Página cargando


Save lives by predicting a future heart failure through remote monitoring

promotional image with two elder people and logos of kaleidos, smartcadia and Ramon y Cajal Hospital
Duration:30 months
Budget:€ 0.5M
Client:SmartCardia (Switzerland)
#machinelearning #ehealth #timeseries #ICU #wearable

Save lives by predicting a future heart failure through remote monitoring

Someone has suffered from a cardiac arrest. Their life has been saved but a huge risk to their life still looms ahead. A second fatal cardiac failure might typically occur in the next month. Monitoring 24x7 is key, but it is also logistically cumbersome and very costly.

This was the brutal briefing we got from Srinivasan Murali and Francisco Rincón, SmartCardia co-founders. "It's about reducing the traumatic re-hospitalization rate that we see these days through real-time remote ECG trend analysis".

At the prestigious École polytechnique fédérale de Lausanne, SmartCardia startup had been kept busy developing a unique wearable.... wearable. Patients can carry it on with almost zero hassle. However, they needed a platform that would collect thousands of data points in real time. Not only that, also process them to early detect troubling trends. It was critical to develop this platform and we loved the challenge. We'd develop a Virtual Hospital Intensive Care Unit (ICU).

doctor, nurse and patient in hospital

The challenge

More than 70% of all medical costs come from chronic disease management. This expense could be drastically lowered if hospitalizations and unnecessary use of emergency services were reduced and therefore diverted into other much needed areas in healthcare. More than 60% of hospital emergencies are not real emergencies. A significant percentage of emergencies due to congestive heart failure fall into this category but there is no way we could be telling patients to not come if they fear they might be suffering from such heart failure; The fear of irreversible consequences is very real. So, how can we be smart about this?

The answer lied in bringing wearable technology and machine learning into a medical doctor oriented platform.

SmartCardia had to finish miniaturizing their wearable patch. Making sure it was up to ICU standards while sporting several days of battery life. If a patient was to be sent home with one of these patches, it had to resemble a remote ICU without all the cables and probes, just sticking it onto your breast. At the same time, Kaleidos had to develop all the software needed to monitor dozens of these wearables, process real-time signals, analyse everything using signal processing and time series machine learning technology and show data, charts and graphs on a robust and accurate web platform.

But, could we actually develop a time machine to predict future cardiac failures? The academic research so far was promising, but just that, promising.

hf predict dashboard

Atrial Fibrillation is the answer, Machine Learning is the solution

Heart diseases come in all shapes and colours but there is an event that is very much linked to a big portion of those, Atrial Fibrillation. Detecting Atrial Fibrillation with enough time to assist the patient remotely can boost prognosis and avoid unnecessary hospitalization.

We developed a Machine Learning engine that would be trained through carefully selected annotated data. We employed a total of 110.135 Electrocardiograms (EGC) segments labelled as "normal" and 129.601 ECG segments labelled as Atrial Fibrillation coming from reliable public databases.

Next, we needed to find out how well our Machine Learning model fared compared to medical doctors, while reducing processing time as much as possible. To do so, we went from 200 defining ECG features for our model to just 43 carefully selected. With those, our Machine Learning model was able to almost outperform three medical doctors in several double-blind experiments. And most importantly, do it so in almost real-time instead of waiting minutes or even hours for a proper diagnosis. There were various hospitals involved throughout this applied research but we would like to thank Hospital Ramón y Cajal in Madrid, and the staff at Cardiology Unit.

The winner classifier (related to Sensitivity and Specificity) used ADA Boost based on Random Forest with 100 estimators. It was able to achieve 95% overlap with medical doctors diagnose (which enjoy 100% success rate if we combine two or more of them).

hf predict dashboard

Machine Learning is great but people need a Platform

Building this medical platform was no small feat either. There were various different modules for a complex platform architecture that should be able to withstand heavy hospital loads and recover from lost Internet signal at the patient's side. The main modules were:

  • Wearable patch and patient assignment module: patches were assigned and activated through a secure protocol.
  • Patients profile & clinical history module: information about previous diseases and conditions and patient registration to the platform.
  • Real-time visualization module: authorized doctors could see the current parameters and ECG signals coming from the patch, labelled and annotated by the Machine Learning.
  • Historic data module: authorized doctors could review thousands of ECG segments in the blink of an eye. Multiple time series could be stacked together to easily follow synchronized data points.
  • Multiple patient module: multiple virtual "beds" where shown to monitor up to 20 patients in one big summary view that would highlight those patients with detected AF.
hf predict dashboard animated

Two pieces of work might also be highlighted:
  • The ability to deploy the HF-Predict platform for any hospital in a matter of seconds due to its cloud-based nature. But also, to be able to make use of existing hospital infrastructure.
  • The pixel-perfect ECG drawing mechanism on any device. Since medical doctors heavily rely on the ECG signal morphology to come up with a correct diagnose, the millimetre-paper that has been traditional used to print these signals had to be digitally reconstructed regardless of screen size, resolution or pixel density.

user history in hf predict dashboard

Blood pressure is still an elusive target

Recent academic research had linked some cardiovascular parameters to blood pressure. High blood pressure combined with other trends are clear indicators of imminent heart failure. Since the wearable patch is not invasive and can't measure blood pressure, we thought we might be able to infer blood pressure using machine learning and complex classifiers. Unfortunately, even though we outperformed all other competitors, we fell short of our major goal. More research and work will be needed.

The platform in the real world (as of January 2020)

With 10 clinical studies in various hospitals around the world, totalling more than 3.000 patients, we have been receiving wonderful feedback from the real professionals in the field, expert cardiologists. The Ramon y Cajal Hospital in Madrid, one of the most prestigious in Spain, has already been using the platform to real-time monitor 20 patients. Other pilots in India will add 100 more patients and another 25 with both HF-Predict and Holter. Denmark is the most recent country to join the list with 30 more patients being monitored via HF-Predict.

This means that HF-Predict performs great under real-world stressful situations where many legacy systems are the norm and resistance to change is common.

It's no wonder than the HF-Predict software platform and the wearable patch were awarded the CE marking, which allows full scale commercialization.

The handover

Kaleidos has a clear policy not to continue building a platform much beyond its market launch since we believe companies should have in-house teams for that. SmarCardia was ready to inherit the platform development and has, in fact, added some new features already mostly related to pdf reporting. They recently launched a brand new website and have unveiled their new platform to the world. Their journey continues!

Building HF-Predict was a major challenge but we knew we had to develop such platform. At Kaleidos we are frequently confronted with choices, some of these choices have to do with project selection. You can be assured that if Kaleidos picks your project, you'll enjoy a fully devoted team, overflowing with motivation and skill, but for that to happen, we need to choose and we need to choose wisely. HF-Predict, even with all the unknown unknowns, was one of easiest picks in Kaleidos history and we are very proud of what we delivered.

Go-to-market time for us was critical, Kaleidos delivered on time, amazing committent and work!
Fran Rincón. SmartCardia CTO and Co-founder
Visit HF-Predict Website
We learned that...Hospitals will accept new platforms when backed by sound science, but it is important that they are also akin to established habits and workflows.
One key takeaway was...Working with hardware prototypes, even if close to the final product, leads to frequent unknown unknowns. A software emulator is key.
Do you like what you see?Let’s talk about your project get in touch
Main officeLa Botánica 4, 1º, 28028 Madrid, Spain
Kaleidos Web SVG sprite