwatch-based diabetes risk screening in studies?
In NeuTigers’ prospective SweetDeep-Watch study (285 participants across the EU, US and MENA, seven days of free-living data on Samsung Galaxy Watch), the investigational model reached 82.5% patient-level accuracy against biochemical ground truth, with 82.1% macro-F1, 79.7% sensitivity, 84.6% specificity and a 5.5% expected calibration error. Allowing the model to abstain on fewer than 10% of low-confidence cases raised accuracy to 84.5%. This is investigational research, not a cleared medical device.
How does wearable-AI screening compare with the FINDRISC questionnaire?
In a patient-level comparison on the same population, the investigational SweetDeep-Watch model outperformed the standard FINDRISC questionnaire on every reported metric: accuracy 84.5% vs 68.8% (+15.7 points), specificity 85.2% vs 66.9% (+18.3 points), sensitivity 81.3% vs 78.1% (+3.2 points), negative predictive value 95.8% vs 93.9%, and AUROC 89.6% vs 84.7%.
Can CGM data alone signal type 2 diabetes risk?
In NeuTigers’ SweetDeep-CGM study, a 13,000-parameter convolutional neural network trained on raw 24-hour CGM traces from 551 participants (304,000 overlapping windows, no standardized meal or medication restrictions) classified type 2 diabetes vs non-diabetes with 90.2% accuracy, 95.1% AUROC, 87.6% macro-F1 and a 3.5% calibration error, with external validation across France + Algeria and United States cohorts. Investigational research.
What signals does the wearable model use?
Daily two-minute recordings of ECG, PPG/blood pressure and bioelectrical impedance (BIA), combined with age, family history and time-of-day. The model has fewer than 3,000 parameters, runs on-device (Edge-AI), aggregates predictions at patient level and abstains when confidence is low.
Is SweetDeep FDA cleared?
No. The planned initial consumer experience is positioned as general wellness, which FDA does not review, and it is not intended to diagnose, treat, cure or prevent disease. The screening capability described on this page is investigational; NeuTigers is evaluating De Novo and 510(k) routes for regulated B2B uses.
What is the KB-PCMHD cardiometabolic knowledge base?
A research knowledge base of roughly 2,000 study participants across the US, EU and MENA that links, at participant level, watch-sensor physiology, CGM glucose dynamics, cardiometabolic biomarkers, biochemical reference tests (FPG, HbA1c, OGTT) and outcomes. It serves as the substrate for developing and validating new models.
Where is the research published?
The SweetDeep study, “SweetDeep: A Wearable AI Solution for Real-Time Non-Invasive Diabetes Screening”, is available as a preprint on arXiv (2512.03471). First pages and abstracts are reproduced in the publications section above.
QUALIFIED DILIGENCE