Scientists are increasingly relying on wearable sensors and machine learning to develop digital biomarkers. However, their successful development requires the identification of physiological events relevant to the disease state. Here's an overview of the method we use to accurately and efficiently label physiological events from biosignals collected from wearable sensors.
1. Raw Data vs. Summary Measures vs. Validated Biomarkers
Wearable sensors assess a subject over minutes, hours, or days to produce waveform data and summary measures that provide a general assessment of their physiological state. We can use the summary measures to determine a general trend of the subject's physiology.
Without accurate labeling, however, summary data may smooth over disease-specific physiologic events, misidentifying or missing them entirely. When critical events are correctly defined and labeled, they can be extracted using developed algorithms and artificial intelligence.
2. Human Expertise
We created VivoSense® wearable sensor data visualization and analytics software to assist highly trained human analysts to label data accurately using well-defined algorithms and AI tools. The software helps our data analysts efficiently find and identify candidate events. They can quickly scroll through vast amounts of physiological data, even at high resolution, and label all events while removing potential false-positives. We call this VivoSense® Human Augmented AI.
3. Event Adjudication
Labeling may be subject to the subtle bias of individual human reviewers. VivoSense® allows multiple scorers to label the same data with 3rd party adjudication of any disagreement (kappa score). It learns the inter-analyst reliability and trains future machine learning-based labeling against a human adjudicated gold-standard.
Efficient and accurate labeling of clinically significant events allows the construction of models that guide the development of novel digital biomarkers of disease.
4. Tools of Human Augmented AI
VivoSense® software has numerous tools that augment human expertise, including:
- Detailed scale and zoom tools to enable clear visualization of the waveform data and events, which helps analysts assess the quality of an event quickly.
- Pairing the visualization tools with complex thresholding criteria algorithms and the derivation of more complex measures from the waveform and summary data.
- Synchronizing and combining measures and waveforms from multiple devices and or different physiological data (such as heart rate, respiration rate, temperature, and activity) to provide context-based review and analysis of multiple signals simultaneously.
5. Machine Learning and Artificial Intelligence
Once sufficient disease specific physiological events are accurately labeled, VivoSense® uses purpose-built machine learning tools to build models that automatically identify the physiological events for the specific disease. Then data analysts can review performance statistics of the models against hold out validation sets by partitioning the data into training and validation data sets.
Successful machine learning can only be accomplished with accurately labeled data. If you'd like to explore using our method of deriving novel digital biomarkers for your clinical trial, contact us today.