Why Almost No Machine Learning Models Ever Reach the Hospital Bedside

Medical journals now hold tens of thousands of studies describing clinical machine-learning models, many of them promising a future of individualised, precision medicine. Yet an editorial published in Frontiers in Medicine argues that the field’s most striking feature is not how much has been built, but how little of it has ever been used. Relatively few models have entered routine care at the bedside, and when a deployment does happen, it is still reported as a significant achievement in its own right. The authors, a group of clinicians and researchers from institutions in Canada and the United States, contend that the usual explanation for this mismatch, a so-called translational lag, is too generous to the field.

The idea of a translational lag implies a temporary delay, one that better engineering, improved infrastructure, or clearer regulation will eventually close. The editorial’s authors reject that framing. In many studies, they argue, translation is not merely postponed but was never part of the research plan at all. Model development routinely ends at retrospective validation, even when the stated objective of the work is to improve real-world clinical care. On that reading, the gap between the computer and the bedside is not a bottleneck in an otherwise healthy pipeline; it is a structural feature of how clinical machine-learning research is currently designed and rewarded.

The scale of the imbalance is quantified in the editorial through a series of recent reviews. By 2024, nearly 250,000 publications describing new clinical prediction models had been reported in the literature. Yet among a sample of 10,462 machine-learning algorithms captured across 220 systematic reviews, external validation, the step in which a model is tested on data beyond the dataset it was built on, was reported in fewer than 1% of models. The evidence thins further along the pathway. A 2024 scoping review identified only 86 randomised controlled trials evaluating clinical machine-learning systems, with a predominance of single-centre studies and outcomes focused on diagnostic performance rather than demonstrable benefit to patients.

At the far end of the pipeline, the numbers shrink to almost nothing. A 2025 systematic review examining real-world implementations of prognostic prediction models found only 37 publications describing 56 models actually in use in practice. The editorial is careful to note that these reviews differ in scope and methodology, so their figures cannot be simply added together, but the pattern they consistently show is unambiguous: the development of clinical machine-learning models substantially outpaces the evidence of their clinical application. Each successive stage of translation, from internal validation to external validation, shadow testing on live clinical data, prospective trials measuring patient outcomes, deployment, and post-deployment monitoring, retains a progressively smaller fraction of the models that began the journey.

To probe why, the editors assembled a Research Topic of six articles examining different points along the model development and translation pathway. The first and perhaps most fundamental issue concerns problem selection. Ideally, machine-learning models should be designed to facilitate specific decisions at key inflection points in patient care. In practice, many studies choose their prediction targets based on convenience, simply using whatever outcomes are available in administrative or electronic health record datasets. The result can be predictions that no one can act on, offering no tangible benefit to patients and leaving little motivation to carry the model forward into clinical testing.

A perspective article by Vinnakota and colleagues highlighted in the editorial argues that investigators should ask not simply whether an outcome can be predicted, but whether the prediction resolves a meaningful clinical uncertainty and can lead to an actionable change in management. A second contribution, by Mo and colleagues, applies a related logic to a different domain: using machine learning to support the assessment of research trustworthiness when conducting systematic reviews. In that setting, algorithmic outputs are used to flag potentially unreliable studies for expert review rather than to replace human judgement. Although the two articles address very different applications, the editorial draws the same lesson from both: the intended role of a model within an existing decision-making process must be defined from the outset. For clinical models, that means specifying where predictions should enter the care pathway, which decisions they are expected to inform, and whether they are meant to supplement or replace the role of clinicians.

A second barrier is generalisation. Two articles in the Research Topic, by Li and colleagues and by Cao and colleagues, largely follow conventional model development pathways but extend their evaluation to include external validation. Their inclusion is notable precisely because external validation remains uncommon in the literature. The exercise matters because it provides evidence that a model’s performance is not confined to the dataset on which it was derived, and it can reveal deterioration in performance when models are applied to different patient populations or different clinical settings. A model that looks impressive on its home data may quietly fail elsewhere, and without external testing neither developers nor clinicians can tell the difference.

Yet even external validation is still a retrospective exercise, and it cannot answer whether a model can function inside a real clinical environment. Questions such as whether the variables a model requires are actually available at the point of care, whether predictions can be delivered in an interpretable form, and whether clinicians will incorporate those predictions into their decisions can only be answered by studying the setting in which the model is expected to operate. Two further articles in the topic took on exactly these implementation questions. Klooster and colleagues examined stakeholder perspectives on predictive models for diabetic foot complications and identified concerns surrounding workflow integration, the presentation of predictions, clinical responsibility, and alert burden. Xie and colleagues studied pharmacists’ adoption of artificial intelligence-based clinical decision support tools and found that adoption depended not only on perceived usefulness but also on trust, as well as organisational and technical support. These considerations, the editorial observes, are often deferred until after a model has been developed, despite being key to determining whether deployment would be feasible at all.

Strikingly, most of the articles in the Research Topic did not address the later stages of evaluation and deployment, such as shadow testing or prospective trials. The editorial is explicit that this does not diminish the contributions themselves; rather, it mirrors the broader gap in the literature. The reasons are likely practical. Retrospective validation can be completed using existing datasets within conventional research timelines, whereas shadow testing and prospective studies demand substantially greater coordination and resources. At minimum, investigators would need to work with healthcare institutions to integrate models into existing clinical systems, obtain the necessary approvals, and maintain those systems throughout the evaluation period. Prospective studies may also expose problems that were invisible during model development, as the implementation studies by Klooster and colleagues and Xie and colleagues illustrate. The prospect of uncovering such problems, the authors suggest, may itself discourage investigators from pursuing prospective designs, creating a self-reinforcing cycle in which the hardest and most informative studies are the least likely to be done.

The editorial closes with two warnings. First, translation does not end with successful deployment: shifts in patient populations, clinical practice, and the underlying data can degrade model performance over time, so healthcare institutions must establish clear responsibilities for post-deployment surveillance, including regular assessment of model calibration, data drift, and performance across clinically relevant patient subgroups, with predefined protocols guiding recalibration or controlled updating and subsequent validation before changes take effect. Second, and more broadly, the six articles together address only small parts of the translation process, and they show that the barriers separating model development from clinical use are not a single technical problem that better algorithms can solve. Translation can fail because the wrong problem was selected, because a model does not generalise, because it cannot be fitted into clinical workflow, because intended users do not trust or adopt it, or because its output has no clearly defined role in decision-making. Progress, the authors conclude, will depend less on producing additional retrospective models and more on designing research programmes that confront these questions from the very beginning. Without that shift, they warn, clinical machine-learning models risk remaining expensive statistical exercises rather than the paradigm-shifting tools for individualised and precision medicine they are frequently claimed to be.

Subject of Research: Barriers to translating clinical machine-learning models from development into routine 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