Decryptions
All scientific publication decryptions on Tatakoto.
Skin lesions: 96% accuracy and 0.99 AUC — on the three quarters of images the model agreed to decide
Two authors combine three networks (MaxViT-Tiny, ConvNeXt-Tiny, EfficientNetV2-B0) with MC Dropout uncertainty estimation to classify seven skin lesion types on HAM10000, and report 96% accuracy and 0.99 AUC. Those numbers are computed after discarding uncertain cases — the results table covers 1,508 images while a 20% test split holds 2,003, meaning roughly a quarter of cases were never decided, a coverage figure the paper never reports. The abstention rate is not uniform: reconstructed class by class, it reaches about one third on melanoma.
ICU sepsis: a reinforcement learning agent that lowers treatment as lactate rises, and that standard metrics score top of the class
Two engineers from Vanderbilt and Duke subject two reinforcement learning agents trained on MIMIC-III to a simple test: artificially raise lactate and watch which way the recommended treatment moves. The model that imitates clinicians best (MSE 0.0098) reduces treatment intensity by about 15% as lactate rises — Spearman correlation of −0.67, the reverse of Surviving Sepsis Campaign guidance. The diagnosis is right; the demonstration rests on a single random seed, with no confidence intervals, and the preprint never states how many patients it used.
Thyroid ultrasound: an auditable "agentic" AI tested across 35 centres, and the three numbers the abstract leaves out
Fourteen authors from Sun Yat-sen University, Zhujiang Hospital and Zhejiang present ThyroidXAgent, a system that chains nodule segmentation, benign-malignant classification and report writing while keeping every intermediate output as a correctable piece of evidence. Segmentation reaches a mean Dice of 87.21%; on 4,999 cases from 35 independent Chinese centres, classification AUROC falls to 0.819. The paper never states how benign versus malignant was established on its test sets, the mean AUROC quoted in the abstract (0.9466) does not match its own supplementary table (0.9001), and the word "overdiagnosis" does not appear once.
AMIE on video: a Google multi-agent system outperforms primary care physicians in a simulated OSCE — with the doctors' cameras off
Google Research and Google DeepMind put AMIE through an objective structured clinical examination over video: 100 scenarios, 15 professional actors, ten US primary care physicians as the comparator. Top-1 diagnosis is correct in 0.91 of cases versus 0.77 for physicians, and adding video lifts the observation-and-examination score from 0.41 to 0.74. But the human comparator had to keep their camera off, cases were restricted to what an actor can portray, and the actors themselves found a significant difference on only 2 of the 29 criteria they rated.
Report-derived labels versus image re-reading: for cardiomegaly in MIMIC-CXR, agreement is close to nil
Radiologists re-read 2,160 chest radiographs from MIMIC-CXR and compared their reading with labels extracted automatically from reports. For cardiomegaly, Cohen’s kappa is 0.011 and the label recovers only 1.3% of confirmed cases; 94.8% of disagreements stem from omission in the report, not negation. On the expert-relabelled cohort, a DenseNet121 reaches a ROC-AUC of 0.853. A useful, honest audit, but limited to one repository, one finding and an artificially balanced cohort.
Five retinal diseases, "experts" that specialise: what a sparse Mixture-of-Experts reveals on fundus photographs
Nagur Shareef Shaik and colleagues propose an architecture that does not apply the same computation to every fundus image: a "sparse" routing mechanism sends each photograph to a specialised "expert" depending on the pathology present. On five classes (healthy eye, diabetic retinopathy, AMD, epiretinal membrane, glaucoma) with patient-disjoint cross-validation, the model reaches a macro AUC of 0.912 and a macro F1 of 0.653, and its experts genuinely specialise by disease (p < 0.001). An elegant contribution to interpretability, but the AUC/F1 gap and the lack of external validation are a reminder that a neat mechanism is not yet a screening tool.
Training a patient-record model without ever pooling the data: federated learning tested on 122,251 ICU stays
Michael Burkhart, Brett Beaulieu-Jones, and colleagues test a concrete question: can you train a patient-record "foundation model" without ever gathering several hospitals' data in one place? Across 122,251 ICU stays from three independent health systems and twelve prediction tasks, their generative event models transfer far better between hospitals than conventional models (a 0.025-point AUC penalty versus 0.079), and federated learning nearly matches centralized training. But the most useful finding is counterintuitive: pooling the data yields only a modest gain over purely local training, and that advantage fades as each hospital accumulates its own data.
Predicting the day's anatomy before scanning: a "digital twin" for adaptive head-and-neck proton therapy
Yizhou Wu, Yuheng Li, Xiaofeng Yang, and Chih-Wei Chang propose a digital twin that predicts a head-and-neck patient's anatomy on treatment day, before acquiring any image, by transferring onto that patient's planning CT the way other patients changed during treatment. On 88 patients, these predicted CTs match the real day's anatomy better than the frozen planning CT: +22.8% image correlation, +20.2% organ-at-risk overlap, −23.4% CT-number error. Anticipating anatomy to speed up adaptive radiotherapy is a clever idea, but the demonstration is limited to image-similarity metrics and never evaluates delivered dose — the one quantity that truly decides benefit in proton therapy.
Representing each patient as a graph to skip data cleaning: PatTree and its 98.5% accuracy on Alzheimer's classification
Four researchers propose PatTree, a method that automatically structures heterogeneous medical data — missing values, multiple modalities, inconsistent formats — into a single per-patient knowledge graph, with no prior harmonization step. On a subset of the ADNI-1 cohort (763 subjects), classifying directly on this graph reaches 98.5% balanced accuracy and an F1 of 0.987 to distinguish Alzheimer's disease, mild cognitive impairment, and normal cognition. Skipping manual harmonization is a valuable idea; but a near-perfect score on a notoriously hard task, from a single cohort and without external validation, reads first as a methodological red flag to investigate, not as proof.
AI super-resolution in brain MRI: does reconstructing thin slices erase, or invent, small white-matter lesions?
A University of Pennsylvania team tests whether AI super-resolution, applied to thick-slice FLAIR MRI before automatically counting white-matter lesions, preserves those lesions. On 29 artificially degraded ADNI subjects, the main danger is not the hallucination feared from generative models but the silent erasure of real small lesions, worse as slices get thicker. Every reconstruction still beats the raw thick slice; the self-supervised model ECLARE preserves signal best, while an implicit-network foundation method does no better than plain cubic interpolation. Small sample, simulated degradation: a methodological warning, not a clinical verdict.
Intraventricular haemorrhage: a "brainstem dorsal line" on the CT scan, read by a tabular foundation model, to predict six-month disability
Nine Chinese hospitals, 728 patients. A team led by Southwest Hospital (Army Medical University) proposes the BSDL, a simple grade of brainstem compression measured on the admission CT, and feeds it with eight other variables to TabICLv2, a tabular foundation model, to predict poor six-month outcome (mRS 4-6). External AUC of 0.90 and good calibration — but external validation rests on just 118 patients, all Chinese, retrospective, and the BSDL is still graded by hand.
Readmissions and multimorbidity: a self-supervised model that learns the patient before it knows the outcome
The AI-MULTIPLY consortium (Newcastle, Queen Mary, Birmingham) builds Self-HR, a model that first learns to "summarize" the records of 99,207 UK patients with at least two chronic conditions, then predicts an emergency 30-day readmission. AUROC of 0.92, external validation on 79,224 patients from a second database — but most of the signal comes from the admission diagnosis, on an unrepresentative population and without prospective validation.
Rare cases in medicine: training an LLM to choose the next action, not just the final diagnosis
A Ben-Gurion team builds MedUPSQA, 21,874 mid-stream clinical decision points drawn from 5,535 real rare cases, and shows that aligning a language model on the next action helps more than making it bigger — provided you look at who defines the right answer.
ChatGPT Health for triage: more conversation does not make the advice safer
A Mount Sinai team tests ChatGPT Health on 255 real and synthetic triage cases. Talking longer does not improve guidance, and when the AI diverges from the nurse standard it under-estimates urgency nearly seven times out of ten — with a dangerous under-triage about one time in ten.
Predicting ICU delirium from one hospital to another: what transfers, and what does not
Shengliang Ni and Kengo Sato (Institute of Science Tokyo) train five model families to predict ICU delirium, then move them between databases — from eICU (208 US hospitals) to MIMIC-IV (a single Boston center) and back. Risk ranking transfers reasonably well (external AUROC up to 0.88–0.95 in one direction), but probability calibration and the alert threshold do not. The contribution is not a score: it is a distinction — a clinical model's transportability breaks into layers that do not travel together.
When a clinical LLM fills in the wrong boxes: measuring, then repairing schema compliance for ICD-10, CPT and FHIR
Jianru Shen (University of Montana) tests three open-source LLMs on 320 clinical scenarios across ten specialties, asking them to produce structured outputs that conform to the ICD-10, CPT and HL7 FHIR standards. Without a safeguard, format compliance plateaus between 85.9% and 91.6%. A validation-repair loop lifts it to 99.0% within two iterations, with about 1% of cases escalated to a human. But the metric captures structural, not clinical, correctness: a perfectly well-formed code can still be medically wrong, and that is the gap the paper does not close.
Whether to operate on a pancreatic cancer: what a multimodal CT + clinical-data model really learns to triage
A team from Basel, Aarau and London trains a deep-learning model that fuses 3D CT with 17 clinical variables to sort 159 pancreatic ductal adenocarcinoma patients into the three NCCN resectability categories. AUC reaches 0.86 in cross-validation and 0.86 on an independent external cohort of 52 patients from Aarau. The model learns vessel-aware anatomy without requiring a mask at inference. But the ground truth is only a tumor-board consensus, the cohort is small and Swiss, and no study compares the model to radiologists.
One week of wrist movement to predict future disease: what two accelerometry foundation models really read in the UK Biobank
A Stanford and Denmark team encodes one week of wrist movement from 97,696 UK Biobank participants with two frozen foundation models, then predicts 390 future diseases from a single survival model. Mean concordance is 0.688, one "vigour" axis explains 76% of predicted risk, and a prodromal Parkinson's disease signature emerges (five-year AUROC 0.90, 428 cases). But everything rests on a single non-representative cohort, with no external validation, and the encoders were pretrained on those same data.
Predicting the heaviest Mohs surgeries before the operation: what thirty machine-learning models mostly learn from tumour size
A Sydney team trains thirty machine-learning algorithms to predict, before surgery, whether Mohs micrographic surgery will need 13 or more tissue sections — a "heavy case" threshold copied from the Australian reimbursement schedule. On 408 procedures from a single centre, the best model reaches a test AUC of 0.884 and 81.7% accuracy. But the test set holds just 82 cases, no external validation is done, and the feature-importance analysis shows the model mostly replays an obvious fact: the bigger the tumour, the more you have to cut.
Screening for sleep apnea from a plain face photo: what a VLM-based multimodal system does, and why its "accuracy" flatters the numbers
A Shanghai team (Shanghai University of Engineering Science, Tencent Youtu Lab, Fudan ENT Hospital) proposes EviOSAHS, a system that gauges sleep-apnea risk by orchestrating, through prompting and with no training, a vision-language model and an LLM: it decomposes a face photo into seven structured anatomical observations, then decides by adding a clinical profile. On 642 patients with polysomnography, it reports 88.47% accuracy and 94.86% sensitivity. But the cohort is imbalanced (584 positive, 58 negative), specificity is never reported — it can be reconstructed at about 24% — and an "always positive" model would beat it on accuracy.
Lung cancer on low-dose CT: what if we trained AI on "degraded" full-dose scans instead of denoised ones?
A team at Politecnico di Milano tests a counterintuitive idea for lung cancer screening: instead of denoising low-dose CT (LDCT) images, degrade the abundant full-dose (SDCT) scans into fake low-dose images to train a nodule classifier. Three degradation methods are compared on the public LIDC-IDRI dataset. Robust finding: a model trained on full dose collapses on real low dose (sensitivity 0.571), while all three degradations restore an AUC above 0.84 — but ranking them against each other stays fragile.
Mixing mammography datasets does not improve screening AI: when the model learns the dataset's "signature" instead of the cancer
A team at Memorial University of Newfoundland tests a widespread intuition: topping up a real mammography screening set (NLBSD, 5,997 exams, 1.29% suspicious) with biopsy-confirmed cancers from public datasets (CBIS-DDSM, CMMD) should improve a triage model. The opposite happens: AUC falls from 0.737 to 0.620 as sources are added, and a diagnostic test shows why — the model separates images by dataset of origin at 0.9998 AUC, learning the dataset's "signature" rather than the cancer.
Are medical imaging foundation models interchangeable? What a controlled dissection of 25 encoders reveals about representational "convergence"
A German academic team tested an idea that has become received wisdom: as they grow, medical image encoders supposedly converge toward a shared representation, making them interchangeable. Dissecting 25 open-weight encoders (18 image, 7 text), from 7 million to 27 billion parameters, across five modalities and 650,982 chest radiographs, the authors show that convergence is real but modest, that it is driven by the self-supervised training objective rather than by scale or clinical supervision, and that it does not reproduce how a radiologist judges the similarity between two cases.
AI-generated pathology reports: why BLEU scores overstate quality, and what the clinical metric CRQS fixes
A team from Stony Brook and the University of Utah built the first standardized benchmark for automatically generating pathology reports from digitized slides, re-evaluating four models and three image encoders across three datasets. Their finding: the usual language metrics (BLEU, ROUGE, METEOR) reward surface similarity and can score highly a report that flipped the diagnosis. To fix this they propose CRQS, a score that compares extracted clinical facts rather than words — but the judge metric itself relies on a language model and is measured against noisy references.
Predicting mortality from an ECG, with no labels: what an autoencoder's "reconstruction error" really measures
A team at the Medical University of Innsbruck trained a masked autoencoder on 7.2 million Brazilian electrocardiograms, never showing it who would die. The idea: the more an ECG departs from the learned "normal", the larger its reconstruction error, and the higher the death risk. Validated on 1.6 million patients across three continents, the score predicts all-cause mortality well — but a careful look shows the word "self-supervised" hides a selection driven by labels, and that the headline figure never isolates what the ECG adds beyond age.
Predicting preterm birth from a cervical ultrasound: why these AI models collapse on a different scanner
A team from the GARBH-Ini cohort (North India), with the University of Oxford, trained several models — including a Vision Transformer — to predict spontaneous preterm birth from a single mid-trimester cervical ultrasound. Internally, the best model reached an AUC of 0.71; tested on an independent cohort acquired with a different scanner, it fell to 0.52, barely above chance. A negative result, honestly published, that mostly says what medical imaging AI cannot yet do: walk out of the lab.
Auditing the hidden biases of a medical AI without metadata: what CAPRA offers, guessing patient subgroups from the image itself
Six researchers at the Beijing University of Posts and Telecommunications propose CAPRA, a method that reconstructs demographic and technical axes — age, sex, device type, view — directly from the images, in order to spot the patient subgroups on which a model fails once the metadata needed for that audit have disappeared. Evaluated across three modalities — fundus (diabetic retinopathy), dermoscopy (HAM10000) and chest radiography (CheXpert) — it surfaces disparities that metadata-only slicing misses, and improves worst-subgroup performance in fourteen of fifteen comparisons when a robust learner reuses its partitions. The idea is useful and honestly presented, but guessing sensitive attributes from the image opens a risk of circularity, and nothing is validated in clinical conditions.
Molecular diagnosis of childhood glioma from MRI: what a contrastive-learning method that fixes its own "false negatives" delivers, and does not prove
Six researchers at the University of Toronto and the SickKids children's hospital propose MSeaCL, a method that learns to link brain MRI scans and radiology reports while accounting for how close in meaning they are, so as to stop treating two genuinely similar tumours as opposites. Pre-trained on 341 image–report pairs from a single hospital, then fine-tuned to classify the molecular subtype of paediatric low-grade gliomas, it gains only a hundredth of AUC internally but jumps 0.23 points on an external test set from different scanners. The generalisation result is real and instructive, but rests on tiny numbers, a single imaging sequence, and an evaluation that stays far from the clinic.
Robustness of pathology foundation models: what a "perturbations plus distribution shift" benchmark reveals about their brittleness
Three researchers at the University of Warwick put twelve digital-pathology foundation models, and two reference convolutional networks, through eleven realistic image perturbations and a validation protocol built to break the resemblance between training and test data. The large models resist better than the older networks, but their robustness plateaus well before the billion-parameter mark, and all lose accuracy the moment they face genuinely different distributions. A useful reminder: clinical reliability will rest on the quality and diversity of training data, not on model size.
Predicting Parkinson's from genetics: when cohort enrichment turns a risk gene into a "protective" one
A researcher at Claremont Graduate University trains, on 1,987 participants from the Fox Insight cohort, a battery of models — from a penalized logistic regression to an ensemble of algorithms — to predict a Parkinson's diagnosis from a handful of targeted genetic variants. The best model reaches an AUC of 0.929 on a held-out test set, a seemingly excellent figure. But the cohort holds 1,861 patients for only 126 controls, and the field's best-known risk variant comes out "protective": the performance mostly measures a recruitment artifact, not a screening ability.
MedQADE: an LLM grading medical answers matches physicians on agreement, never on caution
A team from the universities of Lübeck, Tübingen, Würzburg and Berlin's Charité builds MedQADE, the first open-response clinical benchmark in German: 3,800 questions drawn from the national medical exam, scored by physicians and by nine language models acting as graders. The best, Gemini 3 Flash, reaches agreement with the medical consensus (kappa 0.694) statistically indistinguishable from the human ceiling (0.709). But where physicians abstain when a question exceeds their expertise, frontier models issue a definitive verdict in 100% of cases: surface-level agreement without the clinical caution that normally comes with it.
ALICE: a single pathology foundation model distilled from eight experts — best average rank across 96 tasks, but no prospective validation
A team from Tsinghua Shenzhen, Oxford, and Sun Yat-sen University fuses eight digital-pathology foundation models into a single network through a three-stage "agglomerative distillation" (ALICE). Across 96 tasks over 48 data sources, ALICE achieves the best average rank against specialized models, beating the runner-up by 1.79, 6.39, and 3.04 points depending on task type. But the evaluation is entirely retrospective, the comparators are ALICE's own teacher models, and no data-leakage control is reported even though pretraining and testing both draw on TCGA.
MedPMC: curating the medical literature to train multimodal models — +7.1 AUC points across 26 benchmarks, but still thin clinical validation
A Yale-led team turns 6.1 million PubMed Central articles into 11 million medical image-text pairs via a five-stage automated pipeline (MedPMC). A vision-language model trained on the corpus, MedPMC-CLIP, gains 7.1 AUC points on average across 26 benchmarks spanning 11 specialties, against the best comparable model, with fewer than half the pairs. But the only evaluation on real patient data is an image-retrieval task in dermatology, and the authors warn that nothing is ready for clinical use.
RAG and public health: 99% right on multiple-choice, but barely two-thirds of answers faithful to the official guidance
A UK team extends PubHealthBench — 7,929 questions drawn from 687 UK Government public health guidance documents — to evaluate Retrieval-Augmented Generation. Grounding a language model in the right passages pushes multiple-choice accuracy to around 99%, with small open-weight models beating a large proprietary one; but in free-form answering, the best configuration is faithful to the guidance only 64% of the time, and the automated judge collapses precisely on "factual consistency" (κ near 0).
Turning a network's attention into a biomarker: Grad-CAM-guided "radiomic signatures" to classify tumors — marginal gains, clinical validation barely above chance
A University of Florida team proposes converting a segmentation network's Grad-CAM attention maps into "signature" regions, from which radiomic features are extracted to classify breast, brain and kidney tumors. On public datasets (BUSI, BraTS) the signature improves classification by 0.2 to 3 AUC points; but on the only private clinical cohort (UF Health renal CT), every deep network stays near chance (AUC ≈ 0.49–0.59). The preprint reports no sample sizes, no code, and no funding or conflict-of-interest statements.
Cardiac resynchronization: a digital twin to choose where to place the lead in the coronary sinus — a 74-patient pilot, nothing validated prospectively
A Russian team combines a cardiac digital twin, electrophysiological simulations and an explainable classifier (SVM) to predict, site by site within the coronary sinus, the response to cardiac resynchronization therapy. On 74 patients, the model beats a reference clinical calculator (AUC 0.78 in internal validation, accuracy 0.78 vs 0.58). But validation is purely internal and retrospective, the cohort is small, and the core result — "another position would have worked" — remains an in-silico prediction never checked in a patient.
ER-JEPA: learning the 12-lead ECG without labels — state of the art on a benchmark, nothing clinical yet
A hierarchical self-supervised model, ER-JEPA, is pretrained on about 174,000 unlabelled 12-lead ECG traces from China and Brazil, then evaluated on two public test sets. It matches the state of the art on the ST-MEM benchmark and surpasses it on PTB-XL fine-tuning (AUC 0.936 and 0.943), on a single consumer GPU — but the author admits training is unstable, the code is not yet released, and no figure is translated into clinical performance.
An electronic frailty index built by deep learning on clinical notes: across 193,629 Finns, a sevenfold death risk in the frailest — but a threshold calibrated on the mortality it predicts
A Finnish team builds an electronic frailty index by using deep learning to extract ten functional deficits from free clinical notes, on top of ICD-10 codes and lab results. Across 193,629 people aged 35 to 103, the death risk of the frailest is multiplied by 7.3 and severe-infection risk by 9.2 — but the frailty thresholds are calibrated on the very mortality they predict, the comparators are weak, and validation remains single-centre.
Genosolver: large language models that read clinical notes to diagnose rare diseases — causal gene ranked first in 72% of solved cases, but only 1.7% more diagnoses on unsolved ones
A team from Aachen combines large language models, reasoning models and RAG to extract phenotypes from unstructured clinical notes and re-rank genetic variants in rare disease. The causal gene is ranked first in 72% of already-solved cases and beats Exomiser, but reanalysing 1,875 unsolved cases adds only 1.7% more diagnoses — and the comparator receives less information than the model.
Predicting 10-year ischemic stroke risk: an XGBoost that beats classic scores but whose absolute risk collapses from one hospital to the next
A team from Birmingham and Mount Sinai combines electronic health records, laboratory trajectories and twenty polygenic risk scores in an XGBoost to predict 10-year ischemic stroke. The model easily beats classic scores and its ability to rank patients transfers to an external cohort — but absolute risk collapses from one hospital to another, the genomic contribution is marginal, and self-reported race adds almost nothing.
Segmenting the inner retina in retinitis pigmentosa: two AI models, including the SAM foundation model, trained on just 228 OCT slices
A Göttingen team adapts the SAM foundation model and an nnU-Net to measure the inner retinal layers on OCT in patients with retinitis pigmentosa, using very little annotated data. Reliable on the inner layers, but failing precisely where disease staging happens, and validated on a single cohort.
RadGrounder: a vision-language model for radiology that shows where it is looking, trained on 1.2 million slices without manual annotation
RadGrounder writes reports, answers questions and localises on the image the structure it mentions, trained on 1.2 million CT and MRI slices labelled entirely by other AIs. Strong on two public benchmarks, but the grounding only points to organs — never the lesions.
Severe acute pancreatitis: a random forest beats every deep-learning model on 722 patients
On a Chinese cohort of 722 patients, eleven models predict severe acute pancreatitis at admission: classical models win (Random Forest, AUC 0.877) and deep learning fails, but the cohort is 81% severe cases, the inverse of reality.
Stroke prognosis: six neurologists, a classical model and a deep-learning model compared on the MR CLEAN trial
On the MR CLEAN trial, models predict three-month disability after large-vessel stroke better than six neurologists, whose systematic optimism skews their prognosis.
Classifying an emergency ECG from its image: a ConvNeXt ensemble nears cardiologists on 18,519 tracings
At InCor (São Paulo), a model that reads the image of an ECG classifies 12 emergency tracing types with a macro F1 of 0.807, versus 0.820 for the annotating cardiologists — useful when only a paper or photographed tracing is available.
ICU mortality: predicting death from the first 24 hours on MIMIC-IV, and why calibration matters as much as discrimination
Five tabular models predict ICU mortality on 53,866 MIMIC-IV stays. AUROC of 0.856, but an AUPRC of 0.45 and a single center are a reminder of the limits.
Prostate cancer: segmenting lesions on PSMA PET/CT with a transformer, and stratifying survival before radioligand therapy
Fine-UNETR, a vision transformer, automatically segments PSMA lesions on whole-body PET/CT. Dice of 66.63% internally, but 44.11% on external validation.
StrokeTHG: predicting stroke mortality at 30, 90 and 365 days with a heterogeneous graph of patient records
A heterogeneous graph neural network predicts post-stroke mortality at three horizons from EHR data. AUROC 0.837-0.878, but single-centre, transductive evaluation without external validation.
Foundation models in multimodal oncology: what an audit of pathology and the transcriptome reveals
Five foundation models tested on pathology slides and the transcriptome of 7,600 patients: on omics a plain PCA beats the dedicated model, and fusing modalities does not always help.
AI cervical cancer screening: what a four-country validation actually reveals
A multi-task model triages treatable lesions (CIN2+) from a single colposcopy image. Strong in Germany and India, it drops to chance level in Romania.
Automatically interpreting qPCR Ct values: what a model trained on 41,770 amplification curves shows
An XGBoost model learns "normal" Ct behaviour from 41,770 qPCR curves to flag abnormal amplifications — but its reference is the machine, and it fails from one instrument to another.
When a skin-cancer detector changes country: a cascade classifier of dermoscopic images falls from 0.96 to 0.80 AUC between the ISIC archive and a Russian clinic (arXiv, 2026)
A decryption of a preprint posted on 11 June 2026 on arXiv by Elena Kozachok and colleagues: four deep-learning architectures (ViT-B/16, Swin-S, ConvNeXt-S, EfficientNetV2-S) and three classification schemes — binary, four-class, and a "triage then differentiation" cascade — are trained on the open ISIC archive, then tested on two small Russian clinical datasets. Internally the benign/malignant discrimination is excellent (ROC-AUC 0.952 to 0.966); on the Sechenov University data it drops to 0.797–0.893, sensitivity falls to 0.53–0.67, and calibration error climbs from 0.02 to 0.27–0.39, with the model underestimating malignancy. The cascade adds an explicit sensitivity control, absent from single-stage classifiers. The work is honest about the generalization gap — but the external cohorts are tiny and imbalanced, with no dermatologist comparator and no prospective validation.
Forecasting atrial fibrillation five minutes ahead: a personalized neural network reads wearable ECG (arXiv, 2026)
A decryption of a preprint posted on 9 June 2026 on arXiv by a team at Seoul National University: from a 60-second single-lead ECG segment, a neural network tries to forecast whether an atrial fibrillation episode will start within 5 minutes. Personalizing the model on each patient's first 24 hours lifts the AUROC from about 0.61 to 0.71 on the internal cohort, and from 0.59 to 0.69 on an external Korean cohort. The work is concrete, the code is open, and validation spans three cohorts — but discrimination stays modest, thresholds are tuned on the patient's own test data, the external European cohort is just 6 patients, and two authors are tied to the device maker.
SchistoTrackNet: a neural network reads liver ultrasound to spot the fibrosis of bilharzia (medRxiv, 2026)
A decryption of a preprint posted on 2 June 2026 on medRxiv: a neural network classifies liver-ultrasound images to detect periportal fibrosis caused by schistosomiasis, in rural Uganda. Trained on 3,710 images from the SchistoTrack cohort, SchistoTrackNet reaches 82.2% accuracy across six classes and agrees better with the sonographer who acquired the image (kappa 0.77) than a second sonographer does (0.54). The work takes a neglected disease seriously, with rare rigor — but the ground truth is a single human reader, the data come from one country and one machine, and the most severe fibrosis is caught only half the time.
Diagnosing acute myeloid leukemia from the bone-marrow smear: a "cell-to-patient" pipeline that sidesteps blast counting (arXiv, 2026)
A decryption of a preprint posted on 9 June 2026 on arXiv: a deep-learning pipeline detects and classifies the cells of a bone-marrow smear, then aggregates those observations into a per-patient score to assist the diagnosis of acute myeloid leukemia. Validated on 258 patients from six centers (89 of them held out for external validation), it reaches a weighted F1 of 0.87 to 0.91 on three unseen centers. The demonstration is careful and the cell-to-patient link well designed — but the learned target is a morphological proxy, not the leukemic blast, and the preprint releases no patient-level diagnostic metric, no code, and no comparison to a cytologist.
When breast density skews the evaluation of screening AI: the Mass-Bench benchmark and the hidden degradation (Mathematics, 2026)
A decryption of a study published 10 June 2026 in Mathematics: Mass-Bench unifies four public mammography datasets (32,930 images, 8,245 patients) to measure mass detection and BI-RADS classification not globally, but stratified by breast density. The finding — performance collapses as the breast grows dense, which imbalanced evaluations hide — is real and useful. But the paper reproduces some of the flaws it denounces: a headline figure absent from its own tables, test cells of a single image, and no released code for something billed as a benchmark.
Designing proteins that recognize only one shape of their target: AlloGen and learned conformational selectivity (arXiv, 2026)
A decryption of AlloGen, an arXiv preprint posted 3 June 2026: a framework that generates bespoke proteins binding a single conformation of their target — an enzyme's active form but not its resting form. At its core is a learned scorer, Q_θ, that rates the conformational selectivity of a protein-protein interface. Across eight targets never seen in training it reaches a mean rank correlation of 0.520, and on calmodulin de novo peptides bind the holo form with no detectable binding to the apo form. An elegant, reproducible proof that conformational selectivity is learnable — but a single wet-lab target, a surrogate metric, weak baselines and a non-commercial license place it far upstream of any drug.
Predicting HIV treatment non-adherence with machine learning: what is a "real-world" validation on 192,732 multi-country records worth? (medRxiv, 2026)
A decryption of a preprint posted in May 2026 on medRxiv: machine learning models validated on 192,732 multi-country clinical records to predict HIV treatment non-adherence and quantify gaps in the care pathway. Temporal validation reaches a 0.772 AUC and the study documents a median 74-day delay between diagnosis and treatment start. An honest, useful large-scale demonstration — but modest discrimination, an adherence outcome left opaque in the public abstract, and economic modelling with undetailed assumptions invite reading the numbers for what they are.
Topology meets vision transformers for brain tumor classification: what is 99.1% accuracy on a single MRI dataset worth? (Ahmed 2026, arXiv)
A decryption of the preprint posted on 30 May 2026 on arXiv by Faisal Ahmed (Embry-Riddle Aeronautical University, Arizona): a model fusing a vision transformer with topological data analysis (persistent homology) to sort brain MRIs into four classes. It reports 99.10% accuracy and 99.98% AUC on the public BRISC2025 benchmark — but the gain over existing models sits within the noise, evaluation rests on a single dataset with an image-level split that does not rule out data leakage, and the test set also serves to select the model.
BreastGPT: one multimodal model for the entire breast cancer care pathway — what a 90% score on a home-made benchmark is really worth (Liu et al. 2026, arXiv)
Critical analysis of the preprint posted on 3 June 2026 to arXiv by Yang Liu and colleagues (Alibaba DAMO Academy, Zhejiang University, Hupan Lab, West China Hospital, China Medical University): BreastGPT, an 8-billion-parameter multimodal large language model claimed to cover the entire breast cancer care pathway — screening, diagnosis, treatment planning — across five imaging modalities (mammography, ultrasound, MRI, CT, pathology slides) and text. Trained on 1.86 million question-answer pairs largely built by Alibaba's own large models, it reaches 75.66% accuracy on multiple-choice questions and 89.92% on open-ended questions of its own BreastStage-Bench. A genuine engineering feat, but most of the gap comes from training on the exact test distribution: the fair comparator gains only a few points, nothing was evaluated on real patients or compared against clinicians, and the corpus is largely generated by the in-house models.
MCEN: predicting complete response to breast cancer chemotherapy from a biopsy, with the Mamba architecture (Zhang et al. 2026, npj Digital Medicine)
Critical analysis of the article published on 2 June 2026 in npj Digital Medicine by Wenchuan Zhang, Shuwan Zhang, Fengling Li, Qingjie Lv, Yuhao Yi and Hong Bu (West China Hospital, Sichuan University, and colleagues): MCEN, a Mamba-based deep learning model that predicts, from a needle biopsy read as a digital slide, whether a breast cancer patient will achieve a pathological complete response after neoadjuvant chemotherapy. Trained on 1,023 patients from one Chinese hospital then tested on four independent centers (1,646 patients in total), it reaches an AUROC of 0.923 in training but falls to 0.76–0.81 on external validation, with fusion of clinicopathological data rising to 0.84. Strong for its genuine multicenter validation and Mamba's efficiency on gigapixel images, the work remains limited by a marked train–validation gap, an exclusively Chinese cohort, exclusions that drop atypical forms, and no comparison against pathologists.
SKELEX: a foundation model trained on 1.3 million radiographs to read bone, from cyst to fracture (Kim et al. 2026, npj Digital Medicine)
Critical analysis of the article published on 2 June 2026 in npj Digital Medicine by Shinn Kim, Soobin Lee, Ilkyu Han, Sunghoon Kwon and colleagues at Seoul National University: SKELEX, presented as the first large-scale foundation model dedicated to musculoskeletal radiographs. A masked autoencoder with a ViT-Large backbone is self-supervised pre-trained on 1,296,540 unlabeled radiographs from a single Korean hospital (2010-2016), then adapted to 12 diagnostic tasks across 7 public datasets. It beats five baselines by 6.21% on average (relative), reaching an AUROC of 0.953 vs 0.884 for its own initialization model on bone tumor detection, is better calibrated, and matches the best models with half the labels. Convincing on label efficiency and methodological hygiene, the work is limited by single-center, single-country training data, genuine external validation restricted to the bone-tumor application alone, no comparison against radiologists, a resolution reduced to 224×224, and weights released for academic use only.
PINNOCHIO: predicting the post-operative face in orthognathic surgery with a physics-informed network, as accurate as finite elements but in seconds (Lee et al. 2026, arXiv)
Critical analysis of the preprint posted on arXiv on 1 June 2026 (submitted to MICCAI 2026) by Jungwook Lee, Daeseung Kim, Kevin Gu, Zhangfeng Hu, Tianshu Kuang, Finn Hopeman, Michael A.K. Liebschner, Jaime Gateno and Pingkun Yan (Rensselaer Polytechnic Institute, Houston Methodist, Baylor College of Medicine): PINNOCHIO, a physics-informed neural network that predicts how facial soft tissue deforms after the jaws are surgically repositioned, by separating the bone–tissue interface movement from the volumetric hyperelastic deformation. On 40 real clinical cases (pre-operative CT + post-operative 3dMD surface) it matches or beats the reference finite-element simulator on surface fidelity (Chamfer distance 1.12 mm vs 1.30; 86.55% of points within 2 mm vs 80.90%) while running in 3.24 seconds instead of 3.5 hours. Convincing on speed and biomechanical plausibility, the work is limited by a 40-patient cohort, supervision that only covers the outer surface, fixed mechanical parameters shared by all patients, and no released code or weights.
When an LLM must run the interview itself: an exam-inspired benchmark shows interactive diagnostic reasoning degrades performance (Zhan & Gan 2026, arXiv)
Critical analysis of the preprint posted on arXiv on 21 May 2026 by Chen Zhan, Xihe Qiu, Xiaoyu Tan, Xibing Zhuang, Gengchen Ma, Yue Zhang, Shuo Li, Peifeng Liu, Xiaoxiao Ge, Liang Liu and Lu Gan: an "OSCE-inspired" benchmark in which a standardized patient simulator forces fifteen large language models (LLMs) to run the interview themselves, turn by turn, before reaching a diagnosis. Across 468 cases, moving from information served upfront to active history-taking lowers diagnostic accuracy by 12.75% and supporting-evidence quality by 24.36%, with errors driven mainly by premature diagnostic closure and inefficient questioning. The sober, useful takeaway: rankings on static medical multiple-choice exams likely overstate what these models can do in a real consultation. Caveats: the patient simulator is itself algorithmic, the provenance of the cases is not detailed in the accessible abstract (contamination risk), and the figures are reported as relative values without an explicit human comparator.
GTBIS: a deep learning model that reads the morphology of combined pulmonary neuroendocrine carcinomas to predict prognosis (Yang & Zhou 2026, npj Digital Medicine)
Critical analysis of the npj Digital Medicine paper of 30 May 2026 by Lin Yang, Ruyu Sheng, Zijian Yang, Shilong Liu and Meng Zhou (National Cancer Center / Cancer Hospital of the Chinese Academy of Medical Sciences in Beijing, Wenzhou Medical University and Harbin Medical University Cancer Hospital): GTBIS, an interpretable deep learning model that reads pathology-slide morphology to distinguish small cell lung carcinoma (SCLC) from large cell neuroendocrine carcinoma (LCNEC), then applies that reading to combined cSCLC-LCNEC tumors to stratify their prognosis. Across multicenter cohorts totaling 670 patients, the model splits chemoradiotherapy-treated combined tumors into a favorable-prognosis SCLC-like subgroup (five-year overall survival 100% vs 39.5%, disease-free survival 87.5% vs 36.0%) and a poor-prognosis LCNEC-like subgroup, the classification remaining an independent prognostic factor in multivariable analysis. But the sample is modest, all centers are Chinese, validation is retrospective without an explicit human comparator, and the CC BY-NC-ND license closes adaptation.
Pathog-PDx: a machine learning system to identify 22 pediatric respiratory pathogens from the electronic health record (Su 2026, npj Digital Medicine)
Critical analysis of the npj Digital Medicine paper of 29 May 2026 by Dubin Su, Qun Chen, Ruizhi Xu and colleagues (First Affiliated Hospital of Xiamen University, Zhengzhou University, Nanjing University, Shenzhen Second People's Hospital and UIUC): Pathog-PDx, a diagnostic system that combines 42 clinical and laboratory features from the electronic health record to distinguish 22 pathogen subtypes responsible for respiratory infections in hospitalized children. Development cohort of 134,500 children across three centers and two databases, prospective independent validation on 1,338 children, mean AUC 0.88 across the 22 pathogens and 0.95 for influenza virus, public deployment of a web-based decision support tool. But all development centers are Chinese, the human clinical comparator is absent, the CC BY-NC-ND license blocks academic adaptation, and the very nature of the gold standard for 22 classes deserves a separate discussion.
EpiVLM: a vision-language model for video seizure detection and classification, from hospital to home (He 2026, npj Digital Medicine)
Critical analysis of the npj Digital Medicine paper of 26 May 2026 by Mengqiao He, Leihao Sha, Pengfei Wei, Lei Chen and colleagues (West China Hospital, Sichuan University and Shenzhen Institutes of Advanced Technology, CAS): EpiVLM, a vision-language model (VLM) that combines clinically structured prompts with video reasoning to recognize five seizure semiologies on 232 video recordings from 127 patients (11,666 annotated segments) drawn from two tertiary centers, unconstrained home recordings and an independent public dataset. Accuracy 0.795–0.947, sensitivity 0.842–0.957, video-level false detections 0.47–2.45%, mean onset-to-detection delay under 6 seconds, with prompts and thresholds fixed a priori without site-specific recalibration. But all tertiary centers are Chinese, the home cohort is barely described in the abstract, there is no head-to-head comparison with human annotators, and one co-author is affiliated with a private LLC (Brain Everest) without a competing-interest declaration.
An automated neuroimaging pipeline for personalized post-stroke cognitive prognosis (Brzus 2026, npj Digital Medicine)
Critical analysis of the npj Digital Medicine paper of 27 May 2026 by Michal Brzus, Joseph Griffis, Aaron D. Boes and colleagues (University of Iowa): a fully automated DICOM-to-PDF pipeline that segments ischemic lesions with a 3D Residual U-Net, predicts 28 neuropsychological outcomes via lesion network mapping, and drafts a personalized report via air-gapped LLaMA 3.3 70B in under three minutes. Training on 604 patients from the Iowa Lesion Registry, independent testing on 153 ischemic stroke patients imaged on 17 scanner models. AUCs of 0.74 to 0.90 on five detailed cognitive domains, 96% concordance between predictions from automatic versus manual segmentations. But training and testing from the same center, no clinical comparator (NIHSS, mRS, demographics alone), clinical review of reports by the senior author himself, and four of the seven authors hold the associated patent and co-founded NeuroPred Inc.
SHAP and SVM to predict deep venous thrombosis after endometrial cancer surgery (Zhou 2026, npj Digital Medicine)
Critical analysis of the npj Digital Medicine article of 27 May 2026 by Qing Zhou and colleagues: a four-variable SVM model (postoperative D-dimer, age, fibrinogen, FIGO stage) predicts deep venous thrombosis after endometrial cancer surgery, with AUC 0.828 in internal validation and 0.819 in an external cohort across 841 + 95 Chinese patients. SHAP makes contributions interpretable. But symptom-triggered imaging (detection bias), 100% Chinese cohort, no head-to-head comparison with Caprini/Wells scores, and D-dimer measured after surgery — this is more an early-detection aid than a strict prediction.
UNet-MoE-Cli: a mixture-of-experts to personalize neoadjuvant therapy in rectal cancer (Liu 2026, npj Digital Medicine)
Critical analysis of the npj Digital Medicine article of 26 May 2026 by Xiangyu Liu and colleagues: UNet-MoE-Cli, a mixture-of-experts deep learning model on multiparametric MRI and clinical variables, estimates regimen-specific pathological complete response probabilities for neoadjuvant therapy in locally advanced rectal cancer. AUC 0.827 in internal validation, 0.790 in prospective cohort (ChiCTR2400085797), but sensitivity only 0.45–0.53, single-centre nCT expert, 100% Chinese cohort, and the escalation benefit is computed by the model itself.
When text eats the image: what the Restrepo 2026 study reveals about the contextual fragility of clinical VLMs on MIMIC-CXR
Critical analysis of the arXiv preprint 2605.17436 of 17 May 2026 by David Restrepo (CentraleSupélec-Université Paris-Saclay) and colleagues: eight vision-language models evaluated on 1,000 MIMIC-CXR chest X-rays lose up to 66% of their correct decisions when the clinical text is swapped for that of an opposite-class patient. Image-only tops out at 0.50–0.68, text-only matches multimodal. Even MedGemma, adapted to medical data, collapses. These VLMs are report classifiers disguised as image readers.
PromptRad: labelling liver CT reports with only 32 annotated examples, and matching GPT-4
Critical analysis of the May 2026 arXiv preprint 2605.20052 (BioNLP 2026 @ ACL) by Ying-Jia Lin and colleagues (Chang Gung University, Taiwan): a 110-million-parameter PubMedBERT, fine-tuned via prompt-tuning with a UMLS-enriched verbalizer, achieves 89.2% macro F1 on seven categories of liver lesions in CT — from only 32 annotated reports, and with better negation handling than GPT-4.
10,000 synthetic cases against four frontier LLMs: what Auger 2026 reveals about the clinical blind spots of Gemini 3 and GPT-5 in multiple sclerosis
Critical analysis of Stephen D. Auger's April 2026 medRxiv preprint (Imperial College London): up to 10,000 synthetic multiple sclerosis cases with ground truth, four frontier models (Gemini 3 Pro/Flash, GPT-5.2/5-mini) evaluated on diagnosis, localization, investigations and management. Diagnostic accuracy does not predict therapeutic safety: Gemini under-uses appropriate corticosteroids, GPT-5 recommends intravenous thrombolysis in nearly one out of ten cases.
GPT-4 in radiology: why the format of an LLM's explanation changes physicians' diagnostic accuracy
Decryption of Spitzer et al.'s 2026 npj Digital Medicine paper: a randomized trial with 101 radiologists comparing three formats of GPT-4 explanation. Chain-of-thought adds 12.2 percentage points of accuracy, while differential diagnosis induces automation bias. Implications for the clinical deployment of LLMs.
GigaPath in digital pathology: what changes when a foundation model is trained on 1.3 billion tiles
Critical analysis of the Nature 2024 paper on Prov-GigaPath, a transformer foundation model for digital pathology. Architecture, data, performance on 26 cancer benchmarks, and what it really changes for diagnosis.