No data is no problem!

Speeding up AI integration in remote patient monitoring products

A guide to product and technology design

Neurons Lab
Neurons Lab
Published in
7 min readJun 30, 2021


This story was originally published on the Neurons Lab blog on Jun 30, 2022.

In the inefficient universe of healthcare, with skyrocketing service costs, remote patient monitoring (RPM) looks like a light at the end of the tunnel. RPM like telehealth, telemedicine, virtual care, and digital health is designed to free the healthcare workers from routine tasks and will unlock access to quality care for many more patients than it is right now. Taking into account the aging of our population and the request for a longer healthy lifespan, the use of already available gadgets such as fitness trackers and smartphones provides us with a lot of significant information as a heart rate, and variability, blood pressure, oxygen saturation, physical and social activity, physiological mood, etc. Smartphones allow performing questionnaire surveys, patient video reports, mental and physical exercise tracing [1].

There are ready-made solutions for RPM, but it is highly likely that additional developments and settings will be needed to integrate with the customer’s systems. A customized solution can meet the needs in terms of functionality, but its development is relatively time-consuming and requires significant investment.

Why do we need an RPM?

Illustration from
  • Access to medical care. An RPM solution provides e-visit functionality with the help of wireless medical devices, chatbots, and wearable biosensors to examine patients outside of clinical settings.
  • Data gathering. With RPM, healthcare providers continuously obtain real-time data on the health of their patients. Big medical data is fuel for data science (DS) and machine learning (ML). Today big medical data is a source for new meds and therapies development.
  • Patient’s dynamic. RPM solutions monitor medical parameters and provide up-to-date information about the progression of the disease or injury and the response to prescribed medications. This information can help you customize treatment and avoid medical emergencies.
  • Early detection. Early detection of disease exacerbation with RPM can help reduce emergency department attendance and readmission rates.
  • Patient engagement. Thanks to remote monitoring, patients feel confident in their successful recovery because they understand that their health is monitored around the clock, seven days a week.

In a 2019 study, Fairview Health Services and HealthEast (Minneapolis) did RPM program focused on heart failure showed a 33% reduction in hospitalizations, a 75% reduction in 30-day readmissions, and a 52% reduction in emergency department visits [5]. Using RPM technology, the University of Pittsburgh Medical Center has reduced ER utilization and readmissions by 76% [6]. Due to these reductions, significant cost savings were achieved.

Why does AI integration fail?

However, the collection and visualization of this data is just half of the work done. The real value will come from automated analytics of this data and providing insights about deviation from normal behavior, severity estimation, and decision support. Many product managers have AI embedded into their product as a number-one priority task in their backlog, however, we still do not see the market booming with great and competitive solutions. Why it is so? We have selected top-3 reasons that our customers share with us on the stage of AI product planning:

  • It’s not clear how the AI feature is going to look in the final product
  • It’s hard to sell the vision and the value to the stakeholders, especially to those with a medical background
  • Even if we plan it well, the unavailability of the vast amount of data blocks us from any possible developments.

All this hassle could be easily avoided in the product planning phase, on which your research team is evaluating potential solutions and technical feasibility. In this article, we want to shed light on some of the solutions that can help you to design AI products much more clearly and develop first market-ready solutions faster than your competitors.

Product design

Let’s begin with product design. Most probably you already have a generic software platform in the place that collects questionnaire data and is monitoring some biosignals. One of the reasons why you need any machine learning embedding there — is to add smart functionality. Healthcare professionals want to see more than just nice graphs on their tables. And while your software team is thinking in terms of scalability and generalism of your platforms, your healthcare users, care about the context of every symptom measurement, Anamneses of the patient, and making a treatment decision based on that data. This looks a lot like a standard care protocol and what you need to do — is to map it to the user stories pointing out which pieces should be automated by the AI.

We recommend working through the framework created by Google and their People + AI Research team [2]. It is designed to answer the main questions about the data that your users will provide and receive as the response, to define their success and related metrics, to sketch the interaction scheme, and adjust and control the expectations from the “almighty AI” which is definitely not such one.

The story of BenchSci — a platform for speeding up the drug discovery process might give you a lot of motivation to do what they did — to design the AI product correctly from the very beginning:

Technology design

Now we have a vision and it’s tempting to start development with the state-of-the-art AI algorithms from the latest technical conferences that showed killer performances on the data similar to yours. A class of machine learning algorithms called deep learning indeed can digest any complex data and their combinations including medical images, EHRs, biosignals, drug molecules and produce diagnoses, recommendations, or new insights. However, your technical team will quickly tell you that you need dozens of thousands of data samples to train them, which you most probably don’t have and won’t have any soon. And anyway, you just need to test a feature quickly!

A solution lies in classical approaches to data modeling. We could land on the moon without deep learning, why can’t we monitor wellbeing? While doing the solution research we recommend looking at the not-machine-learning-based solutions as well. It can improve your go-to-market speed at the first stage and make you closer to the actual data collection which will unlock opportunities to use the latest algorithms. What you can look for:

  • Rule-based expert systems (logic, fuzzy logic)
  • Linear models (with already published parameters and coefficients)
  • Differential equations (with already published parameters and coefficients)
  • Pattern-matching models (with published patterns and metrics for them)

For example, let’s assume you need to track patients overnight carefully, but you haven’t collected such data before and open-source datasets aren’t a very good fit here. Most of the tech consultants would just recommend you to wait and collect data, but you, as a product manager, know very well that you won’t be able to push customers to track themselves in the nights for a long time (just for the sake of data collection) without giving them any value. Or it will just be very costly for you. This is a typical “cold start problem”.

Examples of the sleep patterns published in a scientific paper that can be used as a basis for a product feature

Using the idea of mathematical model baselines, you could find the following paper [3] that describes sleep patterns that you can collect from the pulse and accelerometer data. The researchers found the patterns in data for you — now your team needs to implement this knowledge and already have a needed feature implemented with known limitations. Now you can give your customers value and collect data for the smarter ML algorithm development which will provide even more value. Voila! This is how a flywheel of ML works.

Next steps

The stakeholders’ agreement will naturally arise if previous steps were done. Algorithms won’t look risky, the opposite, they will be naturally embedded in the care protocol healthcare professionals are familiar with. The first version will be developed quickly, it will be benchmarked, results easily interpreted, and could be used with real users, who will provide you with the data to train real AI. But that’s for a different story. Now you already can start getting benefits for the regular patients to care automation, allowing up to 40% reduction on the nursing costs [4]. Want to learn more about how to design AI solutions focusing on rapid go-to-market and clear vision? Talk to us!

You can reach us out at or contact Paul, author of this article and Managing Director & Partner, Healthcare Practice directly at





Neurons Lab
Neurons Lab

We co-create disruptive AI solutions with fast-growing companies and empower them to gain a competitive edge. Join our journey here!