Using Horse Sense to Reduce Bedcentricity in Hospitals
When you’re in the hospital, your bed is your home and your haven. You wear pajamas all day and wait for strangers to bring you food, medicine, or drinks. Friends and family gather around your bed during visiting hours. You adjust pillows, foot and head elevations, and blankets over and over, until you feel comfortable and safe.
It turns out that bed is also working against you.¹ When your body isn’t moving, it doesn’t recover as quickly. As a result, patients who stay in the hospital to get better often leave the hospital in worse physical shape than they were in when they arrived.² Research has shown that this is especially prevalent in elderly patients and patients who are not able to walk properly.
Physiotherapists at the Radboudumc in Nijmegen, the Netherlands, are on a mission to send people home from the hospital healthier than they would have been if they had stayed in bed. They call this mission Ban Bedcentricity and is led by Shanna Bloemen and Yvonne Geurts. They want people to leave their beds. They started by telling patients why it is unwise to stay in bed all the time and how it affects them. While they may have felt safe, it was not helping them get better. Another part of the program focused on making it easier for patients to move around. They provided patients with comfortable chairs, added railings to walls, and created break spots in the hallway, where mobile patients could sit down and regain their strength. These interventions aimed to get people out of their rooms and moving around.
The physiotherapists at Radboudumc wanted to go further than improving patient education and their physical environment. First they needed to establish whether the interventions were working or not. Doing that required insight into how much patients were moving around. The physiotherapists and their staff could not keep tabs on every single patient during the day to learn if and how much the patients were moving. They tried using self-reported diaries, but research showed that patients often over-estimated their activity levels.³ Commercially available activity trackers did not perform as well as expected because they were designed for healthy people and only recognized healthy walking gaits.⁴ When a person walked with an adjusted gait due to using a mobile walker, an IV-stand, or crutches, those activity trackers had trouble recognizing what the person was doing.
In 2017, the physiotherapists started pitching their idea for a personalized, self-learning, custom-made activity tracker for the Ban Bedcentricity program to possible partners. They found Totem Open Health, a company that had developed an open source sensor.⁵ They also found us, Orikami, a boutique data science company that specializes in healthcare. This was around the time of the annual Dutch Hacking Health event. Radboudumc, Totem Open Health, and Orikami each sent several representatives to the hackathon and they spent a weekend at the hackathon at the Radboudumc brainstorming and developing to see what was possible.
Operating under the motto, “One weekend, one goal, one nation,” Dutch Hacking Health focusses on innovation in healthcare. They organize simultaneous hackathons throughout the Netherlands and at each location, professionals join forces to see how they can use technology to tackle difficult problems in healthcare. They come from different professions; developers, UI/UX-designers, data scientists, product managers, and more. The event is a single weekend of brainstorming, design, development, sleep deprivation, and a final presentation. Dozens of highly engaged and motivated individuals and organizations attended for fun, networking, or because they love a novel challenge. Dutch Hacking Health awards prizes at the end of the weekend with winners selected by juries or peers. This lends a competitive edge to the atmosphere, which is alleviated by the fact that everyone’s working on a shared theme, so no matter who wins, everyone makes a contribution to the field.
For the Radboudumc, Orikami, and Totem Open Health, our ideal outcome was to verify whether the idea of a custom-made sensor with an activity recognition model was feasible. If it could work, then we could consider moving the project forward. Each part of our team brought special skills and knowledge to the group that weekend.
Totem Open Health brought their open source sensor to use for data collection during the hackathon. The physiotherapists brought their expert knowledge about the human body and movement and their data. Orikami brought specialized experience with data collection and analysis. We used Totem’s sensor to collect all kinds of activity data during the hackathon from ourselves and other attendees. We attached sensors to them and asked them to walk around, sit, stand, and lie down. We used to the collected data, along with previously available data, to try to develop a machine learning model which could recognize what type of activity a person was performing.
At Orikami, we had worked on a similar problem before, although with one important variation. Our subjects weren’t human, they were horses. For them, we developed machine learning models that could identify a horse’s gait, whether it was walking, trotting, or cantering. The products we developed for that specific scenario, weren’t immediately applicable to the Ban Bedcentricity program, but our experience gave us insights into the type of data we could expect from the sensors we were using, and experience with developing a good machine learning model to work with that data. Using a combination of theoretical and practical knowledge we gained working with the horse data, we developed a new model for activity recognition in humans that worked surprisingly well for a first version.
By the end of the hackathon weekend, we had a simple prototype that was trained on the small dataset collected from our fellow hackers. Training in this case meant letting the machine learning model learn to find patterns between the collected sensor data and the related labels. That meant the computer learned to interpret a pattern in the sensor data as a specific movement and labeled it as such. That version still had trouble recognizing unique and specific activities. To use the sensors in a real-world setting we needed a machine learning model that was more robust and could recognize a wide variety of unique situations that might occur only in a hospital. Based on the success of our hackathon experience, we decided to move forward together and won a grant to fund further development work towards this more robust model.
First, we needed to collect a larger dataset that mimicked what actually happens in a hospital. This included how patients move and the specific movements they perform. The physiotherapists at the Radboudumc used healthy volunteers who imitated patient activities in a hospital to collect this data. The movements they mimicked included lying, sitting, standing and walking, both with and without various medical aids like walking wheels, IV-stands and crutches.
During this phase of the project, we investigated the role of sensor placement on a patient’s body. Our goal was to figure out which locations on a patient’s body gave the most accurate activity results, while not hindering the patient’s movement or comfort. Using a combination of sensors is interesting because a single sensor may not always recognize different body positions. Let’s say you have a sensor on your chest. Whether you’re sitting upright or just standing still, the sensor’s movement would be pretty much the same. That means it would not be clear to this sensor which of the two you are doing. The same holds for a sensor on your upper leg, which would be in the same orientation for both sitting and lying down.
Since most conventional activity trackers focus on when a person is moving, they can do their work with one location on the body. But in a hospital setting, even the small amount of activity you get from standing still matters. In order to figure out how we could use multiple sensors, we attached six sensors to various places on each patient’s body, with location choices based on relevant scientific literature.⁶ With these six sensors, we collected a large amount of information during each experiment. At a later stage, we create numerous models based on each sensor location and also tried out combinations of different locations.
Thanks to the data collected by the sensors, we were able to create numerous models for activity recognition based on various sensors and combinations. We found that having two sensors worked significantly better than having just one. But not just any two sensors work in combination. Any combination of the lower back, chest, or lower leg, for example, didn’t improve the results much because they all have trouble differentiating between sitting and standing. We had to combine sensors that struggled to differentiate between activities, like the upper leg and chest sensors, to improve our results. A number of these trained models outperformed all other activity sensors that they tested inhouse for this task.
Orikami and the Radboudumc then developed a prototype solution that can give physiotherapists insight into which activities a patient performs during the day. The next big goal is to put this activity recognition model into production. That requires designing a sensor system which is collects reliable data, is comfortable to wear, and easy for health care staff to use. No matter how well a system like this may perform, it has to fit seamlessly into high-stress, high-stakes hospital environment.
One day, physiotherapists will use this information to give patients personalized and meaningful movement goals and the ability to monitor their progress towards those goals. In the meantime, their physiotherapists would be able to review patients’ progress by checking the tracker data. One day, this physical activity tracker could lead to a physical activity biomarker to indicate patient recovery that all healthcare providers could use.
We envision that in the near future, all patients in the hospital will wear movement sensors collecting data used to encourage them to get out of bed and increase the amount of meaningful activity they get, thus speeding their recovery and putting them on track to get back home, healthier.
Special thank you to Assistant Professor Thomas Hoogeboom and PhD candidate Niek Koenders, both at Radboudumc, for their valuable input on this article.
[1] Koenders, Niek, et al. ““I’m not going to walk, just for the sake of walking…”: a qualitative, phenomenological study on physical activity during hospital stay.” Disability and rehabilitation (2018): 1–8.
[2] http://www.apta.org/Blogs/PTTransforms/2019/7/1/EveryBODYmoves/
[3] Prince, Stéphanie A., et al. “A comparison of direct versus self-report measures for assessing physical activity in adults: a systematic review.” International journal of behavioral nutrition and physical activity 5.1 (2008): 56.
[4] Koenders, Niek, et al. “Validation of a wireless patch sensor to monitor mobility tested in both an experimental and a hospital setup: A cross-sectional study.” PloS one 13.10 (2018): e0206304.
[5] Totem Open Health recently closed down.
[6] Cleland, Ian, et al. “Optimal placement of accelerometers for the detection of everyday activities.” Sensors 13.7 (2013): 9183–9200.
