NWO interview - Predicting and preventing falls

People over the age of 65 have a high chance of tripping and falling, and medication use is most often the cause of this. However, which individuals are most at risk and how can that risk specifically be reduced? The big data project SNOWDROP is developing ways to calculate risks of falling on a case-by-case basis. General practitioners and patients can use this information to jointly decide whether and how they want to adjust the use of medication to prevent falls. This is a project from the Commit2data programme about machine learning and improving doctor-patient consultations.

About one-third of people over the age of 65 fall at least once a year. The frequency of falling increases with age, the number of different conditions a person suffers from and an increasing degree of vulnerability. The medical and psychological consequences of falls are severe. In the Netherlands, a person over the age of 65 ends up at the casualty department every five minutes due to a fall. ‘Although certain groups of medication are known to increase the risk of falling, it was so far impossible to calculate how high the risk of falling is for any individuals using such medication’, says SNOWDROP project leader Prof. Ameen Abu-Hanna from Amsterdam UMC, location AMC. That is because this risk not only depends on the medication a person is on but also on how the patient’s body reacts to that medication. ‘Based on the data in a person’s medical file, we want to be able to automatically identify who runs a high risk of falling. The general practitioner can then talk with the patient and decide how they want to limit that risk.’

Extensive database

There are various ways in which medication can cause a fall. For example, some medication leads to decreased muscle strength, poor balance or a slower reaction time. General side effects such as fatigue, confusion and dizziness can also lead to a fall. Older people are often on several types of medication at once. And the individual sensitivity to side effects changes as a person gets older. The SNOWDROP project uses data from more than 75,000 people aged over 65 years who registered at one of the 109 GP practices connected with Amsterdam UMC. This database contains, amongst other things, information about medication use, age and gender but also lab results, diagnoses, referrals and free text notes in Dutch.

As the electronic file system does not have a separate code for falls, we cannot tell whether somebody has fallen or not. We have to infer that from other information.

Noman Dormosh

Finding and predicting falls

PhD student Noman Dormosh from the Department of Medical Informatics is working on this part of the project. ‘The biggest challenge here is obtaining relevant data’, he says. ‘As the electronic file system does not have a separate code for falls, we cannot tell whether somebody has fallen or not. We have to infer that from other information.’ For example, from the free text sections of tens of thousands of files, Dormosh manually identified expressions that indicate a fall using different words. ‘Examples include trigger words like slipping, collapse or stumbling.’ He subsequently automated this process to identify falls in other files. Based on this information and with the help of machine learning, he produced two models that can predict future falls. The first model uses only structured data, such as age, gender, medication use and lab results. The second model uses this information, as well as the free text notes from GPs. The models were developed using data from GP practices connected with Amsterdam UMC location AMC and were subsequently validated externally on data from practices connected with Amsterdam UMC VUmc location.

Worthwhile consultation

The second PhD student in the project, Leonie Westerbeek, is a communication scientist. She is investigating how GPs can engage in a worthwhile consultation with the patient based on the predictions made by the first prediction model. ‘We are working together with ExpertDoc, supplier of the programme NHGDoc, which already provides decision support for GPs in the Netherlands. We want to add an extra module aimed at falls to that program. In my research, I work together with GPs to find out what is needed for this and what shape it should take.’ For example, during the focus groups with GPs, we investigated how you can best present a risk to the patient and to the GP. ‘The answer was: a combination of the absolute figure that indicates the risk of falling, combined with a colour bar from green to red.’ Intervention recommendations have also been added to the system based on the official guidelines for GPs. Besides the options to adjust medication, the system also provides suggestions, for example, about mobility and making the home fall safe.

SNOWDROP is not only aimed at GPs but also very much at the patient so as to encourage joint decision-making. How can patients best prepare themselves for a consultation about reducing their personal risk of falling? Westerbeek: ‘Prior to the visit to the GP, the patients receive a list of questions that they might want to ask. Patients use this list to pick several questions they would like to discuss with the GP. The GP can see this information before the consultation, so they know what needs to be discussed.’

Going well

The researchers are proud to say that the project is going well. Abu-Hanna: ‘The predictive model was finished sooner than expected. It has already been implemented as a test module in the programme for GPs.’ Westerbeek: ‘We will now evaluate how this works in practice at several GP practices. At this stage, we will mainly examine the communicative aspects: to what extent does a warning from the model combined with the decision support and the patient portal actually lead to a consultation? How does such a consultation go? Is the medication adjusted as a result of this, and if so, to what extent? If this method proves effective, a next step could be to set up a larger-scale trial to find out whether people actually fall less after this intervention. However, that step would fall outside of the scope of this project.’

Dormosh is also already thinking beyond SNOWDROP. ‘In principle, all of the informatics-related work in this project has already been completed and implemented. It would be great if we could take a more detailed look at how different medication affects certain conditions. I’m also curious about whether it is possible to improve the accuracy of the predictions by including other parameters or by changing the hierarchy of the medication used. For instance, is it sufficient to know whether somebody uses a sedative or is it actually important to know that the sedative in question is benzodiazepine? Which level of information is the most suitable?’

It is up to us to develop a system that understands the free text notes and can find new connections between them.

Ameen Abu-Hanna

More widely applicable

Abu-Hanna concludes: ‘Within this project, valuable methods have been developed that can be applied to much more than just falls. For example, we are using several of the algorithms developed in a project with the Dutch Cancer Society to examine whether we can find new indications for a high risk of lung cancer other than the factors age, gender and whether somebody smokes or not. Ultimately, we are of the opinion that GPs must be able to make notes in the free text sections in their own style and not just using standardised codes. And it is up to us to develop a system that understands the free texts and can find new connections between them.’

The SNOWDROP project is part of Commit2data, a platform that organises research programmes and is co-funded by NWO. Commit2data wants to make optimal use of big data possibilities for societal and economic applications.

Ameen Abu-Hanna is Professor of Medical Informatics and Head of the Department of Medical Informatics at the Academic Medical Center at the University of Amsterdam. Amongst other things, his research focuses on artificial intelligence, machine learning, data mining and statistics, prognostic modelling and evaluation and decision support applications in medicine and healthcare. Ameen Abu-Hanna obtained a BSc (cum laude) in computer technology, an MSc in computer sciences (both at the Technion in Haifa), and a PhD in artificial intelligence at the University of Amsterdam.

Ameen Abu Hanna

Noman Dormosh is a PhD student at the Department of Medical Informatics of the Academic Medical Center at the University of Amsterdam. He obtained a BSc in pharmacy at the Jordan University of Science and Technology, an MBA at the EU Business School and an MSc in Medical Informatics at the University of Amsterdam. Dormosh has extensive experience in industry, for example as a business manager at WeDev, product manager at Ibn Al-Haytham Pharmaceutical Industries Co., and system administrator at HostingFest LTD.

Leonie Westerbeek is a PhD student in communication sciences at the University of Amsterdam. She obtained her BSc in communication sciences at VU Amsterdam, and her MSc in communication sciences at the University of Amsterdam.