The Future of Remote Patient Monitoring is in Artificial Intelligence

Introduction

In remote patient monitoring, patients are typically monitored in a home environment using connected measurement devices. The measurements are transmitted from the measurement device to a cloud service, either directly or via a phone or tablet. From the cloud service, healthcare professionals can then review the measurements using a web-based user interface. Wearable devices such as smartwatches are convenient because they can perform measurements automatically, but also non-wearable devices that require patient interaction can be used, such as a blood pressure cuffs or scales. The below picture shows a typical remote patient monitoring system:

Remote patient monitoring systems are not necessarily medical devices. In accordance with MEDDEV 2.1/6 guidance, systems which merely transfer, store and display medical data are not considered to be medical devices in EU. FDA considers them to be Medical Device Data Systems (MDDS) for which FDA applies enforcement discretion, i.e. MDDS are theoretically subject to regulation but it is not enforced, so in practice they are unregulated just like in EU. Of course, additional features like decision support systems can still make them into medical devices. For example, home ECG monitoring typically requires automatic recognition of arrhythmia due to the sheer mass of data from continuous monitoring (tens of hours of home ECG data vs. a few seconds of spot measurement ECG data at a clinic), and the algorithm that recognizes the arrhythmias obviously has to be medically approved.

The Challenge of Remote Patient Monitoring

A basic remote patient monitoring system is not particularly difficult to implement, in particular if you only support low-frequency data like blood pressure or weight that only produce a few measurements per day at most. A couple of gifted university students could probably create a proof-of-concept as a summer project. High-frequency measurements like ECG can produce a million times more data and require paying more attention to the scalability of the architecture. The real difficulty, however, lies in getting the user experience right, and that requires close collaboration with healthcare professionals and patients. New requirements will also emerge at that point: you need to collect more data than just the measurements, e.g. symptoms experienced by the patient, and provide some way of communicating with the patient. Integration to hospital systems will be requested. Still, with some effort all of that is achievable, and it does not require any bleeding edge technology.

Considering the ease of implementation and the relaxed regulatory control, one might expect remote patient monitoring systems to proliferate quickly. And indeed many have been created, and there has been a lot of pilots and experimentation. However, success stories of wide production deployments are still quite rare. There are three reasons for this:

1. Lack of solution,
2. Lack of clinical evidence, and
3. Lack of healthcare professionals.

Lack of Solution

Most companies making remote patient monitoring systems are technology companies, and technologists think in terms of platforms. If healthcare professionals tell an engineer that they want to monitor respiratory volume with a spirometer, the engineer will immediately start thinking about a generic measurement platform that can also be extended to other measurement types such as blood pressure and weight. However, what the healthcare professionals really care about is diagnosing the patient’s asthma, and they want to know what kind of situation the measurement was made in, whether the patient had been smoking, etc. They want a solution that helps them follow the standard care protocol for such patients in the most efficient way possible. Not a platform that enables all possible solutions in theory, but does not provide any specific solution out of the box.

So, often companies offer only a platform, when healthcare professionals are looking for solutions for treating specific medical conditions – and those solutions have to conform to accepted best practice for treating the condition.

Lack of Clinical Evidence

Of course some companies do offer solutions for specific medical conditions. Sometimes the company even specializes in a particular conditon, and the solution may even be medically approved. Such solutions can provide several benefits:

• Improved outcomes: Patients can be monitored more closely since measurements can be done more frequently, or even continuously.
• Improved comfort: Patients can perform measurements at home, during times convenient for them.
• Improved efficiency: The time it takes to process a patient is reduced due to workflow automation, and patient beds are freed since patients can be sent home earlier.

However, in the medical industry, it is not enough to claim that such benefits are obvious. You need to prove them through clinical trials, and that is expensive and slow, which is anathema to most startups. Developing medically certified software is much more expensive than regular software development, and that alone is enough to discourage most startups. Clinical trials, however, make medical device certification look like a bargain. Clinically proven remote monitoring solutions are few and far between.

Moreover, if you can only prove improved comfort or efficiency, you will not be able to charge very much. Only a clinically proven improvement in outcomes will justify a high price. And if you are not able to command a good price, achieving profitability in the relatively fragmented healthcare market is going to be challenging.

Lack of Healthcare Professionals

The most important benefit that remote patient monitoring systems can provide is improved outcomes. But this benefit is mainly due to the ability to monitor the patients more closely than before, and that only helps if a healthcare professional reviews the results and acts of them. Since more data is being generated, more healthcare professionals are needed to review the data, and they are both expensive and in short supply.

Some remote patient monitoring systems only replace a previous manual note-taking system. In those cases, the amount of data remains the same, and consequently no additional monitoring effort by healthcare professionals is needed. Unfortunately, although there may be some improvement in outcomes due to better adherence, interactive patient guidance apps, and faster review turn-around time, the potential for improvement is limited in these systems.

Thus, the potential of remote patient monitoring is limited by the availability of healthcare professionals for reviewing the data. To realize its full potential, we need to automate the review of data.

The Promise of Artificial Intelligence

Even with clinical proof, without artificial intelligence (AI) remote patient monitoring can only provide incremental improvements in efficiency. With AI, we can increase efficiency by an an order of magnitude. With AI, remote patient monitoring has the potential to be a major part of the solution to the problem of skyrocketing healthcare costs caused by an aging population.

AI-based Remote Monitoring

The first step is AI-based remote monitoring: AI handles all patients that do not need attention. The system identifies patients potentially needing attention based on AI analysis of data. Human doctors at least to some degree only review patients identified by the AI, e.g. review interval may be based on AI prediction. A decision support AI system may be used separately to help with diagnosis of such patients.

The same prioritization AI can be used also for inpatients, but it only becomes really valuable with remote monitoring (outpatients) where number of patients can be scaled up easily.

AI-based remote monitoring is actually not particularly controversial, assuming that the AI-based prioritization is sufficiently conservative in its estimation. For example, automatic arrhythmia recognition in ECG is fairly well established technology these days.

AI-based Disease Management

The end goal, however, is AI-based disease management: AI handles all patients that respond to standard treatment as expected. The system automatically provides disease management feedback to the patient, requiring healthcare professionals to intervene only when the AI-based prognosis changes radically or exceeds limits set by the doctor. A decision support AI system may be used separately to help with diagnosis of such patients. For example, the AI might automatically change the medication of the patient, instruct the patient to visit a laboratory for additional tests, or provide the patient with dietary and exercise advice.

AI-based disease management could be highly controversial with both regulators and healthcare professionals, depending on how much freedom the AI has to determine the treatment of the patient. However, it also has the potential to free human doctors from routine disease management work, and allow them to focus their efforts on the patients that do not respond to standard treatment protocols. A doctor would still be needed for diagnosis, but once a patient has been diagnosed, their treatment could be largely automated. This would enable a doctor to manage many times more patients than is currently possible.

The below figure illustrates the difference AI makes in the workload of a Healthcare Professional. Remember that most patients either do not need attention, or will respond to standard treatment:

Future of AI in Remote Patient Monitoring

Solution Components

There are various components of AI-based remote patient monitoring becoming available, though the degree of availability varies, and they each have their own problems:

1. Remote patient monitoring platforms.
   • Availability: Common (although these often also die off quickly, so there may not be many available at a given point in time).
   • Technical difficulty: Easy to do, at least if you only handle low-frequency data and not high-frequency data like raw ECG.
   • Problems: It is an enabler rather than a solution, so the value is limited and you need large customers who are both able to implement their own solutions using the platform and have sufficient scale to make the business worthwhile.
2. Productized, medically approved, medical condition specific customization of a remote patient monitoring platform.
   • Availability: Uncommon.
   • Technical difficulty: Not particularly hard technically, although you need to collaborate with a medical partner to get the functionality and user experience right.
   • Problems:
     • Does not reduce physician workload significantly (but may offer some incremental improvement in workflow efficiency).
     • It is difficult to justify a high price for incremental workflow efficiency improvements, so you need large scale to make it worthwhile, and the market is fragmented so scale is difficult to achieve.
3. Academic AI model that predicts treatment outcomes.
   • Availability: Uncommon becoming common.
     • Number of people studying AI is likely increasing due to both job market demand and government/industrial education programs, so we can expect the academic output in applied AI to increase.
   • Technical difficulty: AI tools are becoming commoditized, so the level of technical difficulty is going down, but data curation remains an issue.
   • Problems: AI development needs lots of data, so access to sufficient amounts of data is becoming the main barrier to entry.
4. Productized, medically approved AI model that provides decision support to healthcare professionals. The AI model may be meant to be used with EHR data, remote monitoring data, or as part of a particular device (e.g. imaging analytics).
   • Availability: Rare becoming uncommon.
   • Technical difficulty: Medical approval of AI is becoming mainstream relatively quickly, with several recent examples of FDA-approved AI (typically via the De Novo route) and AI-specific FDA guidance being developed. On EU side things are less clear, and AI-based medical products may result in a higher risk category.
   • Problems:
     • Need for an extensive (expensive and long) clinical investigation to prove efficacy in various patient populations.
     • Acceptance by the notoriously conservative healthcare professionals, especially if they cannot independently verify the results. Getting recommendation from a recognized medical authority may be needed.

Roadmap

From the point of view of an AI provider, decision support systems based on EHR data are probably more attractive than remote monitoring products, because they are deployed centrally, require no change to existing processes, and apply to entire patient population (as opposed to only outpatients). Likely AI providers will view remote monitoring systems as just another data input.

From the point of view of a remote monitoring system provider, an AI prioritization module is attractive but in-house development is likely not feasible for many of them due to cost of clinical trials. They will therefore likely seek to use commercial, medically approved decision support modules once those become available. This will take some time since such systems are yet rare. Academic systems are already available but the problem of clinical trials remains.

If a remote monitoring system provides both decision support and disease management functionality, then connecting the two is a fairly obvious step to take, although the issue of clinical trial cost will rise again. There is however likely going to be a localization problem: Commercial decision support systems will likely provide diagnosis information and generic recommendations that have to be customized to local treatment protocols. The customization might be applied as a mapping layer on top of the AI results (easy), or it might require the AI model itself to be customized (hard). So this might happen organically, or it might take significant AI development effort.

Thus, likely sequence of events is as follows:

1. Remote monitoring systems are in use but not a huge financial success. Arguably we are in the early stages of this phase already.
2. Medical AI research progresses to the point that clinically approved decision support AIs are commercially available for a variety of diseases.
3. Remote monitoring system providers partner with AI providers to enable scaling up the number of patients and to offer decision support.
4. Remote monitoring system providers with disease management functionality connect the AI decision support to the disease management functionality and first reduce the amount of physician involvement to reviewing and correcting the AI suggestion (i.e. mostly the physician just approves the suggested course of action).
5. The remote monitoring system providers train their AI model to recognize the reliability of the treatment suggestion (how likely the physician is to approve), and enable automatic disease management when confidence is sufficiently high. The AI model can even undergo a clinical investigation in the background using existing customer deployments, by testing how well it predicts the physician’s approval.
6. At first physicians will probably use the AI only for low risk patients and review the AI decisions afterwards.
7. As physicians grow more confident in the system, their use of AI will expand, until they only intervene when the AI-based treatment is not successful.

Conclusion

It is not particularly difficult to develop remote patient monitoring systems, and even certifying them as medical devices is quite feasible. Clinical evidence of efficacy is typically not necessary, as long as the system only optimizes the workflow of a traditional treatment protocol. The problem is that for achieving higher commercial value, clinical evidence of improvement in outcomes is required, and the cost of clinical trials is a significant barrier to entry, especially since the companies developing remote patient monitoring systems are often not used to such costs.

However, in order to realize the full potential of remote patient monitoring systems, clinical evidence by itself is not sufficient. AI is required to free the healthcare professionals from routine tasks and to enable scaling up the number of patients a single doctor can handle. Of course the efficacy of AI needs to be proven through clinical trials, so the required investment is high, but so is the reward. With remote patient monitoring systems evolving into AI-powered disease management systems, we have the opportunity to radically improve the availability and quality of care, and help in solving the healthcare cost crisis.

Comment on this article on Linkedin