Big Data in orthopaedic surgery
J. Kuipers1 J.N. Doornberg2,3 T. Gosens4 1 Department of Orthopaedic Surgery, Erasmus University Medical Centre, Rotterdam, the Netherlands 2 Department of Orthopaedic Surgery, Flinders Medical Centre, Adelaide, Australia 3 Department of Orthopaedic/Trauma Surgery, University Medical Centre Groningen, the Netherlands 4 Department of Orthopaedic Surgery, Elisabeth-TweeSteden Hospital, Tilburg, the Netherlands Corresponding author: J. Kuipers, firstname.lastname@example.org
In recent years, increasing attention has been paid to buzzwords such as Big Data, Artificial Intelligence, Machine Learning, Deep Learning and predictive analytics. Nevertheless, many people do not exactly understand what it means, how these analyses could be used in healthcare and, more specifically, how these techniques can be applied in orthopaedic surgery? In this article, the most frequently asked questions about these topics will be addressed: • Why is Big Data relevant for me as an orthopaedic surgeon? • What is Big Data? • How is Big Data related to Artificial Intelligence and Machine Learning? • Why is it happening now? • What are the pros and cons? • How can I integrate the use of Big Data in my daily routine?
Why is Big Data relevant? “Big Data is like teenage sex: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it.”1,2 Dan Ariely, professor of psychology and behavioural economics - 2013 As medical professionals in the Netherlands we are acting on a challenging playing field, as we have to deal with a lot of highly demanding and different stakeholders: patients, governmental bodies, hospitals, and insurance companies. The increasing use of benchmarking and quality indicators in the Dutch healthcare system leads to a pressure cooker effect: better, faster care and cure for the lowest price. As a reflex, people in healthcare work harder and harder to meet the raising standards. Due to the rising costs of our healthcare system (100 billion last year) and an expected rising of costs up to 140 billion per year in the next decades,3 it is necessary that a system change will take place in the near future. A proposed change is the shift to a Value Based Health Care system. In this system you will not get paid for what you deliver (e.g., surgeries), but the payment is linked to the amount of ‘value’ you add to someone’s life. In the discussion of how value is measured, the initiative should belong to the patient and the healthcare professional, because care and cure is a service transaction based on trust and responsibility between care provider and care recipient. The amount of data collected in the healthcare system, and by patients themselves through wearables and smartphones etcetera, leads to a wealth of data. By exploring these data and combining it with our medical knowledge, we can estimate a definition about the ‘value’ of care and good quality of care. More importantly, we may get insights into the factors that are associated with a good or bad outcome. These insights potentially enable us to deliver the right treatment at the right time to the right patient, also described as coined personalized medicine or precision medicine, two other increasingly used buzz words. So why is Big Data relevant for me as an orthopaedic surgeon? As described above, a system change seems to be inevitable. In many current Value Based Health Care cases, orthopaedic surgery is taken as an example. This is because in our field most conditions are well described and defined, orthopaedic surgery is elective, care pathways are often highly protocolized, and patients are often otherwise healthy. These qualities make orthopaedic surgery also highly suitable for Big Data and other smart applications. As medical specialists we have to think about what the value is we add and how we can measure it, preferably together with our patients. What type of data do I need for what I want to analyse? Does it specifically reflect the patient’s medical condition and recovery, or is it a surrogate measure of overall psychosocial wellbeing? We feel that there is ample data to answer these important questions. We argue that Orthopaedic Surgeons should take the lead to define best Orthopaedic Value Based Health Care.
What is Big Data? “Big Data is indeed the next natural resource – promising to do for our era what steam, electricity and oil did for the Industrial Age.”4 Ginny Rometty, former CEO at IBM – 2013 There is no clear uniform definition for Big Data. Kitchin and McArdle5 investigated the different traits appointed to Big Data. One of the first definitions of Big Data in 2001 described that it should consist of the three V’s: volume (consisting of enormous quantities of data), velocity (created in real-time) and variety (being structured, semi-structured, and unstructured). In recent years, other qualities have also been attributed to Big Data: exhaustivity (n=all), fine-grained, uniquely, indexical, relationality, extensionality, scalability, veracity, value, and variability. Many of the datasets we currently consider as Big Data, are actually ‘Small Data’. For instance, our national arthroplasty registry (LROI) is actually an example of small data. It contains a lot of entries, as almost all the arthroplasties in the Netherlands are included. Nevertheless, we only collect data from patients who underwent an arthroplasty. Importantly, we do not have data of patients who are managed conservatively. There is no real-time data availability and there is not much variety in data. Furthermore, it mainly consists of structured data from questionnaires (Patient-Reported Outcome Measures). So, there is a twilight zone between Small Data and Big Data. Important characteristics which distinguish between Small and Big Data are velocity (both frequency of generation, and frequency of handling, publishing and recording) and exhaustivity. Big Data is all-inclusive and quick. The fact that you can get real-time access to data of all cases sets Big Data apart and has led to its recent popularity. Why is it happening now? “I think AI is akin to building a rocket ship. You need a huge engine and a lot of fuel. If you have a large engine and a tiny amount of fuel, you won’t make it to orbit. If you have a tiny engine and a ton of fuel, you can’t even lift off.”6 Andrew Ng, chief scientist in the Google Brain Project - 2015 True Big Data cannot be analysed with ‘classic’ statistical methods. The term Artificial Intelligence (AI) has been around since the 1950’s and it describes any technique that enables computers to mimic human behaviour. Machine learning (ML) has been around since the 1980’s. It is a technique, under the AI umbrella, that gives computers the ability to learn without being explicitly programmed to do so. Deep learning is a subset of ML which makes the computation of multi-layer neural networks feasible. For this last technique you need extensive computing power. All these techniques use mathematical equations, better known as algorithms. Due to the continuous development of faster computer chips, supercomputers are now able to digest the enormous chunks of data (Big Data) offered to them. In 2003, the human genome was sequenced for the first time, after thirteen years and three billion dollars had been spent. Nowadays, it costs about one hour and 100 dollar to perform whole genome sequencing.7 Andrew Ng eluded to the fact that ML and Deep learning techniques have one thing in common: they need tons of data to fuel them. The magnitude of data created every hour, every minute, and every second, even at the moment you are reading this goes beyond the scope of our imagination. The internet of things, smartphones, wearables, and other gadgets provide a constant stream of data. The magnitude of this stream has the potential to give us unprecedented insights in structures and processes of daily practice. For instance: “Why does nearly 20% of the patients who underwent total knee arthroplasties still have complaints after surgery without any apparent cause?”, or: “Which patient is more prone to develop a prosthetic joint infection?” However, these questions can only be answered based on good quality data. In orthopaedic surgery, current data collection is small and either unstructured or semi-structured. The potential for meaningful bigger data is limited by the very poor quality of current data collection, due to heterogeneity in our Electronic Medical Records (EMRs).
What are the pros and cons? Big Data could help us to provide personalized care by individual risk stratification: precision medicine. This can be compared to the advertising industry, where precision marketing has become mainstream. In these advertorial fields, Big Data and algorithms already play an important role in day-to-day affairs. One could argue whether these technically very advanced applications are used in an ethically responsible manner: do we really want Big Data (fuel) and AI (engine) to make our kids watch too many YouTube videos due to smart advertising, or make us binge on Netflix? The ultimate potential of Big Data in healthcare is to deliver the right care to the right person at the right place at the right time: precision medicine. We will explain the great potential, as well as the important risks in the text below. In the current era, our daily routines are getting more and more streamlined by Big Data companies and their algorithms. For instance, Netflix knows what people like you prefer to watch, Google knows what you are looking for (Search), where you are going (Maps) and what kind of videos you like (YouTube). Some claim the algorithms behind Facebook know you better than your spouse.8 We tend to get scared by things we do not understand, therefore it may be better that most people are not aware that a lot of the techniques mentioned are already incorporated in our lives. It is essential to know that we all use (free) services, for which we (partially) pay with our data. In the near future, Google health,9 Apple healthcare,10 Alihealth11 and some other big players will use their experience in pattern detection algorithms and will apply them on healthcare records. This could be an interesting development, but correlation versus causality is not always straightforward, and therefore challenging for medical specialists and experts to interpret, while these sources potentially are adopted (too) quickly by others (patients, governmental bodies, and insurance companies). The question of debate is: will medical specialists be outperformed by these algorithms in the near future? For example, will computer algorithms be better in selecting if patients should be offered arthroplasty than orthopaedic surgeons? Will an algorithm be better in predicting failure or success of an operation? If so, will surgeons only execute orders and ‘become a slave’ of data and algorithms?12 There are several reasons why this is not likely in the (near) future. First, Big Data is never totally raw or unbiased, as there is always a reason why certain data was collected.13 Hence, it remains important to assess the reasons why data were collected and in what specific context.14 Predictive algorithms may fail, or wrong conclusions may be drawn when the data is not used in the right context. For example, prediction models for the outcomes after surgery based on administrative databases from the US are not applicable on a Dutch population, as none of these applications have been externally validated.15–17 It is well established now, that most algorithms perform very well in the institutions where training and internal validation have been performed. However, external validation of these algorithms is still very rare. Second, Big Data can give us more insights in the ‘when, what and where’, but is not likely to answer questions like ‘how and why’. Thus, Big Data is not meant to inform us about causal relations and cannot answer all research questions. These challenges associated with Big Data show that the knowledge and experience of a medical specialist will remain of utmost importance in the future. We should embrace and utilize the possibilities of Big Data to improve personalized treatment and add more value to care. We can argue that externally validated clinical applications of AI (engine) trained on good quality Big Data (fuel), will have the potential to support surgical decision making. The algorithms have potentially ‘seen’ a multitude of patients as compared to the treating surgeon. And therefore, will add to surgeons’ own experience, and will partially overcome surgeons’ bias.
How can I integrate the use of Big Data in my daily routine? There are several examples of how new technologies are already helping us. Most progress is made in the field of radiology, with pattern detection algorithms, for example in detection of simple fractures.18,19 But for more complex tasks, such as identification of occult scaphoid fractures, experts still outperform the computer.20 However, the combination of a -junior- doctor, with the Convolutional Neural Network algorithm providing a heat map on plain radiographs to point our young colleagues in the right direction during their nights shift in the ED, is currently the field of application with the highest yield.19 So, how can we use Big Data to improve our daily practices as an orthopaedic surgeon? The authors’ proposed horizon is illustrated by the following clinical case that applies multisource structured and unstructured data (fuel) in a ML algorithm (engine) for risk stratification: Patients are referred to an orthopaedic surgeon. Prior to hospital visit, patients fill in their own history, symptoms, state of mind etcetera, through smart questionnaires, using computer adaptive testing (CAT).21 In addition, patients can upload their data from their health app on their smartphone/wearable in their personal health care record. With the help of a motion traction visualisation application using the smartphone camera, the app can already make a judgement about range of motion and movements. At the outpatient clinic, the history will be checked and structured accordingly. The patient-doctor conversation is automatically logged in a WhatsApp-like format by the computer through Natural Language Processing (NLP).22–24 Physical examination findings will be processed through NLP in combination with Video Movement Analysis Using Smartphones (ViMAS).25 This combination of data potentially gives us a more dynamic view of the patient’s complaints. Further examinations (e.g., X-rays, scans) can be added in the personal health record. The algorithms have the potential to make a structured summary of the outpatient visit based on previous experiences (reinforcement learning). Based on the detected patterns, the algorithm cross-references the patient with world-wide data and gives a differential diagnosis including a probability score. Especially in more complicated or rare cases (cases in which the ML algorithm has seen more patients, i.e., has more ‘experience’ than the treating surgeon), this can help to establish a diagnosis or guide personalized treatment decisions, using all available knowledge at this moment in real-time. So, when one is pre-planning an arthroplasty for (robotic) surgery, patient and surgeon will immediately get failure and success probabilities as well as patient-reported outcomes, based on the data of (millions of) previous arthroplasties26. Postoperative follow-up will again be done through smart questionnaires and computerized adaptive testing. As a medical specialist you will have your own arthroplasty dashboard where you can see how well every patient is doing27. You will get a notification if a patient’s rehabilitation negatively deviates from normal, after which a visit at your outpatient clinic or a video consultation will be scheduled. Maybe this sounds like a fairy-tale, but all separate technologies do exist (!). All it needs, is adequate structuring of currently available applications. We feel that because all technologies are commercially available separately, this prohibits integration to a level that truly benefits patients. Therefore, all scientists at Massachusetts Institute of Technology (MIT) work in an ‘open access’ mentality. Their vision is that sharing data, and integrating ‘open access’ algorithms, will facilitate a steep learning curve of AI applications. Medical Information Mart for Intensive Care (MIMICs) is their proof of concept: ‘open access’ safe data sharing of thousands of ICU patients for clinicians and scientists to develop and train their predictive models on. We argue that this ‘open access’ mentality should become the standard across orthopaedic surgeons for collecting and sharing patient outcome data of operative, as well as non-operative treatment, in the Netherlands.
What do we have to do now? “Garbage in is garbage out.” We have to think about data. We have to think about data sharing. Data that has the potential to deliver the best quality of care to our patients. Data can facilitate personalized care by risk stratification. Many new technologies and derived applications make it possible to analyse data from the past and the present. We have to think about which data we should collect for the future and how we structure this. First, we have to define current (unmet) needs and unexplored research questions. For example: are real-world patient activities, collected from apps and wearables, correlated with functional outcomes and patient perceptions? Are patient-reported outcomes and patient experiences associated with success? How do we define and measure success? Nowadays, many data are collected and stored in EMRs, but these data are often difficult to access, not broadly shared, and oftentimes left unused.28 There are a number of conditions and challenges that have to be taken into account before we can use Big Data in orthopaedic surgery:29 1. Standardization of data registration; in order to make data more comparable, reproducible and processable for algorithms. 2. Quantity and quality; sufficient quantity, but especially quality, is the essential building block for an algorithm to make an accurate prediction. 3. Verification; controllability is more difficult with more advanced algorithms. Testing and transparency of algorithms will be the biggest challenge for the future.30 4. Ethics; most discussions nowadays are about the ethical aspect. The issues that arise are, for example: who is in charge? Who owns the data? Who is ultimately responsible? Are decisions made by AI applications fair?30 We live in a time of disruptive changes, which not only causes many challenges, but also leads to new possibilities to improve the value of care, if we adopt and incorporate the use of Big Data in our care processes. “If you always do what you’ve always done, you’ll always get what you’ve always got.” Henry Ford – 1908 Disclosure statement The authors have nothing to disclose.
References 1. Ariely D. Big Data is like teenage sex [Internet]. facebook.com. 2013;Available from: https://www.facebook.com/dan.ariely/posts/904383595868 2. Ariely D. Dan Ariely [Internet]. Available from: https://danariely.com/wp-content/uploads/2020/04/Ariely_CV.pdf 3. Zorginstituut [Internet]. Available from: https://www.zorginstituutnederland.nl/financiering 4. Richard N. Haass, President C on FR. A Conversation with Ginni Rometty [Internet]. 2013;Available from: https://www.youtube.com/watch?v=SUoCHC-i7_o 5. Kitchin R, Mcardle G. What makes Big Data , Big Data? Exploring the ontological characteristics of 26 datasets. 2016;(June):1–10. 6. Carling C. Andrew Ng: Why ‘Deep Learning’ Is a Mandate for Humans, Not Just Machines [Internet]. wired.com. 2015;Available from: https://www.wired.com/brandlab/2015/05/andrew-ng-deep-learning-mandate-humans-not-just-machines/ 7. Gent E. $100 Genome Sequencing Will Yield a Treasure Trove of Genetic Data [Internet]. 08 March. 2020; Available from: https://singularityhub.com/2020/03/08/100-genome-sequencing-will-yield-a-treasure-trove-of-genetic-data-and-maybe-a-dystopian-nightmare/ 8. Youyou W, Kosinski M, Stillwell D. Computer-based personality judgments are more accurate than those made by humans. Proc Natl Acad Sci U S A 2015;112(4):1036–40. 9. Google Health [Internet]. Available from: https://health.google/health-research/ 10. Apple Gezondheidszorg [Internet]. Available from: https://www.apple.com/nl/healthcare/ 11. AliHealth [Internet]. Available from: https://alihealthgroup.com/ 12. Bijlsma J. Slave to the algorhythm [Internet]. Ski. nr. 2-3. 2019;10–4. Available from: www.skipr.nl 13. Beletsky A, Lu Y, Patel BH, et al. Should We Question the External Validity of Database Studies? A Comparative Analysis of Demographics. Arthrosc - J Arthrosc Relat Surg [Internet] 2019;35(9):2686–94. Available from: https://doi.org/10.1016/j.arthro.2019.05.020 14. Dalton J, Thatcher J. What does a critical data studies look like , and why do we care? Seven points for a critical approach to ‘big data.’ Soc Sp [Internet] 2014;1–12. Available from: http://societyandspace.org/2014/05/12/what-does-a-critical-data-studies-look-like-and-why-do-we-care-craig-dalton-and-jim-thatcher/ 15. Ramkumar PN, Navarro SM, Haeberle HS, et al. Development and Validation of a Machine Learning Algorithm After Primary Total Hip Arthroplasty: Applications to Length of Stay and Payment Models. J Arthroplasty 2019;34(4):632–7. 16. Ratliff JK, Balise R, Veeravagu A, et al. Predicting occurrence of spine surgery complications using big data modeling of an administrative claims database. J Bone Jt Surg - Am Vol 2016;98(10):824–34. 17. Navarro SM, Wang EY, Haeberle HS, et al. Machine Learning and Primary Total Knee Arthroplasty: Patient Forecasting for a Patient-Specific Payment Model. J Arthroplasty [Internet] 2018;33(12):3617–23. Available from: https://doi.org/10.1016/j.arth.2018.08.028 18. Langerhuizen DWG, Janssen SJ, Mallee WH, et al. What Are the Applications and Limitations of Artificial Intelligence for Fracture Detection and Classification in Orthopaedic Trauma Imaging? A Systematic Review. Clin Orthop Relat Res 2019;477(11):2482–91. 19. Lindsey R, Daluiski A, Chopra S, et al. Deep neural network improves fracture detection by clinicians. Proc Natl Acad Sci U S A 2018;115(45):11591–6. 20. Langerhuizen DWG, Bulstra AEJ, Janssen SJ, et al. Is Deep Learning On Par with Human Observers for Detection of Radiographically Visible and Occult Fractures of the Scaphoid? Clin Orthop Relat Res 2020;1. 21. Liu TC, Ohueri CW, Schryver E, Bozic KJ, Koenig KM. Patient-identified Barriers and Facilitators to Pre-Visit Patient-Reported Outcomes Measures Completion in Patients With Hip and Knee Pain. J Arthroplasty [Internet] 2018;33(3):643-649.e1. Available from: https://doi.org/10.1016/j.arth.2017.10.022 22. Leids Universitair Medical Centre [Internet]. Available from: https://campagne.lumc.nl/ai-in-de-zorg 23. Tibbo ME, Wyles CC, Fu S, et al. Use of Natural Language Processing Tools to Identify and Classify Periprosthetic Femur Fractures. J Arthroplasty [Internet] 2019;34(10):2216–9. Available from: https://doi.org/10.1016/j.arth.2019.07.025 24. Wyles CC, Tibbo ME, Fu S, et al. Use of Natural Language Processing Algorithms to Identify Common Data Elements in Operative Notes for Total Hip Arthroplasty. J Bone Jt Surg - Am Vol 2019;101(21):1931–8. 25. Finkbiner MJ, Gaina KM, McRandall MC, et al. Video movement analysis using smartphones (ViMAS): A pilot study. J Vis Exp 2017;2017(121):1–7. 26. Zheng G, Nolte L-P. Computer-Aided Orthopaedic Surgery: State-of-the-Art and Future Perspectives. Adv Exp Med Biol 2018;1093:1–20. 27. Ramkumar PN, Muschler GF, Spindler KP, Harris JD, McCulloch PC, Mont MA. Open mHealth Architecture: A Primer for Tomorrow’s Orthopedic Surgeon and Introduction to Its Use in Lower Extremity Arthroplasty. J Arthroplasty [Internet] 2017;32(4):1058–62. Available from: http://dx.doi.org/10.1016/j.arth.2016.11.019 28. Shaha SH, Sayeed Z, Anoushiravani AA, El-Othmani MM, Saleh KJ. Big Data, Big Problems: Incorporating Mission, Values, and Culture in Provider Affiliations. Orthop Clin North Am 2016;47(4):725–32. 29. Wouda F, Hutink H. Artificial Intelligence in de zorg [Internet]. 2019. Available from: https://www.nictiz.nl/rapporten/artificial-intelligence-in-de-zorg/ 30. Keskinbora KH. Medical ethics considerations on artificial intelligence. J Clin Neurosci [Internet] 2019;64:277–82. Available from: https://doi.org/10.1016/j.jocn.2019.03.001