Healthcare & Digitalization

The digitalization of the healthcare sector turned out to be a double-edged sword. On the one hand, it brought us impressive advances in medical imaging and opened the door to effective remote health monitoring. On the other hand, in some cases, instead of improving healthcare provision, digitalization has led to a significant productivity drain. Take electronic health records (EHR), for example. The way EHRs are designed and used these days is not only linked to physician burn-out and decreased work-life satisfaction [1] but also associated with physician turnover [2]. Considering these detrimental circumstances, it comes as no surprise that the term “digitalization” left a sour taste in the mouths of many healthcare professionals.

A new generation of intelligent decision support systems

Still, there is a lot of potential. From the perspective of a machine learning (ML) practitioner, capturing patient visits, treatments, and diagnoses digitally is the foundation for a new generation of more reliable and accurate decision support systems. Such machine learning systems are able to efficiently learn from large troves of historic patient data to answer medical questions with unprecedented speed and accuracy. The possibilities of ML-based applications in healthcare are endless and range from the early detection of sepsis [3, 4] to the automatic identification and classification of heart arrhythmias [5] to the prediction of in-hospital mortality [6]. Improving diagnostic accuracy is only one way in which ML can be utilized in the medical domain. There is great potential in using ML to improve healthcare operations (e.g., predicting wait times/delays and determining their causes [7]) and to support physicians in mundane time-consuming documentation tasks. Figure 1 shows three broad areas within healthcare that we think will benefit the most from ML.

Figure 1 | Three fundamental areas within healthcare which can benefit from ML9.

Challenges of machine learning adoption in healthcare

While more and more industries discover the value that reliable ML systems can add to their business, the adoption in healthcare is slow. This is not because healthcare problems are hard to solve (even though they are) but it primarily hinges on factors revolving around

  1. interoperability,
  2. mindset,
  3. communication, and
  4. systems engineering.


Interoperability refers to the ability of different systems to communicate with each other. In the clinical context, the highest level of interoperability (semantic interoperability) means that systems, whether they’re part of the same organization or not, can seamlessly exchange information (provided relevant privacy rights are respected). A fully interoperable healthcare system would allow all stakeholders to always have the most recent view on a patient. In the context of machine learning, interoperability enables algorithms not only to generate the most accurate predictions but also to communicate them with all involved parties. Even though interoperability is one of the most pressing issues in medical informatics, it is incredibly hard to realize a fully interoperable healthcare network. Luckily, interoperability is not a hard requirement for machine learning to generate clinical value. The right mindset, however, is.


Clinical decision support systems (CDSS) have been around since the 1980s [8] and while they were built to support clinicians, they frequently did the opposite, leading to fragmented workflows and alarm fatigue. It is natural that skeptics see machine learning as “yet another decision support system”. The difference between the 80s and today is that a rigorous science (data science) has been developed around the pain points from the past (e.g., alarm accuracy, data quality, interpretability). Additionally, holistic approaches such as our own effective MLOps framework, stress the importance of being more user-centered. In other words, not only can we build better CDSSs, we co-create them together with the end users such as physicians, nurses, administrative staff, or the patients. We should therefore embrace the advances in machine learning (i.e., more accurate prediction systems) and view them as a chance to build effective support systems, so healthcare professionals can focus on what counts: the patient. Finally, we must get comfortable with the fact that by introducing new intelligent systems, clinical workflows must be reinvented or adapted.


Finding a common language is critical, especially when it comes to expectation management. First, the data scientist must understand the clinical context and its challenges. This concerns the workflow in which the system’s predictions will be embedded as well as the performance metrics that are of clinical relevance. For example, if the priority is to reach a reliable “rule-out” decision, our system must have high specificity and a high negative predictive value. Second, the healthcare professional should understand both the possibilities and limits of machine learning. There are many success stories (mostly of academic nature) which make you believe ML can solve even the hardest medical problems. And while many challenging problems are being solved with unprecedented fidelity, it can take years before such a level of maturity can be reached; there is no silver bullet.

Systems Engineering

Lastly, machine learning adoption in healthcare is being slowed down by the fact that many projects don’t make it past the proof-of-concept stage. Even if we can demonstrate the clinical benefit of a system, many obstacles remain before it can be used in clinical everyday life. Some are of organizational nature (e.g., how will healthcare professionals use the system) and some are of technical nature. The latter refers to challenges such as changes in the patient cohort while running the new system (maybe the system was developed on data from adults, but now we want to use it for adolescents), poor data quality (maybe the system was developed on clean and maintained in-patient data but we want to use it in ambulatory care), or the speed at which a prediction is provided. These and many other challenges can be approached by following so-called machine learning operations (MLOps) practices.


There is no denying the fact that machine learning has the potential to revolutionize healthcare. However, it is a rocky path that necessitates not only technical progress but also rethinking and proactive partnerships. To circumvent some of the challenges we identified, it is helpful to start clinical projects with ML experts experienced in the healthcare sector (like our team at MLAB). This eases the frequently experienced “clash of cultures” of tech and clinical experts and prevents major communication difficulties. Additionally, both sides should be interested and capable of going beyond the proof-of-concept stage. From a technical point of view, it becomes inevitable to understand and practice both software reliability engineering (SRE) and MLOps principles. In an upcoming blog post, we will dive deeper into the interaction between MLOps and healthcare.

Until then, feel free to reach out to discuss the possibilities that machine learning can bring to your healthcare organization.

References and Acknowledgements

  1. Robertson, S. L., et al. "Electronic health record effects on work-life balance and burnout within the I3 population collaborative". [link]
  2. Melnick, E. R., et al. "Analysis of electronic health record use and clinical productivity and their association with physician turnover". [link]
  3. Moor, M., et al. "Early prediction of sepsis in the ICU using machine learning: a systematic review". [link]
  4. Sendak, M. P., et al. "Real-world integration of a sepsis deep learning technology into routine clinical care: implementation study". [link]
  5. Hannun, A. Y., et al. "Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network". [link]
  6. Ponce, D., et al. "Development of a prediction score for in-hospital mortality in COVID-19 patients with acute kidney injury: a machine learning approach". [link]
  7. Pianykh, O. S., et al. "Improving healthcare operations management with machine learning". [link]
  8. Sutton, R. T., et al. "An overview of clinical decision support systems: benefits, risks, and strategies for success".[link]
  9. Icons created by Flat Icons, Freepik, and monkik - Flaticon