Why Almost No Machine Learning Models Ever Reach the Hospital Bedside

Medical journals now contain tens of thousands of studies about machine-learning models for healthcare. Many promise a future where medical care is tailored to each person. But an editorial published in Frontiers in Medicine argues that the biggest story is not how many models researchers have built, but how few are actually being used. Only a small number have become part of everyday patient care. When a model does reach the bedside, it is still treated as a major achievement. The authors, a group of doctors and researchers from institutions in Canada and the United States, argue that the usual explanation for this problem, called a "translational lag," gives the field too much credit.

The term "translational lag" suggests that research is simply taking time to reach patients, and that better engineering, improved systems, or clearer rules will eventually solve the problem. The editorial's authors disagree. They argue that many studies never planned to put their models into real-world use in the first place. Researchers often stop after testing a model on past data, even when they say their goal is to improve real patient care. From this point of view, the gap between a computer model and the hospital bedside is not just a delay in an otherwise working process. It is built into the way clinical machine-learning research is designed and rewarded.

The editorial uses several recent reviews to show how large this gap is. By 2024, the scientific literature had reported nearly 250,000 publications describing new clinical prediction models. Yet a sample of 10,462 machine-learning algorithms covered by 220 systematic reviews found that fewer than 1% of models had been tested on external data. External validation means checking a model using data other than the data used to build it. The evidence becomes even more limited at later stages. A review published in 2024 found only 86 randomised controlled trials that tested clinical machine-learning systems. Most took place at a single centre, and they mainly measured how well the systems made diagnoses rather than whether they actually helped patients.

At the final stages of the process, the numbers become extremely small. A 2025 systematic review of prediction models used to forecast future health outcomes found only 37 publications describing 56 models that were actually being used in real-world care. The editorial points out that these reviews cover different topics and use different methods, so their numbers cannot simply be added together. However, they all show the same pattern: clinical machine-learning models are developed far faster than they are put into practice. At each step, fewer models continue. These steps include testing a model on the data used to build it, testing it on outside data, running it quietly on live clinical data, conducting prospective trials to measure patient outcomes, putting it into use, and monitoring it after launch.

To understand why this happens, the editors brought together six articles that examine different stages of developing machine-learning models and bringing them into practice. The first and perhaps most basic problem is choosing the right question to solve. Ideally, models should help with specific decisions at important moments in patient care. In reality, many studies choose what to predict simply because the information is easy to find in administrative records or electronic health records. This can lead to predictions that nobody can act on. Such predictions offer no clear benefit to patients and give researchers little reason to continue testing the model in clinical settings.

A perspective article by Vinnakota and colleagues, discussed in the editorial, argues that researchers should ask more than whether they can predict an outcome. They should also ask whether the prediction answers an important medical question and can lead to a useful change in care. A second article, by Mo and colleagues, applies a similar idea to another area: using machine learning to help assess whether research studies can be trusted during systematic reviews. In this case, algorithms flag studies that may be unreliable so experts can examine them. They do not replace human judgement. Although these two articles focus on very different uses, the editorial draws the same lesson from both: researchers must decide from the beginning how a model is supposed to fit into an existing decision-making process. For clinical models, this means deciding where predictions will be used in patient care, which decisions they should help guide, and whether they should support or replace decisions made by clinicians.

A second barrier is whether a model works well beyond the data and setting where it was created. Two articles in the Research Topic, by Li and colleagues and by Cao and colleagues, follow common model-building methods but also test their models using external data. This is important because external validation is still rare in published research. It helps show whether a model works beyond the dataset used to develop it. It can also reveal when performance drops with different groups of patients or in different healthcare settings. A model may appear to work very well on the data it was built with but fail elsewhere. Without outside testing, neither its developers nor doctors can know whether this is happening.

However, even external validation looks back at existing data. It cannot tell researchers whether a model will work in a real healthcare setting. To answer that question, researchers must study the place where the model is meant to be used. They need to know whether the required information is available when care is being delivered, whether predictions can be shown in a way people can understand, and whether clinicians will use those predictions when making decisions. Two more articles in the Research Topic examined these practical questions. Klooster and colleagues studied what different groups thought about prediction models for diabetic foot complications. They found concerns about fitting the models into everyday work, how predictions should be displayed, who would be responsible for decisions, and how many alerts clinicians would have to handle. Xie and colleagues studied pharmacists' use of artificial intelligence tools that support clinical decisions. They found that pharmacists' willingness to use these tools depended not only on whether they seemed useful, but also on trust and the support provided by their organisations and technical systems. The editorial notes that researchers often leave these questions until after a model has been built, even though they can determine whether using it in practice is possible at all.

Notably, most of the articles in the Research Topic did not examine the later steps of testing and putting models into use, such as running them quietly alongside real clinical care or conducting prospective trials. The editorial makes clear that this does not make the articles themselves less valuable. Instead, their focus reflects the wider lack of research at these later stages. Practical difficulties are likely part of the reason. Researchers can test models on existing records within the usual time limits of a research project. Quiet testing and prospective studies require much more planning, teamwork, and funding. At a minimum, researchers must work with healthcare organisations to connect models to existing clinical systems, get the required approvals, and keep those systems running throughout the study. Prospective studies may also uncover problems that were not visible while the model was being developed, as the studies by Klooster and colleagues and Xie and colleagues show. The authors suggest that the risk of finding such problems may itself discourage researchers from running prospective studies. This can create a cycle in which the most difficult and most informative studies are also the least likely to happen.

The editorial ends with two warnings. First, putting a model into use successfully is not the end of the work. Changes in patient groups, medical practice, and the data used by a model can make its predictions less reliable over time. Healthcare organisations must therefore clearly assign responsibility for monitoring models after launch. This includes regularly checking whether predictions match real outcomes, whether the data have changed, and whether the model works well for different patient groups that matter in clinical care. Organisations should also set clear rules for when a model needs to be recalibrated or updated in a controlled way, and for testing it again before any changes take effect. Second, the six articles cover only small parts of the process of bringing research into practice. Together, they show that the barriers between building a model and using it in healthcare are not one technical problem that better algorithms can fix. Progress can fail because researchers chose the wrong problem, a model does not work well in other settings, it does not fit into clinical routines, its intended users do not trust or adopt it, or nobody has clearly explained how its predictions should guide decisions. The authors conclude that progress will depend less on building more models tested only on past data and more on planning research that tackles these challenges from the start. Without this change, they warn, clinical machine-learning models may remain costly statistical exercises instead of the major tools for personalised and precision medicine that they are often claimed to be.

Subject of Research: Barriers to bringing clinical machine-learning models from development into everyday patient care

Article Title: Editorial: Translating machine learning innovations into clinical practice

Article References: Deng, J., Zhou, F., Heybati, K., & Bozzo, A. (2026). Editorial: Translating machine learning innovations into clinical practice. Frontiers in Medicine, 13, Article 2003342. https://doi.org/10.3389/fmed.2026.2003342

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.2003342

Keywords: machine learning, clinical prediction models, external validation, knowledge translation, precision medicine, artificial intelligence, clinical decision support, model deployment, prospective trials, healthcare implementation, model generalisation, translational research