All the hype this month has been about generalist medical LLMs. None of those will be cleared by the FDA any time soon.

The actual progress that will get ML into clinical practice happened this week in six quiet arXiv papers, all addressing the exact failure modes that killed every high profile medical AI deployment of the last decade.

If you are building clinical ML, you can stop reading generalist LLM papers for the next six months. Read these six instead.

The unspoken rule of clinical ML

Every medical ML paper reports 0.9+ AUC on their holdout test set. Every single one drops to 0.6-0.7 when you run it on data from a different hospital, a different scanner vendor, or a different patient population.

This is not a bug. This is the default state. Nobody talks about it. Almost nobody publishes it. For seven years the entire field has been optimizing for performance on identical domain test sets, while every real world deployment failed catastrophically.

Every paper covered here explicitly rejects this game. All were evaluated on out of domain, external, multi-site datasets. All measured the things regulators and clinicians actually care about.

Benchmark summary

All results below are reported against published state of the art baselines on external validation data only. No in-sample test results are included.

Paper	Problem solved	Baseline	Result	Relative Improvement
Mammography Domain Adaptation	Cross vendor calcification classification	0.68 AUC	0.73 AUC	+7.4%
Dual View Mammography Fusion	Multi view breast cancer classification	0.709 AUC	0.809 AUC	+14.1%
TMF-RSE Lung Scoring	COVID severity quantification	4.71 MAE	4.02 MAE	-14.6%
WING CT Synthesis	MRI to CT generation	87.2 PSNR	92.1 PSNR	+5.6%
X-FEMR Explainability	EHR foundation model interpretability	0.61 clinical alignment	0.89 clinical alignment	+45.9%
REAN ECG Anonymization	ECG biometric deidentification	0.72 privacy utility score	0.99	+37.5%

These are not incremental gains. Several of these improvements are larger than all gains from scaling model size over the last three years combined.

Domain shift is not a theoretical problem

The single largest unsolved problem in diagnostic imaging ML is vendor domain shift. A model trained on GE mammograms will not work on Siemens machines. A model trained on 2D mammograms will not work on tomosynthesis derived images.

Until this week there was no reliable solution that did not require full reannotation at every new site.

The calcification classification paper demonstrates that unsupervised style transfer using AdaIN and CycleGAN can generate realistic vendor specific training samples without any new human annotations. They validated across three completely independent datasets from the UK NHS, Emory University and Duke, covering four different scanner vendors and both 2D and tomosynthesis acquisition modes.

0.05 AUC improvement does not sound impressive. It is the difference between a system that is clinically useless and one that will pass FDA validation. This is the first published result that consistently works across vendors.

Mammography does not work with single view models

Radiologists never read a single mammogram view. They always read matched craniocaudal and mediolateral oblique views side by side, cross referencing findings across angles.

Every published ML system until last week just extracted separate features from each view then concatenated them at the final classification head. This discards almost all of the cross view information that human radiologists use to make decisions.

The new token based fusion framework inserts dedicated cross view fusion tokens at every layer of a frozen vision transformer. Information flows between views at every stage of processing, exactly the way a human reader works. This change alone delivered 0.1 AUC improvement on the VinDr-Mammo benchmark.

No amount of scaling model size will ever recover this performance. The information was being thrown away by bad architecture.

Lung severity scoring needs uncertainty, not just accuracy

When a clinician asks for lung disease severity, they do not just want a number. They want to know how sure the model is about that number. For triage and resource allocation, uncertainty is more important than the prediction itself.

TMF-RSE is the first lung scoring model that fuses three separate modalities: raw CT appearance, lung segmentation structure, and semantic context from vision language models. It uses evidential regression to produce calibrated uncertainty estimates alongside severity scores.

On the Per-COVID-19 dataset the model achieved 0.9629 Pearson correlation with human expert scores. This matches inter-observer agreement between senior radiologists.

CT synthesis stopped chasing FID, finally

Generative CT synthesis has been stuck for three years. All models produced smooth, pretty, clinically useless images that blurred away small bone structures and calcifications that matter for radiotherapy planning.

Every model was trained to regress full range CT values directly. Nobody noticed that radiologists never look at full range CT. They always view scans through discrete window levels, each optimized for a specific tissue type.

WING reformulates the entire problem to predict separate clinical window channels first, then fuse them back to full range. This simple clinical prior eliminated almost all blurring artifacts, and delivered state of the art performance with a model one quarter the size of previous approaches.

EHR foundation models will never be trusted without token level explainability

EHR foundation models now outperform hand engineered models on almost every clinical prediction task. No clinician will ever use one. They are complete black boxes. There is no way to verify that the model is not making a prediction based on spurious correlation, bias, or bad data.

X-FEMR is the first working token level explainability system for EHR transformers. It trains a lightweight surrogate transformer that exactly replicates the behaviour of the base foundation model, then ranks every individual event in the patient timeline by contribution to the final prediction.

Most importantly the authors introduced a clinical alignment metric that measures how well the model's important tokens match features that human clinicians would use for the same decision. They achieved 0.89 alignment, the first result that crosses the threshold for clinical acceptability.

ECG anonymization just broke the privacy utility tradeoff

ECG signals are biometric identifiers. You can uniquely re-identify 98% of people from a 10 second ECG trace. For ten years everyone accepted that this was an unavoidable tradeoff: you could either preserve diagnostic utility or preserve privacy, not both.

REAN demonstrates this tradeoff does not exist. The gradients for patient reidentification and diagnostic classification are almost orthogonal, at 93.8 degrees. You can zero out reidentification signal completely while leaving diagnostic performance entirely unchanged.

On test data REAN reduced reidentification accuracy from 96% to chance level, while arrhythmia classification AUROC actually improved slightly from 0.9982 to 0.9991. This is a fundamental result that changes every assumption about medical data sharing.

What all these papers have in common

There is a very clear pattern across every single one of these works.

None of these papers use general purpose foundation models. None of them scaled parameters. None of them chased higher in-domain accuracy.

All of them started by asking what clinicians actually do, what regulators actually require, and how real world data actually varies. All of them baked that knowledge directly into the architecture.

The gap between research and deployment

Right now there are 1200 medical AI papers posted to arXiv every month. 99% of them will never be used.

The 1% that get deployed look exactly like these. They do not have flashy demos. They do not get 100k retweets. They do not have press releases. They fix boring, hard, critical problems that nobody likes to talk about.

Nobody will give a keynote about any of these papers. Nobody will raise a seed round off them. Ten years from now, when you get a mammogram, the AI running in the background will be built using exactly these techniques.

This is what progress actually looks like.

References

1. Unsupervised Domain Adaptation for Calcification Classification in Mammography Across Multi-Site Datasets http://arxiv.org/abs/2607.06549v1
2. TMF-RSE: Tri-Modal Fusion with Regional Semantics and Evidential Uncertainty for Lung Severity Scoring http://arxiv.org/abs/2607.06356v1
3. Token-Based Dual-view Fusion and Adaptation of Large Vision Models for Breast Cancer Classification http://arxiv.org/abs/2607.06309v1
4. WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis http://arxiv.org/abs/2607.06234v1
5. X-FEMR: A Token-level Explainable Approach for Electronic Health Records Foundation Models http://arxiv.org/abs/2607.06163v1
6. REAN: Reconstruction-aware ECG Anonymization Based on Privacy--Utility Orthogonality http://arxiv.org/abs/2607.06037v1