Clinical Predictive/Surveillance Scores: Past, Present, and Future

The world has embraced consumer clinical devices for some time now – the Apple Watch is the latest example of consumer fascination with technology-based self-diagnostic tools.

The advent of the Analog Devices AD620 low-noise, high-gain operational amplifier thirty years ago meant it was now possible to build very small amplifiers well-suited for physiological applications, such as ECG, EMG, and EEG, with just a handful of components. It wasn’t long until simplistic single-lead ECG devices flooded the market with various incarnations of this hardware, except now it has been updated to include some simplistic measurement tools and filtering. It wasn’t long until someone figured out that you could couple the ECG to Bluetooth and the consumer ECG industry was born.

But ECG capture specifications are not all created alike, and in the hospital, an ECG/EKG means a 12-lead ECG/EKG. A single-lead ECG leaves a lot of data on the table. Early predictive algorithms for medicine included the Early Warning Score (EWS) with its many permutations that were an early attempt to view multiple parameters to identify patterns in five parameters. Next, the Rothman Index added the concept of searching structured parameter data along with non-structured nursing notes which at one time was considered the new frontier of clinical data mining.

In 2005, after leaving Marquette Electronics (now GE Medical), Dr. Richard Crane, (retired) wrote a small piece of software for GE Unity patient monitoring network, a small piece of software that would open up the live UDP packets and extract data including waveforms, which is converted to XML. The GE Carescape connects to the patient monitoring network, converting up to 200 patient beds into 2-second streaming XML in a protocol called high-speed data interface (HSDI). It should be noted that each of the bedside monitors requires its own streaming license (~$1K). Philips has a nearly identical system called PIIC six that outputs a similar system that outputs an XML file (as exampled on the right).

The advantage this provides is the ability to import waveforms in a signed-integer format. ECG waveforms especially provide a great deal of promise for discovery. Heart Rate Variability (HRV) computation is easy to accomplish, as it has been analyzed for approximately 40 years in 24-hour ambulatory monitoring (Holter” monitoring). To determine heart rate variability, first identify the tallest peak on the QRS complex, which is typically the R-wave; measure the time between beats to find the R-R interval. To get a high-resolution HRV report, one may use open-source software such as Kubios. From there, one may use a separate AI algorithm to help identify episodes of atrial fibrillation. Those who have studied the Apple watch AFib algorithm, such as this cardiologist, say that it is fairly accurate in older subjects but that the positive predictive value (PPV)  of the Apple Watch AFib drops to approximately 40% for individuals under the age of fifty. This product has set the entry-level consumer health device expectations and price-point as well as the acceptance of algorithm-interpreted clinical data.

Thus far, everything previously discussed could be filed under “heart-rate variability and sepsis-into-AFib detection” (Yet Another HRV Algorithm (YaHRVa) ) category of predictive analytics, which makes up the majority of today’s FDA-cleared analytics on the market. Beyond HRV, the concept of incorporating other clinical parameters has been implemented successfully at OBS Medical where their “Visensia Safety Index” boasts an 85% positive predictive value for identifying patient deterioration which, for an in-patient, translates to whether a patient will “code” or needs to be resuscitated. The interesting story behind the Visensia Safety Index is that their algorithm came out of an Oxford University study of failure analysis modes in the Rolls-Royce jet engines. OBS Medical translated the algorithm into one that can process five vital sign parameters as inputs. The algorithm uses ECG only for R-wave-triggered heart rate variability and also incorporates other parameters including SPO2, NIBP, and body temperature.

The current state of the art in cardiology AI is still advancing and is limited to single-lead AFib prediction such as the algorithm created by Cardiologs in conjunction with Apple for the Apple watch which performs the AFib algorithm using Edge AI. Many focus on HRV for the the simplicity of having an easy technical target as full continuous data analysis has not been heretofore attained, with or without AI.

Visualization: One of the difficult aspects of creating an algorithm is to concurrently design what needs to be an instantly discernable visualization object, preferably with some values and reasoning statements. These visualizations will often need to go beyond the custom capabilities of most BI programs such as Power BI and offer the ability to create a dynamic 3D objects such as the CoMet score, created by Dr. Randall Mooreman. The CoMet Score is a sepsis prediction tool for the NICU.

Object-Oriented Algorithms: One of the difficulties in using predictive surveillance algorithms is that there are many of them. The patient monitoring system must be programmed to out to a specific, singular instance of an algorithm at a particular port and IP address. One way to encapsulate both the algorithm and the visualization object is Unified Mark-Up Language (UML). UML provides a method of sharing these algorithms and visualization objections together, allowing them to be shared via a community.









  Existing Data Transfer Protocols (for those firmly mired in the minutia):


MFER ISO/TS 11073-92001 – Used by Nihon Kohden in their PM data export systems, otherwise no commercial users.
MQTT – Message Queing Telemetry Transport. More of an IoT transport protocol, but secure.
“X73” or ISO 11073 Use is non-existent probably due to lack of real-time waveform support.
XMPP – Extensible Messaging and presence protocol: “pub/sub”
XML – The current method de rigueur for high-speed data interfacing (HSDI) such as is present on the GE Carescape and Philips patient monitoring interfaces. Current dynamic clinical data transport protocols include: CoMet Score


This Project includes:

1. A headless miniaturized patient monitor to capture diagnostic-quality data and stream to the cloud as well as to an
2. A new algorithm to make use of special parameters that PICSI can provide in an ambulatory setting for general
3. A new transport protocol for high-speed medical IoT data: secure, lightweight, cool acronym (kidding)
4. A new visualization object to coincide with the algorithm.
5. A new use of UML to store both the algorithm and the visualization object making them portable.

The mission is to capture the data parameters and waveforms from the device, store them in SQLite with the other
parameters such as SPO2, NIBP received from BT, encapsulate in a continuous-feed protocol sent securely to the cloud
where they are processed into reports and a visualization object that can be sent back to the device/user.