Residual-Based Analytics

Human physiology is dynamic. Your blood pressure, heart rate, respiration, blood chemistry, peripheral circulation, oxygenation, skin temperature and everything else that can be measured about you changes constantly in response to your activity, the time of day, your environment, your last meal, and so on. Moreover, your physiology can react differently than someone else's physiology.

Conventional thresholds do not have visibility into anomalous measured values inside the population "bell curve".

So making a single measurement, a "spot check" at just one moment, doesn't provide a very clear picture of your health. But this is standard medical practice. Your biosignals are expected to be within a certain range considered "normal" for the population. The range limits are set wide to accommodate the variation seen in the population. As a consequence, only extremely anomalous measurements are detected. Any differences between an in-range measured value and its ideal value are missed.

What goes largely unacknowledged in the conventional approach is that human biosignals of every kind undergo enormous variation throughout the day even in a perfectly healthy subject. This normal variation confounds healthcare's "spot check" approach, as a parameter must first be outside of its normal range of variation before it can reliably be deemed abnormal. As a consequence, analytics in medicine today are hardly “predictive” at all, but instead merely diagnostic.


Residual-based analysis is a better way to monitor biosignals. In this methodology, a measurement of a biosignal is compared to a dynamic baseline that accounts for a person's dynamic state. Discrepancies between the baseline and the measured biosignal — called residuals — indicate incipient deviations from normal health.

Comparison to a dynamic baseline allows for detection of subtle discrepancies, even when the biosignal is well within its customary range of variation.

Thresholds are applied to the residuals (instead of the measured values), resulting in much tighter detection limits. Such deviations can be detected even though the measured signal is still within the normal range of its variation. Conventional medical limits would be blind to such changes. As a result, nascent health issues can be detected significantly earlier with residual-based analytics.


This methodology has been successfully deployed in many industrial monitoring applications (aircraft engines, power plants, vehicles, refineries, computers), where it is used to identify maintenance needs well before the equipment breaks down. In order for residual-based analytics to work, an effective dynamic baseline is required that characterizes normal behavior of the monitored system. In industrial applications, this has been available in the form of design blueprints and "first principles" engineering models of the monitored equipment, which can simulate normal machine behavior over the span of designed operational states.

More recently, these engineering models have been surpassed by data-driven learning models based on artificial intelligence methods. These models learn normal behavior from example data, and have greater fidelity than engineering models. Furthermore, such learning models are not constrained to work with systems for which design blueprints exist. This is particularly important in medicine, where no "first principles" model of the human system exists.


VG-BIO's predictive analytics are built on a foundational data-driven artificial intelligence technology called Similarity-Based Modeling, or SBM. This technology "learns" from exemplary data characteristic of the normal dynamic behavior of any target system, and thereafter can be used as a model of that system (such as a human patient). Measured biosignals from the patient can be compared to estimates from the model of what the biosignals should be, providing the dynamic baseline for detection of health issues long before the patient becomes overtly symptomatic.

SBM does this by inherently learning the relationships that couple the behavior of the parameters comprising the exemplary data, without any explicit mathematical definition of the monitored system. For example, if interrelated measurements of pressure, temperature and flow rate (as measured by sensors) characterize a mechanical system, SBM can learn that system's behavior from a set of several example measurements of those parameters, without any need to define equations that relate the pressure, temperature and flow rate from a "first principles" perspective.

We believe we can accomplish the same result for the human "system", which is a finely tuned machine with a sophisticated control system and internally balanced chemical environment. Please see our applications pages to see how we are applying this advanced technology to improve healthcare.