Background

Globally, 463 million people have diabetes, 4 million in the United Kingdom [237]. Seventy-seven thousand diabetics in England have foot ulcers today [238]. The longer ulcers progress, the longer they take to heal. Untreated ulcers can lead to diabetic foot disease (DFD) and amputation [239]. Twenty-five lower limb (toe/foot/leg) amputations occur in England daily; 85% are preventable [240, 241].

From 2015 to 2018, 27,465 lower limb diabetes-related amputations happened in England with 147,067 DFD-related hospital admissions and an average 8-day stay totaling 1,826,734 bed days costing >£686 million [242, 243].

DFD accounts for £1 in every £100 the NHS spends [244]. Reduction in ulceration prevalence by 33% would save NHS England £250 million/year [9]. Foot checks are a chance for potential problems to be identified and assessed and action taken. NICE NG19 states people with diabetes should have diabetes foot checks annually and patients with active foot problems referred to a multidisciplinary team within one working day and triaged within another [245].

Innovation can democratize patient care and reduce the economic and health burden of ulceration/DFD.

Cognitive Vision

Cognitive vision or image recognition is a machine learning technique designed to mimic how the human brain functions. With this technique, machine learning models are taught how to recognize visual components that make up an image. Through learning on large datasets of images and identifying patterns, models are able to make sense of their input and determine relevant tags and classifications. Image recognition is not an easy task. To successfully perform image recognition, neural networks are the technique of choice.

However, even still, the computation resource cost is expensive in practice. For instance, a 20-pixel by 20-pixel image received by a neural network would receive over 400 inputs. Although this sounds achievable in a typical hardware setup, images of greater size, say 1,000 pixels by 1,000 pixels, would require a powerful computational resource to process the higher number of parameters and inputs. In a traditional neural network, each pixel would be linked to a single neuron. This is computationally expensive.

Proximity of pixels within images has a strong relationship with their similarity. Convolutional neural networks specifically make use of this feature. Rather than treating two nearby pixels as distinct, convolutional neural networks assume there is more likely a relationship between these pixels than two that are further apart.

Through bypassing less significant connections between pixels, convolution solves the computational and time problems faced by traditional neural networks.

Convolutional neural networks improve the computational resource required for image recognition through filtering relationships by proximity.

Clinical evidence suggests that diagnostic tests and risk stratification can predict the risk of ulceration and amputation, with early referral reducing amputation rates and time to heal [251].

Solutions for diabetic foot care must be broader than the standard healthcare worker-dependent service and be deliverable at a lower cost for the patient and healthcare system.

Together with my colleague Yang Wang, we applied state-of-the-art AI techniques to automatically detect signs of diabetic foot concerns, alert the patient, and support communication and assessment by the healthcare team. Image recognition technology for diabetic foot concerns would be revolutionary, particularly considering this machine learning experience could improve ulceration healing time, cost, and risk of amputations—as this is related to how early ulceration is detected. The model has achieved high accuracy in detecting subtle areas of bruising and cracked heels, which can soon develop into areas of foot ulceration.

A cloud-based system was developed to allow remote use of the application such that any device with an Internet browser was able to upload an image alongside a standardized protocol. The output would be a list of any concerns detected, the referral pathway most suited to the concern, and education, if relevant.

The model is constantly refined through a machine learning feedback loop that optimizes and retrains the model periodically with random subsets of labeled data, including the data received from the web-based system.

Project Aims

The aims of our project are to further develop and validate this machine learning technology, including the deep learning neural network used to predict the risk of foot concerns and provide precise medicine to patients. The convolutional neural networks used for multiple predictions can become more precise through their use and feedback loop from users.

As the user base is largely healthcare professionals and carers, the learning can be scaled up and validated by the healthcare professional for appropriate triage. The project also aims to deliver a service to patients, which is only possible when the model has reached maximum, and ideally near-absolute, accuracy. Reducing error is a key priority for the project.

Grading schemes and standardized protocols support the model to develop greater precision and recall. The ability to identify ulceration and other foot concerns earlier would accelerate diagnosis and treatment and save the lives of many. A cloud-based system would enable easier sharing of the application and its outcomes. A catchall event log was developed to identify all events taking place within the ecosystem. The platform includes behavior health mentoring and integrated tracking for blood glucose, blood pressure, mood, food, weight, and sleep through Bluetooth-enabled devices, wearables, and self-inputted data.

AI models provide boundless opportunity in predictive analytics. By analyzing and generalizing on thousands of images, medical images are being turned into science. There are few ways as engaging and immersive as digital, which enables healthcare to collect enormous amounts of data; analyze and provide useful feedback to inform patients; and prevent, predict, diagnose, and treat disease.

By the very nature of imagery, de-identification for aggregated learning is technically simple. However, the requirement to be able to explain the mechanisms of deep neural network prediction is technically difficult. A more general understanding of the application alleviates most stakeholder concerns, which can be explained through demonstration.

As a novel innovation, there has been a requirement to explain how models are generated, validated, and continually evaluated. Regulatory approval has proved more laborious than previous projects due to the loose definitions that exist around product categorization.

Error minimization has been a priority, and regulatory approval has focused on ensuring patient safety. The risk of false positives was determined to be more likely: early ulceration detection means quicker healing and less risk of amputation.

Thresholds of concerns were developed within the model, with any doubt referred to a human.

Outputs of the project are being documented through clinical papers to highlight the value of the project to patients, healthcare providers, and global healthcare providers. Perhaps of most significance, the outcomes will demonstrate how people with diabetes can achieve significant health improvements by preventing and detecting diabetic foot problems.

Challenges

There were a number of inertia points that were observed throughout the duration of the project:

Conclusions

Cloud-based intelligence that is sensitive and accommodating to differences in photography hardware and formats can be combined with patient clinical, demographic, behavioral, and genomic data to develop a precision medicine framework for detection of diabetic foot concerns.

The prototype was welcomed by healthcare professionals, with several remarking on elements that improved their own knowledge on the topic. Healthcare professionals felt more confident in diagnosing foot concerns when partnered with the digital tool. Digital assistance can confirm and support decisions and challenge healthcare professionals on how they arrive at their decisions. Applied across the spectrum of medical images and data collected by healthcare and research organizations, the technology will affordably enable large-scale prognostic, risk stratification, and best treatment decisions.

From our experience, having more data is almost always more important than the algorithm itself.

Overcoming the concerns of text digitalization and label recognition alongside medical imagery proves to be a challenging area within this field. Handwriting and human error have been the biggest problems to mitigate to reduce false recall. Conducting a large-scale clinical trial is the next phase of the project. A pilot project assessing the feasibility of the project is pending funding approval.

Background

Type 2 diabetes has serious health consequences including blindness, amputation, stroke, and dementia; and its annual global costs are more than $800 billion.

Although typically considered a progressive, nonreversible disease, some researchers and clinicians now argue that type 2 diabetes may be effectively treated with a carbohydrate-reduced diet, which could improve type 2 diabetes management and potentially even lead to remission [252]. Indeed, previous research with carbohydrate-reduced diets for type 2 diabetes does show improved outcomes (such as glycemic control, weight loss, and reductions in the use of hypoglycemic medications), for both very-low-carbohydrate diets (roughly 20% or less of total dietary calories derived from carbohydrates) [253, 254, 255] and lower-carbohydrate diets (roughly 40% or less of total dietary calories derived from carbohydrates) [256, 257]. Although dietary interventions have historically been in person, online programs can be just as effective for some participants, as suggested by research that has examined diet and lifestyle interventions in adults with prediabetes [258]. Therefore, it is perhaps not surprising that the beneficial results of carbohydrate-reduced diets for people with type 2 diabetes (glycemic control, weight loss, and reductions in the use of hypoglycemic medications) have been replicated using online programs [259, 260].

Objectives

Our objective was to evaluate the 1-year outcomes of a digitally delivered Low Carb Program (LCP), a nutritionally focused, ten-session educational intervention for glycemic control and weight loss for adults with type 2 diabetes. The program reinforces carbohydrate restriction using behavioral techniques including goal setting, peer support, and behavioral self-monitoring.

Methods

The study used a quasi-experimental research design comprised of an open-label, single-arm, pre- and post-intervention using a sample of convenience.

From adults with type 2 diabetes who had joined the program and had a complete baseline dataset, we randomly selected participants to be followed for 1 year (N = 1,000; mean age 56.1, SD 15.7, years; 59% [593/1,000] women; mean HbA1c 7.8, SD 2.1, %; mean body weight 89.6, SD 23.1, kg; taking an average of 1.2 diabetes medications).

The Low Carb Program is a completely automated, structured, 10-week health intervention for adults with type 2 diabetes. Participants are given access to nutrition-focused modules, with a new module available each week over the course of 10 weeks. The modules are designed to help participants gradually reduce their total carbohydrate intake to < 130 g per day to meet their self-selected goals. The program encourages participants to make behavior changes based on “action points” or behavior change goals at the end of each module. These goals are supported with resources that are available to download including information sheets, recipes, and suggested food substitution ideas. The Low Carb Program online platform also includes digital tools for submitting self-monitoring and device-driven data on a number of different variables including blood glucose levels, blood pressure, mood, sleep, food intake, and body weight. Weekly automated feedback is provided to users based on their use of the program through email notifications, and participants are notified when the next week’s module has been opened. Lessons are taught through videos, written content, or podcasts of varying lengths.

The program stresses the importance of regular contact with the participants’ healthcare providers for adjustments in medications in weeks 1, 2, and 10. After the 10 weeks’ worth of modules have been opened, participants continue to have access to the education content as well as the ability to continue to track their health (glycemic control, weight).

The content and strategies used in the program build off of prior research and theory. For example, evidence suggests that goal setting can act as an effective behavior change strategy used to improve adherence to lifestyle intervention programs in obesity management programs. The program therefore encourages participants to select a goal at the beginning of the program (such as to lose weight, reduce medication dependency, or make healthier choices for their whole family). Participants are also prompted to consider how their health would benefit from attaining their goal.

Throughout the program, participants are periodically prompted to consider how close they are to attaining their goal. The program further reinforces behavior change through integrated tracking, whereby program users are encouraged to track their health data including mood, food intake, blood glucose levels, weight, sleep, and HbA1c.

According to the control theory of behavior change, monitoring goal progress—that is, evaluating one’s ongoing performance relative to the standard—and responding accordingly is critical to goal attainment. Recent findings suggest that program interventions that elevate the frequency of progress monitoring are likely to induce behavior change.

Results

Of the 1,000 study participants, 708 (70.8%) individuals reported outcomes at 12 months, 672 (67.2%) completed at least 40% of the lessons, and 528 (52.8%) completed all lessons of the program.

Of the 743 participants with a starting HbA1c at or above the type 2 diabetes threshold of 6.5%, 195 (26.2%) reduced their HbA1c to below the threshold while taking no glucose-lowering medications or just metformin.

Of the participants who were taking at least one hypoglycemic medication at baseline, 40.4% (289/714) reduced one or more of these medications. Almost half (46.4%, 464/1,000) of all participants lost at least 5% of their body weight. Overall, glycemic control and weight loss improved, especially for participants who completed all ten modules of the program.

For example, participants with elevated baseline HbA1c (≥ 7.5%) who engaged with all ten weekly modules reduced their HbA1c from 9.2% to 7.1% (P < .001) and lost an average of 6.9% of their body weight (P < .001).

Observations

Conclusions

Especially for participants who fully engage, an online program that teaches a carbohydrate-reduced diet to adults with type 2 diabetes can be effective for glycemic control, weight loss, and reducing hypoglycemic medications.

Background

Epilepsy is a neurological condition of the brain. Different epilepsies are due to many different underlying causes [263]. The causes can be complex and sometimes hard to identify. A person might start having seizures because they have one or more of the following: a genetic tendency, structural changes in the brain, or genetic conditions [263].

Epilepsy is sometimes referred to as a long-term condition, as people often live with it for many years or for life [264]. Although generally epilepsy cannot be “cured,” for most people, seizures can be “controlled” (stopped) so that epilepsy has little or no impact on their lives. Treatment is often about managing seizures in the long term.

Most people with epilepsy take antiepileptic drugs (AEDs) to stop their seizures from happening [264]. However, there can be side effects with such medications, and there are other treatment options for people whose seizures are not controlled by AEDs, including the ketogenic diet (KD) [265].

This case study documents how the organization eliquates machine learning project ideas within teams and how an evidence-based approach is used to support decision-making to solve real-world problems.

Implementing the Evidence Base

The positive effects and therapeutic mechanisms of dietary ketosis and those of a ketogenic diet (KD) on human physiology have been well documented in literature. The KD is a high-fat, low-carbohydrate (usually less than 50 g/day), adequate-protein diet [266], whose metabolic effects originate back in the 1960s; however, the therapeutic effects of KD can be traced back to the early 1920s when it was successfully used in the treatment of epilepsy [267].

Ever since, the potential clinical utility of KDs has been investigated by several studies, resulting in an accumulation of scientific evidence on the therapeutic role of KDs on various physiological disorders. Several studies have shown that the KD does reduce or prevent seizures in many children whose seizures could not be controlled by medications.

Over half of children who go on the diet have at least a 50% reduction in the number of their seizures. Some children, usually 10–15%, even become seizure-free [268]. The Epilepsy Society supports the KD as a treatment option for patients over 12 months old. Research highlights the therapeutic effects of the KD on patients with epilepsy as primarily reduced medication and reduced seizures. In a recent Cochrane systematic review of the evidence regarding the effects of KDs, Levy and Cooper found no randomized controlled trials (RCTs) [269].

This demonstrates that epilepsy can be managed; and in some cases, patients can live seizure-free lives through the sustained application of a ketogenic way of eating [270]. Through reducing sugar in the diet and at the same time improving blood glucose control, people with epilepsy can achieve significant health benefits such as reducing their risk of seizures and number of medications [271].

Sensor-Driven Digital Program

The Ketogenic Program for Epilepsy (KPE) is a structured education behavior change program for people with epilepsy. The KPE empowers users to sustainably adopt a lower-carbohydrate lifestyle with the appropriate personalized education, health tracking facilities, support, health mentoring, and resources to maintain a safe and sustainable lifestyle. As a result, people with epilepsy could expect to improve glycemic control and as a result reduce the incidence of seizures, reduce medication dependency, and improve confidence in managing their condition.

Health data is collected in real time through blood glucose monitoring, medication monitoring, and seizure tracking modules within the application. Wearables are used to collect data to sense seizures and falling. Unstructured data is particularly useful in understanding sentiment and the patient’s psychology.

Research activity takes place over the course of 12 months, reporting on 3-month, 6-month, 9-month, and then 1-year epidemiological health outcomes and engagement data followed by an 18-month and 2-year follow-up to demonstrate efficacy and adherence to the program over the short to medium term.

This is the world’s largest and longest study on patients engaged with a digital platform for epilepsy.

Research

As part of this project, we seek to learn the most efficient and effective methods to implement and accelerate at-scale delivery of our technology to the United Kingdom’s NHS population and international bill payers. The project is led by an award-winning and experienced team, recognized for innovating digital health, who are determined to revolutionize the health and well-being of people across the globe.

The technological challenge is to create a scalable, engaging, and effective solution for global implementation. This challenge is mitigated through the project extending on the infrastructure of DDM’s Low Carb Program, with clinically validated outcomes and 71% retention at 1 year. At the time of writing, out of 300,000 patients globally registered within the Low Carb Program for type 2 diabetes, there are 3,112 people with epilepsy (0.5% of the UK population)—showing the readiness for an online, nutrition-focused intervention for this population.

The potential cost-saving impact to health bill payers is significant. Patients who are empowered and have improved health and well-being are also more active members of their communities and have lower social care needs.

Project Impact

The greatest impact will be felt by patients with epilepsy and their families through empowering patients to manage their epilepsy. Evidence demonstrates that a KD in people with epilepsy can be maintained by infants and adults with improvements in number of seizures and medication consumed. Evidence also demonstrates positive effects on cardiovascular disease risk [272]. The health improvements and potential cost savings are sizable.

In addition to health improvements, the project is of tremendous value to the UK economy and the NHS, with reduced complications, improved mental and emotional health, reduced hospital admissions, and reduced medication having a direct impact on the budget of the UK Treasury.

Positive social impacts experienced by the patient have a repercussion for wider communities, businesses, and the UK economy.

As the result of improved patient and population health, organizations will benefit from fewer sick days and reduced absenteeism, improved mental health, and improved perceived quality of life, as well as a reduction in the perceived burden of managing epilepsy. The positive impact is also received by the clinical healthcare community—through reducing physician and healthcare professional burden and enhancing the evidence base.

Preliminary Analysis

The objective of the research arm of this project is to evaluate the 1-year outcomes of a digitally delivered Ketogenic Program: a nutritionally focused, 16-session educational intervention for epilepsy control through provision of a KD.

The program assists patients to achieve ketosis using behavioral techniques including goal setting, peer support, and behavioral self-monitoring. Interesting correlations between elevated blood glucose and mood enabled predictive algorithms to be developed to identify symptoms of impending seizures, alerting patients through notifications to take action.

Especially for participants who fully engage, an online program that teaches a carbohydrate-reduced diet to adults with epilepsy can be effective for glycemic control, weight loss, and reducing medication and the number of patient seizures.

A randomized controlled trial is required to understand the clinical impact of such a digital therapeutic.

Background

Junior doctors arrive at a university eager to help mankind, yet often they are unaware of how to deal with high-pressure, critical scenarios. Dealing with medical concerns appropriately—whether emergencies or not—requires years of experience and learning.

Over time, exposure in doctor surgeries, hospitals, and clinics develops the self-assurance, readiness, and experience required for junior doctors to realize their vocation. High-pressure scenarios, such as a patient having a cardiac arrest, require pinpoint precision accuracy in decision-making and actions.

For some time after the first arrival to a university, junior doctors are not ready to make such decisions. This is where augmented and virtual reality provides an opportunity to engage students in learning that readies them for life as a doctor. When you are a doctor, every decision has consequences—and that can be burdensome for some.

What’s more, with 1,500 students in any one year, group trips to hospitals are expensive and logistically difficult to manage.

The project was recognized for its pioneering nature, winning an Association for the Study of Medical Education (ASME) Education Innovation Award.

Aims

Project Description

It was agreed that the project has as wide a reach as possible. This eliminated virtual reality hardware such as the HoloLens or Oculus and focused the delivery onto mobile devices. The Samsung Gear and Google Pixel were able to facilitate the delivery of the content with the use of a headset to extend the view into virtual reality. This was logical considering almost every student owns a phone, whereas the use of HoloLens and Oculus is still relatively niche.

Simulations of a variety of clinical experiences were filmed with 360-degree videography using both doctors and patients as actors. These simulations were short 5- to 10-minute videos that featured decision moments where students needed to make an active choice—including a patient suffering from cardiac arrest and a patient with schizophrenia becoming increasingly agitated.

These videos were integrated into a virtual reality application that enabled students to explore a clinical setting in virtual reality and engage in the environment and encouraged them to make a variety of decisions during the process. The application encouraged users to “wear” the headset for a full, immersive experience.

The use of this augmented reality application to educate junior doctors was then evaluated with students.

What Is 360° Video?

360° video is a recent technology that uses omnidirectional cameras to capture a spherical video space rather than the landscape dimensions of traditional photography. These videos are slotted together to create an immersive viewing experience, placing the viewer in a scene within an environment rather than an observer or fly on the wall. The viewer has the ability to explore the scene, turning a full 360°. Capturing 360° video requires specialist cameras.

Playback of 360° video is most often experienced on mobile devices, where the smartphone display acts as the window to the 360° environment. Viewers can physically move their phone to view other areas of the 360° video—controlling both the orientation and viewing.

Physical devices which allow the viewer’s device to be worn as a headset can be obtained to provide a fully immersive experience. Desktop computers can also play 360° video, in a similar manner to mobile devices. Interaction with 360° video on desktop usually involves a mouse to move through the environment. Hardware-driven headsets can also be used to engage with 360° videos; however, this was deemed to be prohibitive given the objectives of the project.

Conclusions

Immersive education is a route of delivery that should be further explored to improve students’ learning experiences, reduce costs, and further innovate healthcare.

Background

The United Kingdom, like many developed countries, is experiencing increasing rates of chronic conditions associated with diet and lifestyle choices such as obesity, type 2 diabetes, and cardiovascular disease [273]. Obesity has been shown to be a COVID-19 risk factor [274]. Evidence to support adoption of healthy lifestyles in the prevention and management of these and other long-term conditions is strong [275]. Less than one in ten patients newly diagnosed with type 2 diabetes attends structured education in the United Kingdom [276]. There are a number of reasons for this, including the fact they are in the working week and often inconveniently timed [277]. Over and above this, learned behaviors diminish after 6 months of attending an offline education program [278]. Improvements in management of type 2 diabetes can be achieved by reduction in carbohydrates, and this approach has been demonstrated to be the most effective approach to achieving normal blood glucose control and sustainable weight loss. Weight loss is a key factor in improving self-management of obesity and metabolic health conditions such as prediabetes [279–281].

Patient-facing digital health applications (apps) provide the opportunity to change the way individuals take responsibility for their own health by enabling more effective delivery of health information, allowing better monitoring of health markers, and encouraging lifestyle behavior change. Wearables (and wearable apps) have been demonstrated to increase activity in older adults and can contribute to weight loss [282–284].

Enthusiasm about the potential of health apps has grown rapidly over the last few years with many NHS services now commissioning digital health apps. While health apps offer the potential to augment care for a breadth of health conditions from mental health to medical conditions like hypertension, cardiovascular disease, and type 2 diabetes, in comparison with uptake, surprisingly little is known about their functionality or impact on health outcomes.

From the data that is available, there is a suggestion that some factors mainly predictive of uptake are not necessarily the same as those predicting completion or health outcome and very few health apps demonstrate long-term adherence [285] or report long-term health outcomes. This may, in part, be due to the lack of standards for collecting data across health apps [286].

Aspects of digital health apps that are demonstrated to impact behaviors and health outcomes should be evaluated to improve health service design and delivery. Research is needed to extend previous health app studies and understand the effectiveness of particular app features on engagement and health outcomes. Previous publications on Low Carb Program, a structured education and behavior change app, demonstrate its clinical effectiveness for supporting obesity management including statistically significant weight loss [287].

Patients’ behavior directly contributes to their treatment success, with doctors relying on patients to take their prescribed medication alongside making and maintaining dietary and lifestyle changes. Many of the most significant challenges in healthcare, specifically in long-term or chronic conditions, such as obesity and type 2 diabetes, will only be resolved if we can influence behavior and support sustainable behavior change among other things such as the food environment.

There has been a detailed review of the features of Low Carb Program with an analysis of the theoretical underpinning for the presence of each app feature. The aim of this study was to assess some of these app features and understand how the engagement with these features impacts health outcomes, in particular, weight loss.

Obesity is associated with significantly worse cardiovascular risk factors, suggesting that more active interventions to control weight gain would be appropriate to help address the increasing burden of obesity on the NHS. Regardless of the interventions used to lose weight—pharmacological or behavioral—the weight is commonly regained [288–290].

Typically, half the weight lost is regained in the first year. Weight regain often continues up to 3–5 years after treatment; and, on average, 80% of people return to or exceed their pre-intervention weight [291].

Similarly, relapse rates are high for individuals who initiate attempts to stop smoking and those who try to reduce alcohol consumption [292–294].

Therefore, effective interventions that consider known factors associated not only with initial weight loss but also critically with weight loss maintenance such as building on internal motivations to lose weight, establishing social support mechanisms, identifying coping strategies, or providing support for self-efficacy and autonomy can all enhance weight loss maintenance, which is crucial for the long-term success of any weight loss interventions [295].

Methods

Low Carb Program

Low Carb Program is a behavior change app for patients with type 2 diabetes, prediabetes, and obesity. The app provides NHS-approved structured education for patients with type 2 diabetes, prediabetes, obesity, metabolic syndrome, polycystic ovarian syndrome (PCOS), and non-alcoholic fatty liver disease (NAFLD). The program educates patients to reduce the sugar in the diet to 130 g/carbs per day over a core 12-week implementation phase.

The platform provides users the ability to track their health (weight, HbA1c, blood glucose, food, mood, medication, ketones, cholesterol, blood pressure), track their nutrition with a food diary, and connect the app with a wearable (Apple Health, Google Fit, Fitbit, Nokia, Garmin) and prompts users to check in at regular intervals. The app is used within the NHS and has been demonstrated to facilitate weight loss in patients with obesity, prediabetes, and type 2 diabetes [287].

The platform provides native-language education and culturally relevant support for South Asians (Punjabi) and is available on responsive Web, iOS, Android, Apple Watch, Android Watch, Alexa, and VR (Oculus).

The platform is built on a peer-reviewed behavior change architecture developed to support long-term behavior change within the app. Components include goal setting, tailored education, community support, integrated tracking, tailored resources, coaching, and AI-powered engagement based on behaviors and health status. Figure 10-1 visualizes the behavior change architecture.

Wearable App: Low Carb Program for Apple Watch and Wear OS Watch

The Low Carb Program’s wearable app is available on Apple Watch and Android Watches, also known as Wear OS. The companion Low Carb Program watch app provides wearers with four functionalities.

• Health tracking: Companion app users can track their weight, blood glucose, mood, activity, steps, blood pressure, and heart rate.

• Relax: Participants can choose a set duration of time to focus on breathing. The smartwatch displays an animation intended to soothe and displays instructions on when to inhale and exhale.

• Lifestyle: Participants can read education and send articles to their phone from a daily-updated library of articles.

• Motivation: A screen within the companion app that displays random motivational quotes curated to prompt reflection and goal reappraisal.

Results

Discussion

This was not a randomized controlled experimental trial, so we cannot compare the 6-month results to a control or standard-of-care group. However, of 200 patients with obesity referred to the platform, 84% of those who used a wearable were still engaged at the 6-month follow-up which supports previously reported outcomes of the effectiveness of Low Carb Program.

Although the study design does not support inferences on the causality, there are statistically significant correlations between use of the wearable and reduction of body weight.

The study found that members who either complete (>9 modules) or partially complete (>4 modules) the program report greater weight loss than those who do not complete the program.

Members who complete the program lose a significant 7.38% of body weight at 6 months. Building on prior evidence for the use of the platform by patients with type 2 diabetes to achieve improvements in glycemic control and sustained weight loss, the results of this study go to demonstrate that for participants who fully engage, an automated online program delivered by smartphone and companion app teaching a carbohydrate-reduced diet and providing tailored on-demand workouts to adults with obesity can facilitate weight loss.

Conclusions

The aim of this study was to evaluate whether a companion wearable app contributes to weight loss. Participants tracking their health using a companion wearable app demonstrate significant weight loss at 6 months, as well as increased levels of activity.

The findings support previous studies that demonstrate wearable apps can improve health outcomes. Further analysis should be done to explore differences among populations, in particular if different ages, genders, ethnicities, and socioeconomic statuses impact health app engagements and subsequent health outcomes.

Background

The variety and velocity of patient data has exponentially increased over recent decades. This data has not progressed the robustness of decision-making. This leads to questions as to why decision-making is not becoming more data driven and what can be done to validate and expedite its adoption.

Diabetes.co.uk is the world’s largest diabetes community, based at the University of Warwick, and the platform welcomes over 45 million visitors a year. As the world’s diabetes community, the organization provides a range of services that focus on the patient data ecosystem to empower patients to manage and control their health.

By collecting patient data, or real-world evidence, over time, the ecosystem is created such that it offers insights, education, and feedback on particular aspects of the patient’s health.

Platform Services

The platform collects a wealth of data. From the moment of platform access, behavioral and engagement metrics are collected, with patients progressing onto registering for the platform, opting in to part with over 120 variables that cover their diabetes and overall health status.

Once inside the platform, a variety of data types are collected—unstructured data in the form of conversations, interactions, engagements, and behavior and structured data in the form of self-reported, device-driven, wearable, IoT, and clinical health record data. The platform provides a number of services to patients and bill payers.

Medication Adherence, Efficacy, and Burden

Medication adherence, adverse effects, and wastage are concerns that pharmacology seeks to overcome. Patients with diabetes are typically on at least one medication on entry to the platform.

Within the platform, patients evaluate their medications at regular intervals, sharing their own opinions on the use of medicine in the real world, and report more structured data such as side effects, adherence, efficacy, and burden of using medications. This data is mapped geographically, by condition, medication, and other markers, and fed back in a de-identified, aggregate manner to pharmacological and academic partners for research purposes.

Data is used by partners to identify new drugs and understand interactions, usage, adherence, and real-world adverse effect prevalence. The same data is also used to improve patient safety. For instance, the platform alerts members on particular medications or using particular devices should there be an issue or recall.

Big Data: Pooling People to Empower Health Decisions

The Diabetes.co.uk community forum is the world’s largest support community for people with diabetes and provides an unrivaled insight into the global diabetes population. The platform welcomes people with all types of diabetes, carers, and healthcare professionals. The concept is simple: real people with diabetes talking about their own experiences and speaking to others who are in a similar position. Patient engagement is high; there is over 2 million years of cumulative experience among members, who spend over 17 minutes each visit on the platform.

The impact of participation in online health communities on well-being has been mainly studied in relation to the informational and emotional support that members receive from their community and the positive impact that such support has on their health condition [298, 299].

Community members provide and receive informational support by sharing practical information that can help them manage their own health condition. Emotional support involves community members seeking mutual understanding and comfort by sharing their story about how it can be frustrating and painful to live with a health condition [300].

Equally, not everyone is positive about health communities. There has been a long-standing debate among healthcare professionals regarding the quality of information shared online, as well as unintended consequences of health self-management, particularly among those with low health literacy [301, 302].

As a novel digital community, data and respective insights from the online forum and dissemination of academic findings are constantly shared within the Diabetes.co.uk patient support network, organizational medical advisory board, partners, and other interested networks.

Bernardi and Wu developed a model to investigate a) the factors that influence the level of engagement in online communities and b) the impact of engagement in online communities on members’ cognitive, behavioral, and emotional experiences of health self-management. The model was tested empirically through an online survey with the users of the Diabetes.co.uk forum [303].

The study aimed to provide a holistic perspective on the impact of a properly managed online health community on patients’ well-being.

AI Prioritization of Patient Interactions

It was decided that to meet our obligations, the use of AI bots would speed up the process, enabling humans to spend time with the concerns that mattered. This was realized through the development of a neural network–based classifier to determine sentiment and concern based on contents, user profiles, and frequency of other key metrics.

This information also assisted the customer support teams to efficiently deal with user queries by knowing exactly what a conversation was about, typical responses to such a query, and what to say to a user to ensure a favorable outcome for all parties involved. Additionally, by gauging the sentiment of users and storing it, user data can be implemented to build additional systems in the future.

Real-World Evidence

Through the analysis of patient data, we can determine a discrepancy between official and real-world health statistics. For instance, 14% of people with diabetes in the United Kingdom are expected to have a mental or emotional health concern, whereby the platform sees 44% of users with this health condition [304]. This highlights that further exploration and understanding of the patient population is severely needed to ensure critical aspects of patient care are not neglected. Partners have become more interested in real-world evidence as the years have progressed, partially to do with concerns over research funding and conflicts of interest.

Ethical Implications of Predictive Analytics

Predictive analytics enables intelligent classifiers to predict the risk of disease. The collection of a varied patient dataset—including conversations, health markers, demographics, and behavior—enables the development of a sophisticated model that can be used to refine predictions further. One of the unintended consequences of the project was discovered when deploying an AI model that predicted the risk of a patient having pancreatic cancer.

Pancreatic cancer typically affects 1 in 10,000 people and is frequently misdiagnosed. Often, patients receive a diagnosis when it’s too late. A pancreatic cancer AI model was developed using a clinical dataset of health metrics, engagement, and behavior data and first predicted the likelihood of poor patient metabolic health. Should patients meet the threshold for poor metabolic health, the model evaluates the risk of the patient having pancreatic cancer.

Communicating this information to users was a first-time problem for the organization. All relevant stakeholders, including all members of the Medical Advisory Panel, were prompted for their input on this topic. Predictive analytics had previously been used to inform patients of possible hypertension or increased risk of conditions based on lifestyle.

However, informing patients of their risk of pancreatic cancer was considered to be unethical, particularly as we did not want to create additional mental health concerns or anxieties.

How should a user be told of a possible life-changing concern? The key action for this communication is to nudge the patient to see their doctor or physician as soon as possible. The factor for success in this case was determined to be user friction. The less friction there is between the user and communicating party, the more successful the engagement.

Stakeholders from behavioral economics were vital in testing and confirming an engaging communication to inform the user that their data fell outside expected norms without raising fear. To do this, the prediction was framed against patient and population demographic and medical data. Rather than being told they were the only patient to see anomalous data, patients were reassured that others have also seen similar patterns and been successful in improving their health.

Integration of the IoT

Sensors in wearables such as accelerometers, altimeters, and heart rate and skin temperature monitors provide a tremendous amount of useful information. For instance, motion, sweat molecules, and conversations are all useful and relevant data points. Such metrics can be used to take a snapshot of metabolic health, cardiovascular health, and mental health. For this to be conducted at scale, there must be little to no friction between the patient and providing parties.

The less friction there is between the user and whom they choose to share data with, the more successful the product. Wearables are becoming increasingly used in incentivized wellness propositions. Validating patient behaviors and general physiology facilitates the personalization of premiums and outcomes-based reimbursements.

Patients have the option to connect their devices and applications to the Diabetes.co.uk architecture. What is interesting is that younger users typically adhere to wearable usage and find it easy, while elderly users typically do not. This is of relevance to insurance and incentivized wellness propositions, as most claims come from older policyholders.

Conclusions

The realization of precision medicine is not without technical and ethical challenges. IoT and conversational and wearable data are just examples of a variety of data that can be collected and used to empower patients in real time, facilitating improvements in health and well-being.

With data and the application of AI come ethical concerns, particularly in the prediction of future disease risk. The lack of processes and governance on such topics is urgently required as the abilities of AI mature.

Background

In order to provide personalized healthcare benefits and services to our users, our AI attempts to learn and interpret the user’s activity of daily living. This includes the behavioral pattern relating to their movement across the home, regular routines such as bathroom habits, time they wake up and go to sleep, and so on.

This forms the key foundation for all the predictive analysis that is carried out to create situational awareness and insights about the person living alone at home and the detection of abnormal activities that would warrant an alert to be raised.

In the previous trial where we achieved accuracies above 90% in identifying abnormal behavior based on the “states” and “transitions,” we had briefly discussed the potential of applying the Bayesian learning framework to automate user behavior prediction and enable proactive monitoring capabilities. In this paper, we define the problem statement for this application and outline the Bayesian approach taken to assess the effectiveness of a new predictive algorithm to add in our library.

Our initial assessment of the algorithm based on the data from a selected group of users shows a prediction of 82.2% accuracy, as an average. The accuracy relates to 10-minute windows within which the system is required to predict the location of the person. However, if the window is increased beyond 30 minutes, the accuracy reaches a maximum of 93.75%.

As the study progresses and more data about the subject is collected and the model trained on a wider dataset, we anticipate the accuracy of the 10-minute window to also increase. Some efforts are also being spent on the development of another predictive solution which uses state models for critical-time predictions without online connectivity.

Problem Statement

miiCUBE network collects movement data in the form of sensor triggers. For every user’s home, the network comprises specialized sensors that detect motion installed in each room and passageways.

Every time the user passes by a sensor, the network registers the activity, with the timestamp of the event and the location where the sensor was triggered.

Expectation-Based Modeling Approach

This section outlines how we define the concept of a prediction window and train a Bayesian model for every such window to obtain a probability distribution of where the user is expected to be. This approach also allowed us to find events which are “unexpected” with respect to the model and continuously update our model with new data that is collected every day.

Multinomial–Dirichlet Models

Consider a prediction window of 15 minutes between 9 AM and 9:15 AM in the morning. If our user generally wakes up at around 9 AM, then we can expect sensor triggers in the Bedroom and the Bathroom. Perhaps, after the user has visited the Bathroom, they go to the Kitchen to make a cup of tea or coffee. So there can be Kitchen sensor triggers as well. If we are only monitoring for these three locations, we can visualize the distribution of the number of sensor triggers observed as shown in Figure 10-2.

This distribution of sensor trigger counts is an example of a multinomial distribution. Since the distribution values are sensor trigger counts, they directly translate into probability values when normalized.

For a prediction window of a day ?, we defined a multinomial distribution (???????? = ?) which tells us the probability of observing sensor triggers at a location ?. From Figure 10-2, we can say that

• (???????? = ???????) = 20/(20+10+12) = 10/21 ~ 47.6% between 9 AM and 9:15 AM.

The Dirichlet distribution is a powerful addition to a multinomial distribution: it serves to embed our prior personal knowledge about where we expect the user is at any point in the day. In the Bayesian framework, we always start with a prior knowledge of the probability of the event in question happening.

For instance, from our example in Figure 10-2, we know that the user will be in the Bedroom between 9 and 9:15 AM because they wake up at this time. We formalize this knowledge by defining an initial informative value for the parameter ? where ?_??????? > 1 and ?_{??ℎ?? ?????} = 1.

The value of ? is defined by the Dirichlet distribution (?) as

???(?) = {?? } ??? ? ∈ {??? ?????? ?????????}

and it represents concentration power, alike to weightage. ?_? > 1 represents a higher concentration allocation to the probability of the user being in location ?, while ?? = 1 represents an uninformative unbiased concentration, that is, the state of not knowing any special knowledge about that particular event happening.

Using the Dirichlet distribution allows us to specify certain known characteristics of the user and therefore add behavior personalization to the model.

For prediction window ?, our multinomial–Dirichlet distribution (?)(???????? = ?) tells the expected probability of the user being in a location ? by virtue of triggering the sensor at that location, given an initial prior knowledge. This definition forms the basis of our expectation-based predictive model ??:

?? = ????(??)(???????? = ?)

where

1. 1.

   r ∈ {rooms where sensors are installed}

    
2. 2.

   t ∈ {sequence of n-minute time blocks in a 24-hour day)

    
3. 3.

   n = Duration of one time block in minutes

    
4. 4.

   α_t = Concentration parameter for the Dirichlet distribution for the predictive window t

    

Model Evaluation

In this section, we evaluate the performance of our predictive model. We use six different durations of the prediction window so as to provide a comparison of variations in the predictive accuracy because of it.

Please do note that uniquely identifiable details about the clients have not been disclosed for privacy reasons.

For each of the evaluation table, 1) each row represents a day in the first week of April 2020, 2) each column heading represents a prediction window duration chosen for the trial, and 3) each cell value is the “hit rate,” which is computed as the number of times the model has a good prediction for a day.

We also expected a negative trend in the accuracy as we predicted further into the future as captured in Figure 10-3

. The shaded region highlights the range of accuracy difference for using windows of 10 and 90 minutes.

Similar to User A, a negative trend is also observed for User B, as shown in Figure 10-4

. The shaded region highlights the range of accuracy difference for using windows of 10 and 90 minutes.

Improving the Solution

In our trial for the two users, we set the value of α for all predictive windows as 1. This represents an uninformed unbiased state of the model before it has seen any data. Instead, we can manually define an informed biased value for α which reflects our knowledge of the user's presence during different points of the day. This can be a tedious manual task, but once defined can reduce the amount of data required for the model to fully condition itself to arrive at the real and precise expectation probabilities.

Increasing the duration of the predictive window boosts predictive accuracy; however, our objective is to enhance the model to bring the accuracy to exceed 90% prediction for 30-minute windows. Achieving such a high accuracy is very challenging given an elderly person may not necessarily stick to daily routine. However, with more data points, the predictive model will be able to learn about certain external factors which trigger an elderly person to deviate from their routines, for example, sickness, the effects of medications, and so on.

Put within the context of the intended use within a care home or a home setting, the model will predict where a person is “supposed” to be, and then the algorithm will compare this with where the person is for that particular window. In most cases, having a mismatch for one or two cycles of 30-minute duration is not going to be detrimental.

However, if, for example, we predict that the user will be in the kitchen in the next 30 minutes (first cycle) and then in the living room for another 30 minutes (second cycle), but the system detects that they are in the bathroom for more than 30 minutes, then an alarm will be triggered. At the same time, the system monitors their body vitals, and if something abnormal is observed, these will also be notified to a family member/carer.

Anything shorter than 30 minutes will very likely result in excessive false alerts given the behavioral pattern of an elderly tends to vary a lot. That said, anything abnormal with the body vitals is notified instantly, for example, low heart rate.

The two focus areas as we grow the fleet of miiCUBEs are user duration and data flows.

Conclusion

This approach has three operating parameters: the training period, the duration of the predictive window, and prior values chosen for α for every window of the day.

The approach was able to achieve an average predictive accuracy of 82.6% in a trial involving two of our miiCUBE users.

For fair comparison, the training and testing periods along with the duration of the prediction window and values for α for all windows in a day were kept constant for both cases. The trial has been rolled out to all of our users.

Background

Regular and moderate physical activity practice provides many physiological benefits. It reduces the risk of disease outcomes and is the basis for proper rehabilitation after a severe disease [305]. Most importantly, regular activity can improve quality of life [306]. Physical activity has been demonstrated to increase fitness, physical function, cognitive function, and positive behavior in people with dementia and related cognitive impairments [307]. One study found that gaming systems are motivating and can have a positive effect on strength, balance, and overall fitness for elderly populations [308].

Patient-facing digital health applications (apps) provide the opportunity to change the way individuals, in particular the elderly, take responsibility and care for their own health by enabling more effective delivery of health information. New platforms and devices allow better monitoring of health markers and can act as tools to support and encourage lifestyle behavior change. Apps and wearables have been demonstrated to empower older patients to increase activity and can contribute to weight loss [309].

Objective

The aim of this study was to examine the profile characteristics and activity outcomes for participants using a conversational agent as a health management tool as part of their lifestyle.

Methods

A total of 27 patients with obesity, prediabetes, and type 2 diabetes were referred to Low Carb Program by their NHS healthcare professional in February 2020. Thirteen patients also downloaded the platform’s Amazon Alexa Skill.

All participants were from England, United Kingdom. Health conditions that would exclude healthcare professional referral were carnitine deficiency (primary), carnitine palmitoyltransferase (CPT) I or II deficiency, carnitine translocase deficiency, beta-oxidation defects, medium-chain acyl dehydrogenase deficiency (MCAD), long-chain acyl dehydrogenase deficiency (LCAD), short-chain acyl dehydrogenase deficiency (SCAD), long-chain 3-hydroxyacyl-CoA deficiency, medium-chain 3-hydroxyacyl-CoA deficiency, pyruvate carboxylase deficiency, acute porphyria, history of mental health disorders, or pregnancy.

All users were selected for analysis. Patient health and engagement data from use of the mobile and companion apps were tracked, collected, and analyzed over 4 months to determine if use of the Alexa Skill contributed to change in health-promoting behavior.

Data analyzed included demographic data (age, gender, ethnicity, location, employment status), weight, and amount of minutes spent in on-demand classes. Qualitative feedback on features was also assessed.

Results

For the 13 participants syncing their account with the Alexa Skill (requiring an Amazon account), 92% (12/13) were engaged and had outcomes at 4 months defined as actively inputting an item of data within the last 7 days.

Discussion

This was not a randomized controlled experimental trial, so we cannot compare the 4-month results to a control or standard-of-care group. However, of 13 participants using the Low Carb Program Alexa Skill, 92.8% were still engaged at the 4-month follow-up which supports previously reported outcomes of the engagement of Low Carb Program [91].

Conclusion

The aim of this study was to assess whether use of the Low Carb Program Alexa Skill would contribute to improvements in positive health behaviors. Participants using the Alexa Skill demonstrated improvements in total active minutes.

The findings support previous studies that demonstrate voice-activated assistant apps can improve health outcomes. Participants who use the Alexa Skill were motivated to try innovative technologies, older in age, and desire training to overcome low comfort levels. Further analysis should be done to explore differences among populations, in particular if different ages, genders, ethnicities, and socioeconomic statuses impact engagement and subsequent health outcomes.