VEDES AI™
Physics-Aware Causal AI Platform for Very Early Detection of Emerging States
VEDES AI – Physics-Aware Causal AI Platform for Very Early Detection of Emerging Psycho-Physiological States from Low-Resolution Data Collected 24/7 by Your Wearables (FitBit, Apple Watch etc)
(mobile app coming soon)
Problem
- Innovations in biomedicine enhance our life quality and longevity by offering effective and affordable means for physical and mental health maintenance
- However, many fundamental challenges remain hard to solve
- “Magic pills for every developed abnormality may not exist even in a distant future
- Modern generic treatments are often ineffective or have side effects for x% of patients
- No universal treatment of complex / rare abnormalities and tedious trial-and-error process
- Early diagnostics and effective personalization could solve many problems, however
- Clinical diagnostic is not available 24/7 making early detection of intermittent patterns impossible
- Biomarkers detectable from express blood tests are limited in frequency and covered abnormalities
- Efficiency of genome-only personalization is very limited due to multi-gene nature of many abnormalities and ignorance of epigenetic factors including life styles and environments
- Most of these procedures require extra effort which limits their chance of regular usage in healthy state
- Is 24/7 objective detection of wide range of emerging abnormalities and treatment effect trends currently realistic?
- Yes, we offer such a platform that does not require any extra effort from users
Promise of Physiological Signal Analysis
- Physiological signals, collectable by wearables and portable clinical devices, quantify dynamical state aggregating impact from all factors, not just genetics
- Many physiological patterns (e.g. in ECG waveforms) are effectively used in clinical diagnostics
- Variability analysis of physiological signals (heart rate, gait, EMG, EEG) are effective not only for complementary diagnostics but also for very early detection of emerging abnormalities not observable by other clinical modalities
- Variability analysis (e.g. heart rate variability - HRV) provide practical objective means of detecting not only developing medical abnormalities but also wide range of psychological / psychiatric conditions (e.g. depression and anxiety) and various stress states including overtraining in sports
Fundamental Challenges of Physiological Signal Analysis
- Performance of HRV and other variability indicators significantly deteriorate when applied to short data segments which is necessary for early detection of intermittent patterns
- Even the best variability indicators completely lose their discriminative abilities when applied to low-resolution (averaged) data such as heart rate continuously collected by wearables which makes affordable 24/7 monitoring impossible
- Indicators discriminative ability for normal vs many abnormal states are verifiable except for rare and complex conditions with very limited data
- However, there are simply no data for unlimited number of possible transition paths from healthy to abnormal states which may take many months or years to develop
- Thus, there are no objective means to discover, select or confirm performance of the indicators that are not only capable of normal-abnormal discrimination but also capable of detecting subtle initial trends well before even early stages of abnormality
Our Solution
- Our indicators based on deep integration of physics-based models and non-standard customization of advanced ML/AI techniques perform well on low-resolution data such as heart rate collected 24/7 by most modern wearables
- Our multi-expert indicators have capacity to describe wide range of novel states and subtle changes without re-training which includes unique personalized complex states, e.g. states of optimal performance or rare abnormal states
- Our collection of physics-based models capable of producing realistic synthetic heart rate data can be used to generate any number of virtual patient data including complex / rare abnormalities and millions of long-term transition paths from healthy to abnormal states where real data is simply unavailable
- With these verifiable capabilities, we can use 24/7 low-sampling data from any wearables already used by our customers and offer very early detection of negative (or positive) subtle trends undetectable by standard HRV or more sophisticated clinical check-ups
OUR SOLUTION: Verified Capabilities of Early Detection
- While superior discriminative abilities of our multi-expert indicator can be verified given data for normal/abnormal states, the key capability for very early detection of emerging abnormalities or treatment effects cannot be directly verified since normal/abnormal transition data is not available
- However, using simulated transition paths with varying sampling rates from real normal/abnormal heart rate data, we can verify that our multi-expert indicator clearly detects emerging trends while standard HRV indicators completely fail in early trend detection (see right charts)
- Our realistic synthetic data can be used to test indicator effectiveness on even more subtle and multiple normal-abnormal transitions where we also clearly observe early detection capability of our indicator and complete failure in this task of standard HRV indicators (see left chart)
- Monitoring distance of our full multi-expert state vector to different stress and optimal performance states provides early detection of even more subtle personalized trends in psycho-physiological condition where standard HRV indicators are not applicable at al
Computed on synthetic data generated for different driver parameters in our physics-based model
Computed on real data sampled from normal & abnormal segments: 100% abnormal before orange line and linearly decreasing to 95% abnormal after orange line
Our Discovery Framework: Smart Usage of All Domain Knowledge
Hybrid Framework for Discovery of Multi-Expert Personalized Biomedical Models & Indicators
- Direct usage as aggregated ensemble
- Using state vector of ensemble components for objective representation of complex and rare psycho-physiological conditions
- Using ensemble components as complementary features for further enhancements with deep learning and other ML techniques
- Using ensemble components for discovery of robust invariants with computational topology techniques
Our Physics-based Models: Generation of Realistic Synthetic Data
- Availability of large amount of physiological data (e.g. heart rate) is limited to healthy subjects and patients with common abnormalities
- Data for various rare /complex conditions are either very scarce or not available at al
- There are simply no data for huge number of possible transition paths from healthy state to various common or rare abnormalities
- Our physics-based models are capable to generate realistic synthetic data for millions of transition paths and rare / complex states
- Our synthetic data are used for optimal selection of indicators with early detection capabilities and discovery of additional predictive features
© 2019 Medpobs. Powered by VOOV.