HealthBeacon
Nominated Award: Best Use of AI in Sector
Website of Company: https://healthbeacon.com/
Patients with chronic conditions often self-administer their injections at home. In its 2003 report on medication adherence, the World Health Organization (WHO) quoted the statement by Haynes et al that as many as 1-in-2 patients fail to adhere to their medication schedule. The effects of poor medication adherence are well documented in both the clinical trial and real-world setting, contributing to treatment failures with resultant increases in hospitalizations and healthcare costs. Staying on track with medication for chronic conditions is a challenge and there is inadequate data available to accurately measure the adherence to the medication.
HealthBeacon is a medication adherence technology company which develops smart tools for managing medication. HealthBeacon’s Injection Care Management System (ICMS) provides an innovative alternative to traditional methods by accurately measuring patient’s adherence. HealthBeacon’s FDA cleared smart sharps bin tracks patient injection history, provides personalized interactive reminders and safely stores used injectables. It encases a traditional waste bin and provides patients with an easy, elegant, safe, and connected way to dispose of their used needles, syringes, vials or injectors. HealthBeacon uses customized reminders and provides real time support to individual patients with the help of a dedicated customer care team to help patients start and stay on track with their medication. HealthBeacon tech enables us to identify patients around the globe who miss their injection on the scheduled day and initiates targeted outreach by our Customer Support team.
The HealthBeacon was designed using customer feedback with patient empowerment as the top priority, leading to 90% patient acceptance and documented improvements in adherence and persistence to therapy. HealthBeacon has been globally adopted and regulatory approved on the market since 2014 and has tracked > 500,000 injections in over 14 countries to date. Addressing the needs of individuals who self-inject medications at home is critically important. Through our experience and research, we are leading the delivery of digital health solutions to this growing and vulnerable population.
To improve the patient experience, we opened the HB Labs which facilitates extensive collaboration between the data and product teams. Once devices have been tested in the user lab and are released for patient us e, the data team constantly monitors and identifies new behaviours and insights from the actionable data that we capture using the devices in the market. The data also helps us to understand the relationship between different treatments and the patient outcomes and this in turn helps us to design the most effective and optimized patient support programs. The use of Artificial Intelligence in combination with the data that we collect has also led to significant improvements in areas such as automated clinical decision-making.
Reason for Nomination:
When an injection is dropped into the HealthBeacon device it breaks a sensor which in turn triggers the camera to take a picture of the injection device. The captured image is sent to the server in real time using a ma chine-to-machine sim card and is classified to determine what injection the patient has dropped into the smart sharps bin for that injection event. Initially, IBM Watson Visual Recognition service was being used to train the model and use it for the image classification. However, due to challenges mentioned below the system’s overall performance for classifying images was not at a threshold that we were comfortable with.
Challenges:
1. Some of the injection devices look very similar to each other and it can be very difficult to tell the difference between them.
2. The Quality of some of the images that we receive can be very poor due to external factors such as intensity of the ambient light surrounding the unit.
3. Not all images from the same device will look the same. The visibility of the device in the image depends on user behaviour and how the patient drops their medical waste into the device. In some of the images only half of the device may be visible or sometimes only the cap of the injector or syringe will be present.
Objective:
The objective here was to build a robust in-house object detection platform to detect the injector and syringes from the images by overcoming the challenges mentioned above. The platform is being used to accurately distinguish between medication s and to report the type of injection and brand of the medication.
Methods/Approach:
The main reason for moving to the in-house solution is for scalability and to have control over the architecture of the platform. Scalability was a very critical factor here as we are launching with new clients and new drugs each year. We aimed to build a platform that we can scale both vertically and horizontally without impacting the overall performance. We required a clever approach to deal with the challenges without making any major changes to the device functionality or business operations. The multi-level architecture made up of 7 unique object detection mode ls is built to detect the injection device from the image with highest confidence. All the models are deep learning models which are trained using the transfer learning approach where YOLOV3 is the base model which is fine tuned for the custom objects. The 2-level architecture has been designed in such a way that:
1. At level 1: The “Blue Sky” model which is trained to detect all injection devices (currently 13). 2. At level 2: A set of 6 unique object detection models are present which are being triggered based on the result of the Blue-Sky model and the related drug configured in the system for that patient. The model at level 1 is used for generalization and to guide the platform to level 2. Based on the result of level 1, confidence and drug configurations, either ensemble averaging, or ensemble max voting will be triggered to verify the object more closely and to avoid any error in the detection. At level 2 a high level of diversity is achieved between models by grouping the injection devices in each model differently and by using distinct training data. Each unique injection is only included in 3 models at level 2 so that ensemble machine learning techniques can be used to improve the accuracy further. Ensemble learning helps improve machine learning results by combining several models with a high level of diversity amongst them.
Impact:
(1) Accuracy: By moving to the in-house object detection plat form, we have overcome the challenges faced and are processing images with minimal error.
(2) Performance: The overall performance of the system is better than the old classification system using IBM Watson.
(3) Scalability: The scalability challenges are addressed as we can scale the platform horizontally without the requirement to add additional resources and without impacting the existing performance.
(4) Customisation: Moreover, custom use cases can also be accommodated without altering the existing architecture of the platform.
(5) Efficiencies: Solution is helping us to automatically detect the injection devices from ap proximately > 10,000 images per month which would have required human revie w and classification. In some of the business cases, the solution also helps to verify the medication used by the patient which is evidence for the adherence of the patient to a specific drug.
Additional Information:
HealthBeacon is a major catalyst for change in the Healthcare industry by connecting patients, medical practitioners, and drug development companies through data. HealthBeacon is deploying smart devices in patient homes that connect patients real time to its platform and is a radical paradigm shift in traditional healthcare pathways.
From the previous research studies, we have seen that the overall adherence of patients drops drastically after initiation and is being reported in the range of 46% to 66% over the treatment duration with variation in adherence due to gender, age and therapeutic area. By enhancing patient support programs and mapping those with the individual patient’s routine and behaviour, HealthBeacon aims to improve the adherence further. The HealthBeacon’s Injection Care Management System is helping to improve not only adherence but also the persistence across multiple therapeutic areas with overall adherence rates of >80% being reported.
Prior to the HealthBeacon, there was inadequate independent data available to prove that patients had actually taken their medication. As such, a key feature of the HealthBeacon is the ability to independently track and record the different types of medications that are taken by patients and disposed of in the smart sharps bin. Our in-house image detection platform is us ed to accurately detect injection devices from images. Our innovative AI platforms along with the technology running in the backend are built based on the robust data analytics and understanding of the patient journey and the relevant factors. The extensive use of AI integrated technologies is benefitting both the patients and the system. All components connected together are helping us to digitally monitor the patients’ health more effectively and to intervene smartly at the right time.