Appearance
Clinical AI scores well and still misses the patient in front of it
Why a 90% score stops mattering at phase three
A model can stage a nasopharyngeal tumor correctly and then say nothing about the follow-up scan that decides whether the patient gets salvage radiation. Both are clinical reasoning. Only one shows up in most published numbers.
That's the thread running through five recent preprints on clinical AI. Three are about measuring the right thing: a phase-aware rubric benchmark for cancer care, a masking study for multimodal prognosis, and a multimodal breast biopsy model. Two are about making the internals legible: a patch-attributing ViT and a position paper on decentralized clinical agents. Different modalities, different diseases. Same failure, where the number you report and the behavior you need drift apart.
NPC-RubricBench grades the pathway, not the encounter
Most clinical LLM evaluations score a single encounter. Here's a vignette, here's a question, did the model answer it. Cancer care doesn't work that way. It runs through work-up, staging, treatment planning, delivery, and years of surveillance, and an error at any step compounds into the next one.
NPC-RubricBench is built from longitudinal records of 405 patients at two cancer centers. It slices those histories into 941 patient-phase units across six sequential care phases. Each unit gets scored against atomic criteria generated for that specific patient, 17,688 of them across five evaluation dimensions. The design detail that matters most: every assessment is restricted to evidence available at decision time. No peeking at the pathology report that came back the next week.
| Property | Value | What it changes |
|---|---|---|
| Patients contributing records | 405 | Covers phase transitions, not just initial presentations |
| Cancer centers | 2 | Two institutions means two documentation styles, which is where most evals break |
| Patient-phase units | 941 | The scoring unit is a decision point, not a chat turn |
| Sequential care phases | 6 | One model gets graded six times on the same patient history |
| Patient-specific atomic criteria | 17,688 | About 19 checks per unit, written for that patient instead of a generic checklist |
| Evaluation dimensions | 5 | Decision quality is one dimension among several, not the whole score |
| Models evaluated | 30 | Mix of general-purpose and medical LLMs |
Seventeen thousand criteria is what keeps this from collapsing into a multiple-choice exam. "Did the model recommend the right next step for this patient given what was known on that date" is a different question from "did it pick option B."
Accuracy doesn't carry over, and neither do the errors
The headline from the 30-model sweep: strong diagnostic performance did not translate into better downstream care. Models that nailed staging fell apart later, and the dominant failure mode flipped depending on the phase.
At anatomical staging, explicit decision errors led at 48.2%. The model said something, and the something was wrong. In initial work-up and post-treatment follow-up, decision omissions led at 33.4% and 30.4%. The model said nothing at all.
Those two modes need different mitigations. A wrong recommendation can trip a rule, a second read, or a guideline check. A missing recommendation passes in silence. Nothing in a chat window tells the clinician that the model forgot to order the scan.
The result I didn't expect: in the evaluated set, medically adapted models showed higher omission rates than general-purpose ones. My read is that instruction tuning on medical question-answering rewards producing an answer to the question in front of you. It doesn't reward volunteering the second step nobody asked about. If that read holds, the fix is data and objective design, not more specialty fine-tuning.
Prompt granularity trades decision quality for faithfulness
You'd expect tighter prompts to improve everything. They don't. Broad open-ended requests degraded Decision Quality across all 30 models. Checklist-like prompts raised Decision Quality, then pulled down Evidence Faithfulness and Safety Adherence.
That is a specification trade-off, and you only find it by varying the prompt as part of the evaluation instead of fixing it once. Ask for a checklist and you get a checklist, including items the record can't support. Ask for free reasoning and you get careful evidence attached to an incomplete plan.
The practical move is to pick prompt style per phase. Work-up and follow-up, where omissions dominate, want structured prompts that enumerate required actions. Staging, where wrong calls dominate, wants prompts that force evidence to be stated before a commitment is made.
Quick Take: a model that reads a chart well and a model that manages a care pathway are different products, and only the second one belongs near a clinic.
Explanations that cite tests nobody ordered
The dementia prognosis paper tests something the field has mostly assumed: that explanation narratives stay honest when inputs are incomplete. Amyloid PET and CSF biomarkers are expensive, and plenty of patients in the Alzheimer's Disease Neuroimaging Initiative never got them.
The setup: 307 participants with complete data, predicting conversion from mild cognitive impairment to Alzheimer's within 36 months. The authors masked every subset of four costly modalities, then restored the masked data to obtain interventional ground truth. The restoration step is the good part. It lets you ask whether the explanation would change with the missing test in hand, and check that answer against what actually happens when you add it.
All five missing-data strategies assigned 12% to 23% of attribution to modalities that were never observed. With four absent, that climbed to 45%. Missingness-aware strategies cut it without any accuracy cost, which makes the 12% to 23% baseline a design flaw rather than an unavoidable price.
Then the narration layer: 51,200 reports across four language models and four prompting regimes. Telling the narrator which modalities were absent raised acknowledgement from 48% to 88% and cut phantom evidence twenty-fold.
Key numbers from the missing-modality study12% to 23% of baseline attribution went to modalities that were never observed, rising to 45% with all four absent. 51,200 narratives generated across four language models and four prompting regimes. 48% → 88% acknowledgement of missing inputs after naming absent modalities in the prompt. 27% to 41% accuracy when narrators guessed whether acquiring a missing test would change the prediction, versus 69% for open-weight narrators and 99.9% for one proprietary model when answers were checked against executed restorations.
The number that decides deployability is that last one. A free-text narrator asked "would this test change your mind" is guessing. The 99.9% figure comes from handing the narrator the counterfactual at inference time, which means running the model again with the modality filled in. Nothing in prompt wording gets you there.
| Intervention | Measured effect | What it doesn't fix |
|---|---|---|
| Naming absent modalities in the prompt | Acknowledgement 48% → 88%, phantom evidence cut 20x | Counterfactual advice still correct only 27% to 41% of the time |
| Checking advice against executed restorations | Advice accuracy to 69% for open-weight narrators, 99.9% for a proprietary model | Requires a second model run per question |
| Structured report template | Eliminated the case where narrators declared observed modalities absent | Direction fidelity, which tracked narrator scale rather than prompting |
One more finding worth internalizing: higher phantom attribution went with more phantom evidence and fewer faithful direction claims. The attribution and the narration fail together, so you can't patch one layer and call the explanation trustworthy.
Breast biopsy: what 0.72 AUC actually buys
The breast study is the one closest to a product. ViT for imaging features, DeBERTa-V3-Large for the radiology report, late fusion, Grad-CAM for heat maps. Trained and tested on 1,632 in-house patients, four mammography views plus a report each.
| Operating point | Sensitivity | Specificity | What it means per 100 benign biopsies worked up | What it means per 100 cancers |
|---|---|---|---|---|
| Conservative | 90% | 40% | About 40 spared | About 10 missed |
| Balanced | 95% | 36% | About 36 spared | About 5 missed |
| Aggressive | 99% | 18% | About 18 spared | About 1 missed |
| Maximum catch | 100% | 11% | About 11 spared | None missed |
The 90/40 row is the one I'd actually consider deploying. Multimodal AUC was 0.72 on test, against 0.66 for image-only and 0.68 for text-only. A six-point lift from adding the report says the text carries real signal, and also that neither modality is close to solving the problem. Late fusion is part of why: the text embedding never gets to redirect where the image encoder looks, so the two streams disagree in private and meet at the classifier.
Grad-CAM on a ViT produces patch-grid heat, which on a full-field mammogram means blobs spanning centimeters. Useful for pointing a reader toward a quadrant. Not useful for telling them a 4 mm spiculation is the reason.
SAP-ViT puts attribution in the forward pass
Standard ViT classification reads a global CLS token, and interpretability comes from attention maps. Attention rollout is a proxy for what the classifier used. It isn't a decomposition of the decision, and radiologists notice the difference fast.
SAP-ViT drops the CLS token and aggregates last-layer patch embeddings directly. The class prediction becomes a sum over patches, so per-patch contributions fall out of the forward pass instead of being reconstructed afterward. Sparsity regularization on the patch-weight distribution keeps the evidence short enough to read. Because self-attention mixes information across tokens, the authors add an input-space back-mapping step to recover which input patches drive the influential output patches.
Across several medical imaging datasets and modalities, classification performance holds while patch-level explanations stay consistent. Ablations confirm both the sparsity term and the back-mapping alignment contribute to explanation quality, so neither is decoration.
If you've ever had a radiologist point at a Grad-CAM blob and ask what it means, the appeal is obvious. A per-patch contribution you can read off the forward pass is auditable. A saliency overlay is a guess about the guess.
The limit: patch-level attribution is not lesion-level attribution. A 16×16 grid still tells you which tiles mattered, not which finding.
Decentralized agents swap gradients for reasoning traces
This one is a position paper with no experiments, so treat the claims accordingly. The argument: clinical expertise is distributed across thousands of clinicians working with different populations and resource constraints, and current scaling approaches treat knowledge as static parameters to aggregate. Federated learning averages gradients across homogeneous architectures, which requires everyone to run the same model and throws away the part that differs between hospitals, namely the sequence of decisions.
Decentralized Agentic Systems exchange abstracted reasoning trajectories instead: observation, hypothesis, tool invocation, outcome feedback. No raw data, no weights.
| Centralized training | Federated learning | Decentralized agentic systems | |
|---|---|---|---|
| What moves | Raw data | Gradients | Reasoning trajectories |
| Architecture must match | Data schema | Yes | No, agents can differ |
| Captures local workflow | No | Partly | Yes, that's the thesis |
| Published evidence | Extensive | Moderate | None yet |
| Main governance risk | Record-level leakage | Gradient inversion | Trajectory leakage, which is denser than gradients |
The governance risk is the part I'd push on hardest. A reasoning trace contains hypotheses, drug names, doses, and timelines. Abstraction helps, but "abstracted" is doing a lot of work in that sentence, and until someone publishes a leakage study on trajectory exchange, the privacy comparison to federated learning is unproven.
What trips people up
Reporting one score for a model. If your evaluation number comes from a single task, it says nothing about phase three of a care pathway, where 30.4% of failures were silent omissions. Grade per phase or don't claim pathway competence.
Treating an omission as a softer error than a wrong answer. Omissions are harder to catch precisely because nothing looks broken. A wrong staging call gets flagged in tumor board. A missing follow-up recommendation leaves the room quietly.
Grading explanations by how plausible they read. The missing-modality study shows a 12% to 23% phantom attribution floor across five standard strategies. If you haven't masked inputs and checked what the explanation does, you don't know your faithfulness number.
Assuming medical fine-tuning buys safety. In the rubric benchmark, specialty-adapted models omitted more than general-purpose ones. Specialty tuning buys vocabulary and format compliance, not pathway discipline.
Asking a narrator whether a test would change the prediction without running the counterfactual. 27% to 41% accuracy is close to a coin flip with a bias. Re-run the model with the modality restored and you get 69% minimum.
One Thing to Remember
Every one of these papers is measuring the same thing from a different angle: what a system does when the evidence it needs isn't there yet. The missing phase, the missing scan, the missing report, the missing patch. Retrofitting explanations and prompts doesn't fix that, because the counterfactual has to be executed somewhere. Budget for the second inference call, the phase-wise eval set, or the modality mask, and put it in the design instead of the post-hoc analysis.
The Bottom Line
If you're deploying an LLM into a care pathway, evaluate per phase and expect the failure mode to flip between them. A model that hits 90% on staging can still miss a third of the required work-up actions, and only a phase-split eval surfaces that before a patient does.
If your multimodal model runs on incomplete records, pay for missingness-aware attribution and structured narration templates, and skip free-text rationales for any question that asks whether a missing test would change the answer. Those narratives run 27% to 41% accurate until you execute the restoration, and the executed version costs one extra forward pass.
If you're choosing between a Grad-CAM overlay and a decomposable classifier for a regulatory or audit-facing use case, pick the decomposable one, because patch contributions read out of the forward pass survive scrutiny that saliency maps don't. And if you're at BI-RADS 4/5 triage with 0.72 AUC, run it as a second reader that can only escalate to biopsy, never clear a patient.
One thing to watch: decentralized reasoning exchange will get pilots within a year, most likely in pathology or radiology networks with existing data-sharing agreements. Watch for the first leakage study on reasoning traces, not the first accuracy claim. That paper will decide whether the approach has a future.
Source references
- Patient-specific rubric-based evaluation of large language models across longitudinal cancer care. arxiv.org/abs/2610.35847
- Beyond Federated Learning: Decentralized Agentic Systems for Self-Evolving Clinical Intelligence. arxiv.org/abs/2610.35453
- Faithful Explanations under Missing Modalities in Multimodal Dementia Prognosis. arxiv.org/abs/2610.35430
- Explainable Multimodal AI Approach to Reduce Unnecessary Breast Biopsies. arxiv.org/abs/2610.35704
- SAP-ViT: A Self-Attributing Patch Vision Transformer for Medical Image Classification. arxiv.org/abs/2610.35593