Nivedit Majumdar Nivedit Majumdar

Heart Rate Tracking in 2018: Wearables, Neural Networks & mHealth

Surely, there’s an upper limit to the information that can be gleaned from one parameter?

Heart rate data, for all this time, has proven to be beneficial in providing data regarding a person’s blood flow and heart conditions. But then, there’s the work Cardiogram is doing. Through intensive heart rate tracking, which takes into account a neural network built around the heart rate data of millions of Apple Watch users, Cardiogram hopes to be able to diagnose and predict a host of conditions, just from heart rate data.

And that shall be the crux of my article here: the work Cardiogram is doing to detect condition – mainly atrial fibrillation and hypertension – from heart rate.


I’ve already covered how heart rate is tracked using specialised sensors, and there’s a detailed post on the Cardiogram blog talking about how it is actually disadvantageous. But in a nutshell, the Apple Watch uses green LEDs paired with photodiodes, and measures heart rate data using photoplethysmography (PPG).

Image Source: Apple.

According to Cardiogram, however, PPG measures only pulse waves and ignores P waves or the intervals between heart contractions (PR, QS and QT), owing to which there’s a reduction in efficiency in measurement.

Image Source: Cardiogram

Enter, neural networks.

Cardiogram has built a neural network system that taps into the data of Apple Watch users worldwide, and is also conducting research programs that uses the data gleaned from the trained neural networks, to detect specialised heart and body conditions.


Cardiogram states that while atrial fibrillation can affect a lot of people, seldom is it detected in time, owing to symptoms vanishing before the person actually goes to a cardiologist.

The solution for this? Integrating a deep learning mechanism onto devices – such as the Apple Watch – which are actually worn by regular people, and not just people already suffering from heart ailments. And this is exactly what Cardiogram’s DeepHeart architecture is all about.

Heart rate patterns of a person suffering from Atrial Fibrillation (left), and normal rhythm (right). Image Source: Cardiogram.

DeepHeart results a probability number between 0 and 1, indicating the risk for a disease. In the case of atrial fibrillation, any score above 0.9 indicates a confirmed risk factor for fibrillation, while a score under 0.9 indicates a negative test result. Cardiogram has published their detailed studies on their blog, and it’s really quite insightful as far as data crunching in terms of medical applications is concerned.

So, could a regular Apple Watch help in successfully diagnosing and predicting an onset of atrial fibrillation? Cardiogram certainly thinks so, and with a 97% c-statistic, their research is quite ahead of conventional ECG trackers.


A different research was conducted using the DeepHeart architecture, to recognise hypertension and sleep apnea from Apple Watch data. Going by Cardiogram’s blog post on the same, they’ve done so with an accuracy of 82% and 90% respectively, which is mighty impressive for data driven diagnoses.

The statistics in this regard are this: globally, 1.1 billion people suffer from hypertension, and 80% of people with diagnosable sleep apnea don’t actually realise that they are suffering from it. Within the US alone, hypertension and sleep apnea drive $46 billion and $150 billion, respectively, in direct medical spend, lost productivity, and accidents.

But, and I think I’ve never used this phrase before, the numbers are actually boring in this regard. What’s more interesting, is the correlation between heart rate data and hypertension.

Image Source: Cardiogram

Cardiogram states that the link lies in the autonomic nervous system, which links the heart to blood vessels. Two separate studies correlated heart rate variability with increased risks of hypertension and sleep apnea, and these researches, aided with modern day heart rate sensors and AI techniques such as semi-supervised sequential learning, have allowed researchers to gather patterns from data, such as time series of heart rate measurements.

The organisation has listed out their research findings and modus on their blog, so I’ll leave that here.


Million dollar question, maybe quite literally even – given that so many industries are working in this direction.

While data crunching algorithms are definitely becoming more robust, and chances of errors are reducing astronomically, I still think there’s a lot of work that needs to be done. Chances of false positives exist aplenty, as do an occasional misreading of data.

Then there’s also the availability of specialised hardware for the masses, since most people would rather shell out money for a reasonably priced, all-in-one wearable / smart watch, rather than a full blown fitness tracker.

This being said, if companies such as Cardiogram are able to incorporate their algorithms into the gamut of devices in the wearable spectrum in 2018, it would actually pave the way for people to diagnose conditions through the data gathered from their wearables, and then go to registered medical practitioners for further treatment.

Wearables and crunching algorithms can only help in diagnosing and predicting onset of diseases in a quantified mannerism. Treatment? You’ll have to go to the doctor for now.