Diagnostics & Imaging
279 gains92 Traces84 problems
Every measured gain and problem on the record for this topic, newest first. Updated .
AI in diagnostics and imaging
One square per record, darker for stronger evidence:Evidence: 7 high401 moderate47 emerging
Most records report AI performing well on scans, slides, signals and patient data, usually in retrospective or internal tests. TRV-2026-0486TRV-2026-1490 Few records show that patients did better, and reviews repeatedly note missing outside validation and missing outcome data. TRV-2026-1924TRV-2026-1490TRV-2026-0486
A randomized trial of 300 people with suspected inherited retinal disease shows specialists using an AI tool named the right gene category in their top four answers 81.8% of the time versus 53.1% alone. TRV-2026-0575 In a randomized crossover test, 21 physician assistants reading 50 ECGs reached 92.9% accuracy on heart attack diagnosis with AI versus 79.0% without it, though they took longer. TRV-2026-0651 A systematic review shows AI detected STEMI heart attacks on ECGs with higher sensitivity than clinicians (0.91 versus 0.78), and another review shows AI help raised doctors' fracture detection to 87% from 73%. TRV-2026-2312TRV-2026-1145 A meta-analysis of urological cancer imaging found AI reached a pooled AUC of 0.91 against 0.83 for clinicians, mostly in retrospective studies. TRV-2026-0640
In a randomized trial of 171 veterans after a heart failure hospitalization, AI-guided monitoring did not significantly cut rehospitalization (16.3% versus 18.7% for controls). TRV-2026-1587 A review of 48 studies found generative AI was 13.0 percentage points less accurate than physicians at radiology diagnosis. TRV-2026-1284 Four commercial chest X-ray tools did not improve five readers' accuracy across 1200 patients, and eight dental AI systems caught only 32.7% to 48.7% of decay next to fillings. TRV-2026-0854TRV-2026-1673 A study of 826 chest X-rays in Nepal found only 17.89% of the AI's abnormal flags were correct at a threshold set for high sensitivity. TRV-2026-1265
Many models lose accuracy outside the hospital that built them: a review of knee arthritis AI found median AUC fell to 0.75 from 0.81 on outside data, and Alzheimer's brain-scan AI performed worse outside single centers. TRV-2026-0638TRV-2026-1244 Reviews of heart, lung and eye AI say most studies are retrospective and single-center, with little data on workflow, patient outcomes or cost. TRV-2026-1434TRV-2026-1017TRV-2026-1490 Small pilots add little certainty: a 13-patient simulation suggested patient feedback cut false seizure alarms, while in about 60 children none of seven models using brain-scan signals beat chance at spotting autism. TRV-2026-1616TRV-2026-2016 Performance across patient groups is rarely tested; cuffless blood pressure estimates were less accurate for older and Black ICU patients, and a Brazilian sick-leave model was less accurate for Black workers. TRV-2026-1497TRV-2026-1528
At one hospital, an AI model diagnosed chronic atrophic gastritis better than 15 novice and 16 expert endoscopists, and in 33 hard-to-read ECGs an AI app flagged all 13 occlusion heart attacks. TRV-2026-1937TRV-2026-1923 Other records point the other way: on 1000 prostate MRI cases AI did not confirm non-inferiority against routine practice, and dentists predicted tooth survival more accurately than an AI model (75.6% and 74.9% versus 69.2%). TRV-2026-0402TRV-2026-0856 A pooled analysis found the edge depends on the task, with AI less sensitive than clinicians in the prostate cancer and MRI subgroups. TRV-2026-0640
Tools that read scans, X-rays and slides often score well in internal tests and lose accuracy at outside sites, and few have been tried prospectively on patients. TRV-2026-0257TRV-2026-1244TRV-2026-0638
See these 157 recordsModels built on slides, scans and gene scores sort cancer patients by risk, mostly in retrospective data from one center. TRV-2026-1589TRV-2026-1865TRV-2026-2138TRV-2026-0806
See these 54 recordsA randomized trial of 14,665 adults found AI-guided ECG monitoring picked up slightly more new atrial fibrillation, though the authors call it statistically fragile, and a review found AI-read ECGs detect weak pumping in people without symptoms. TRV-2026-2462TRV-2026-1837
A Trace is one record with both a gain and a problem, so it counts on both sides: 371 records show a gain and 176 show a problem. 172 context records are in the record list but not counted, because they report no measured gain or problem.
| Evidence | Gain side | Problem side |
|---|---|---|
| HighStrong study design, backed by at least two independent sources | 5 | 4 |
| ModerateA measured study or an official source | 342 | 136 |
| EmergingEarly, small, unreviewed or reported second-hand | 24 | 36 |
By type of evidence: 157 observational studies; 129 controlled comparisons; 73 benchmark or lab tests; 45 systematic reviews or meta-analyses; 35 commentaries or narrative reviews; 4 randomized trials; 4 news reports; 3 regulatory or legal actions; 2 sets of official statistics; 1 interview study; 1 survey; 1 case report. How evidence is scored
In a review of 150 clinically validated studies, AI reached expert-level accuracy in medical imaging, with AUC up to 0.94 for cancer detection.
In a model development and validation study of cancer patients, the highest-risk quarter flagged by a face-photo AI had more than triple the mortality of the lowest (adjusted hazard ratio 3.23); its age tool erred by 5.1 years on average.
In 12 externally validated studies of people without symptoms, AI-read ECGs detected weak heart pumping with pooled AUROC of 0.87 to 0.93, similar to results in people with symptoms.
In a study of kidney patients in the US and China, models built to classify heart disease performed worse when tested on the other country's patients, and key risk factors differed: age in the US, hypertension in China.
A single-center study found the software's sensitivity fell to 73% for distal M1 blockages, with only moderate agreement with experts; non-target occlusions were excluded from the main analysis, so it needs expert review rather than standalone use.
17 reviews in this topic printed a pooled result: one figure calculated across many studies. They are the reviews’ own numbers, checked word for word against their abstracts. They often measure different things, so they are listed side by side, not added up.
In a 110-study review, AI beat clinicians on urological cancer scans but lagged on prostate · TRV-2026-0640
Specificity 0.83 (95% CI 0.80 to 0.86)
110 studies · search to June 2026
AI models achieved pooled sensitivity and specificity of 0.85 (95% CI: 0.83-0.87) and 0.83 (95% CI: 0.80-0.86), with an AUC of 0.91 (95% CI: 0.88-0.93)
Sensitivity 0.82 (95% CI 0.79 to 0.85)
110 studies · search to June 2026
Clinicians demonstrated a pooled sensitivity of 0.82 (95% CI: 0.79-0.85) and specificity of 0.68 (95% CI: 0.62-0.73), with an AUC of 0.83 (95% CI: 0.80-0.86).
AUC 0.91 (95% CI 0.88 to 0.93)
110 studies · search to June 2026
AI models achieved pooled sensitivity and specificity of 0.85 (95% CI: 0.83-0.87) and 0.83 (95% CI: 0.80-0.86), with an AUC of 0.91 (95% CI: 0.88-0.93)
General chatbots trail clinicians on image-based diagnosis and often miss detail, though some do better on text tasks. TRV-2026-1284TRV-2026-2253TRV-2026-2004TRV-2026-1679
See these 33 recordsMachine learning on urine, breath, blood and brain-wave signals often reports high accuracy, but in small or internal samples. TRV-2026-1256TRV-2026-0616TRV-2026-0896TRV-2026-0801
See these 45 recordsHospital prediction models catch many events but raise false alarms or lose accuracy on outside data. TRV-2026-2093TRV-2026-1733TRV-2026-1655
See these 25 recordsEach record sits in one thread at most. A record with both a gain and a problem counts on both sides. 108 records fit none of these threads.
Blue Ortho recalled Predict+, a machine-learning tool that predicts outcomes after shoulder replacement surgery, on 2026-02-05 because it lacked pre-market clearance/approval.
AI help raised fracture detection to 87% from 73%; junior doctors gained most, a review shows · TRV-2026-1145
Sensitivity 73% (95% CI 69% to 78%)
17 studies · search to September 17, 2025
Compared to unassisted diagnosis, AI significantly improved pooled sensitivity (87%, 95% confidence interval [CI]: 84-89%, versus 73%, 95% CI: 69-78%) and maintained high specificity (95%, 95% CI: 92-97%, versus 94%, 95% CI: 89-96%).
Specificity 95% (95% CI 92% to 97%)
17 studies · search to September 17, 2025
Compared to unassisted diagnosis, AI significantly improved pooled sensitivity (87%, 95% confidence interval [CI]: 84-89%, versus 73%, 95% CI: 69-78%) and maintained high specificity (95%, 95% CI: 92-97%, versus 94%, 95% CI: 89-96%).
Specificity 94% (95% CI 89% to 96%)
17 studies · search to September 17, 2025
Compared to unassisted diagnosis, AI significantly improved pooled sensitivity (87%, 95% confidence interval [CI]: 84-89%, versus 73%, 95% CI: 69-78%) and maintained high specificity (95%, 95% CI: 92-97%, versus 94%, 95% CI: 89-96%).
Machine learning's edge over standard statistics in predicting trauma deaths is small and may be inflated · TRV-2026-1068
AUC difference 0.026 (95% CI 0.009 to 0.043)
17 studies · 243,324 people · search to January 15, 2025
The pooled AUC difference favoring ML was 0.026 (95% CI 0.009-0.043)
AUC difference 0.017 (95% CI 0.005 to 0.029)
search to January 15, 2025
the co-primary estimate was 0.017 (95% CI 0.005-0.029)
AI-made MRI scans match real ones overall but lag on diseased areas, a 41-study review found · TRV-2026-0904
SSIM 0.92 (95% CI 0.90 to 0.93)
15 studies · search to January 16, 2025
Meta-analysis of 15 brain tumor studies (30 models) yielded pooled SSIM of 0.92 (95% CI, 0.90-0.93) and pooled PSNR of 30.6 dB (95% CI, 28.6-32.6).
PSNR 30.6 (95% CI 28.6 to 32.6)
15 studies · search to January 16, 2025
Meta-analysis of 15 brain tumor studies (30 models) yielded pooled SSIM of 0.92 (95% CI, 0.90-0.93) and pooled PSNR of 30.6 dB (95% CI, 28.6-32.6).
Only 1 of 20 studies of AI fertility predictions had low risk of bias, a review shows · TRV-2026-0674
Sensitivity 0.737 (95% CI 0.662 to 0.799)
14 studies · search to July 2026
The pooled sensitivity was 0.737 (95% CI, 0.662-0.799) and the pooled specificity was 0.789 (95% CI, 0.709-0.851), with substantial heterogeneity (I2 = 97.7% and 99.0%, respectively).
Specificity 0.789 (95% CI 0.709 to 0.851)
14 studies · search to July 2026
The pooled sensitivity was 0.737 (95% CI, 0.662-0.799) and the pooled specificity was 0.789 (95% CI, 0.709-0.851), with substantial heterogeneity (I2 = 97.7% and 99.0%, respectively).
Diagnostic odds ratio 10.49 (95% CI 6.28 to 17.53)
14 studies · search to July 2026
The pooled DOR was 10.49 (95% CI, 6.28-17.53), and the SROC curve indicated acceptable overall discrimination.
For narrowed bile ducts, a review of five studies found AI hit 95% sensitivity · TRV-2026-0679
Accuracy (SROC) 97% (95% CI 95% to 98%)
The pooled sensitivity and specificity were 95% (95% CI: 85-98) and 88% (95% CI: 76-94), with a diagnostic accuracy (SROC) of 97% (95% CI: 95-98).
Accuracy (SROC) 97% (95% CI 95% to 98%)
4 studies · 538 people
Sensitivity analysis of CNN studies (4 studies, 538 patients) demonstrated a pooled sensitivity, specificity, and accuracy (SROC) of 95% (95% CI: 82-99), 88% (95% CI: 72-95), and 97% (95% CI: 95-98), respectively.
Specificity 88% (95% CI 76% to 94%)
The pooled sensitivity and specificity were 95% (95% CI: 85-98) and 88% (95% CI: 76-94), with a diagnostic accuracy (SROC) of 97% (95% CI: 95-98).
AI on heart ultrasounds spots a hard-to-diagnose heart failure, but evidence is thin · TRV-2026-2311
Area under the summary ROC curve 0.875 (95% CI 0.791 to 0.930)
2,770 people · search to June 2026
Pooled sensitivity was 84.0% (95% CI: 73.0-91.0%) and specificity 83.0% (95% CI: 73.0-89.0%), with an area under the summary ROC curve of 0.875 (95% CI: 0.791-0.930).
Diagnostic odds ratio 24.73 (95% CI 18.46 to 33.12)
2,770 people · search to June 2026
The diagnostic odds ratio was 24.73 (95% CI: 18.46-33.12)
Sensitivity 84.0% (95% CI 73.0% to 91.0%)
2,770 people · search to June 2026
Pooled sensitivity was 84.0% (95% CI: 73.0-91.0%) and specificity 83.0% (95% CI: 73.0-89.0%)
AI hit 91% sensitivity on children's ear infections; a nine-study review shows it matched GPs · TRV-2026-2182
Sensitivity 91% (95% CI 77% to 97%)
search to July 30, 2025
The pooled sensitivity and specificity were 91% (95% confidence interval [CI]: 77%–97%) and 99% (95% CI: 94%–100%), respectively.
Specificity 99% (95% CI 94% to 100%)
search to July 30, 2025
The pooled sensitivity and specificity were 91% (95% confidence interval [CI]: 77%–97%) and 99% (95% CI: 94%–100%), respectively.
AI scan tools for glioblastoma show useful but imperfect accuracy in outside tests, with AUC 0.81 · TRV-2026-2075
Sensitivity 0.78 (95% CI 0.63 to 0.88)
219 people · search to October 1, 2025
Pooled sensitivity was 0.78 (95% confidence interval [CI], 0.63-0.88), specificity was 0.75 (95% CI, 0.63-0.84), and summary area under the receiver operating characteristic curve (AUC) was 0.81 (bootstrap 95% CI, 0.69-0.91).
AUC 0.81 (95% CI 0.69 to 0.91)
219 people · search to October 1, 2025
Pooled sensitivity was 0.78 (95% confidence interval [CI], 0.63-0.88), specificity was 0.75 (95% CI, 0.63-0.84), and summary area under the receiver operating characteristic curve (AUC) was 0.81 (bootstrap 95% CI, 0.69-0.91).
Specificity 0.75 (95% CI 0.63 to 0.84)
219 people · search to October 1, 2025
Pooled sensitivity was 0.78 (95% confidence interval [CI], 0.63-0.88), specificity was 0.75 (95% CI, 0.63-0.84), and summary area under the receiver operating characteristic curve (AUC) was 0.81 (bootstrap 95% CI, 0.69-0.91).
In 12 studies, AI-read ECGs spot weak heart pumping in people without symptoms, a review finds · TRV-2026-1837
AUC 0.89 (95% CI 0.86 to 0.92)
search to 8th September 2025
Symptomatic cohorts showed AUROCs of 0.89 (95% CI 0.86-0.92), 0.90 (95% CI 0.88-0.92), and 0.85 (95% CI 0.83-0.86) respectively.
AUC 0.89 (95% CI 0.85 to 0.97)
search to 8th September 2025
In asymptomatic cohorts evaluating left ventricular systolic dysfunction (LVSD), AI-ECG demonstrated promising but heterogeneous discriminatory performance (I²>85%) with pooled AUROCs of 0.87 (95% CI 0.75-0.99) for LVEF ≤35%, 0.93 (95% CI 0.88-0.97) for LVEF ≤40%, and 0.89 (95% CI 0.85-0.97) for LVEF ≤50%.
AUC 0.93 (95% CI 0.88 to 0.97)
search to 8th September 2025
In asymptomatic cohorts evaluating left ventricular systolic dysfunction (LVSD), AI-ECG demonstrated promising but heterogeneous discriminatory performance (I²>85%) with pooled AUROCs of 0.87 (95% CI 0.75-0.99) for LVEF ≤35%, 0.93 (95% CI 0.88-0.97) for LVEF ≤40%, and 0.89 (95% CI 0.85-0.97) for LVEF ≤50%.
A review pooling nine studies shows AI detects eye toxoplasmosis in photos with 0.965 sensitivity · TRV-2026-1770
Specificity 0.955 (95% CI 0.929 to 0.972)
search to January 2026
Pooled sensitivity was 0.965 (95% CI: 0.944-0.978) and pooled specificity was 0.955 (95% CI: 0.929-0.972).
Sensitivity 0.965 (95% CI 0.944 to 0.978)
search to January 2026
Pooled sensitivity was 0.965 (95% CI: 0.944-0.978) and pooled specificity was 0.955 (95% CI: 0.929-0.972).
Diagnostic odds ratio 594.1 (95% CI 263.3 to 1340.4)
search to January 2026
The diagnostic odds ratio reached 594.1 (95% CI: 263.3-1340.4), with positive and negative likelihood ratios of 21.6 and 0.036.
Surgical AI built on a US registry scored 0.71 vs 0.80 for other data in a 295-model review · TRV-2026-1489
Adjusted regression coefficient (β) for AUROC -0.056 (95% CI -0.105 to -0.007)
NSQIP remained associated with lower AUROC (β=-0.056, 95% CI -0.105 to -0.007; p=0.026)
Generative AI was 13.0 points less accurate than physicians at radiology diagnosis, a review of 48 studies found · TRV-2026-1284
Accuracy 42.9% (95% CI 35.8% to 50.1%)
search to March 2025
Pooled diagnostic accuracy of generative AI was 42.9% (95% CI, 35.8-50.1%) for free-text tasks and 58.1% (95% CI, 50.0-66.2%) for choice tasks.
Difference in accuracy (percentage points) -25.9 (95% CI -39.1 to -12.8)
search to March 2025
Text-only input was associated with higher accuracy than image-only input (difference in accuracy [image-only minus text-only], -25.9 percentage points [95% CI, -39.1 to -12.8]; P = .001)
Accuracy 58.1% (95% CI 50.0% to 66.2%)
search to March 2025
Pooled diagnostic accuracy of generative AI was 42.9% (95% CI, 35.8-50.1%) for free-text tasks and 58.1% (95% CI, 50.0-66.2%) for choice tasks.
Review shows AI spots hip bone death at 88.8% sensitivity; only five tests used outside data · TRV-2026-1280
Sensitivity 0.888 (95% CI 0.789 to 0.944)
Pooled sensitivity was 0.888 (95% CI, 0.789-0.944), specificity was 0.937 (95% CI, 0.906-0.958), and SROC AUC was 0.967.
Specificity 0.937 (95% CI 0.906 to 0.958)
Pooled sensitivity was 0.888 (95% CI, 0.789-0.944), specificity was 0.937 (95% CI, 0.906-0.958), and SROC AUC was 0.967.
AI predicts death from heart valve infection well, but half the studies had bias risk · TRV-2026-1049
AUC 0.85 (95% CI 0.81 to 0.89)
search to April 2026
Pooled AUC was 0.85 (95% confidence interval [CI], 0.81-0.89) for in-hospital/early mortality (I2 = 35.3%) and 0.85 (95% CI, 0.82-0.88) for 6-month mortality (I2 = 0%)
AUC 0.85 (95% CI 0.82 to 0.88)
search to April 2026
Pooled AUC was 0.85 (95% confidence interval [CI], 0.81-0.89) for in-hospital/early mortality (I2 = 35.3%) and 0.85 (95% CI, 0.82-0.88) for 6-month mortality (I2 = 0%)
AI models flag TB treatment failure, but only 23.5% of 34 studies tested them externally · TRV-2026-0817
AUC 0.836 (95% CI 0.799 to 0.868)
100,790 people · search to October 2025
The pooled AUC was 0.836 (95% CI 0.799-0.868), with substantial heterogeneity (I² = 97.9%).
AI predicts brain bleed outcomes well in 83 studies, but mostly in internal tests · TRV-2026-0611
C-index 0.850 (95% CI 0.830 to 0.869)
search to September 2025
models integrating both clinical and radiomics features achieved the highest discriminative performance across key prognostic prediction tasks: predicting HE (pooled C-index 0.822, 95% confidence interval [CI] 0.789-0.855), poor functional outcome (C-index 0.850, 95% CI 0.830-0.869), and mortality (C-index 0.860, 95% CI 0.809-0.911)
C-index 0.860 (95% CI 0.809 to 0.911)
search to September 2025
models integrating both clinical and radiomics features achieved the highest discriminative performance across key prognostic prediction tasks: predicting HE (pooled C-index 0.822, 95% confidence interval [CI] 0.789-0.855), poor functional outcome (C-index 0.850, 95% CI 0.830-0.869), and mortality (C-index 0.860, 95% CI 0.809-0.911)
C-index 0.822 (95% CI 0.789 to 0.855)
search to September 2025
models integrating both clinical and radiomics features achieved the highest discriminative performance across key prognostic prediction tasks: predicting HE (pooled C-index 0.822, 95% confidence interval [CI] 0.789-0.855), poor functional outcome (C-index 0.850, 95% CI 0.830-0.869), and mortality (C-index 0.860, 95% CI 0.809-0.911)
33 records · page 2 of 3
Showing the records in one thread of the summary: Chatbots and language models. Show every record
Researchers built RARE-PHENIX, a framework that chains together three steps of rare disease phenotyping: large language model extraction of features from clinical text, mapping of those features to Human Phenotype Ontology terms, and supervised ranking of the terms most relevant to rare disease.
npj Digital Medicine
Chronic kidney disease (CKD) presents a growing public health challenge in China, exacerbated by low patient awareness and limited nephrology resources.
Researchers built a framework called CALF that combines counterfactual data augmentation, in which surveillance-relevant attributes of a post are deliberately altered, with multitask learning.
Researchers tested three reasoning-capable multimodal language models on ten periodontal vignettes, each combining a panoramic radiograph with electronic health record text.
Researchers gave 125 anonymized standing whole-spine radiographs to GPT-5.1 and Gemini 3.1 Pro, each prompted once in a separate chat session, and compared their Cobb angle measurements with a joint expert-consensus…
Researchers at multiple centers retrospectively tested ChatGPT-5.2 and Gemini 3.0 against calibrated endodontists on 645 teeth needing a first root canal.
Researchers tested GPT-5 under four API reasoning-effort settings and the small open-weight medical model MedGemma-1.5-4b-it on eight anonymized dental case reports containing images and clinical…
A systematic review and meta-analysis of 48 studies published through March 2025 synthesized diagnostic accuracy of generative AI in radiology and compared it to physician performance using multilevel…
Objective To evaluate the real-world diagnostic performance of a multimodal large language model (LLM) for image-based assessment of oral mucosal lesions compared with clinicians of varying expertise.
On September 9, 2026, a peer-reviewed comparative study reported testing ChatGPT, Gemini, and Microsoft Copilot on 20 radiological cases split between congenital anomalies and tumors.
On September 4, 2026, a peer-reviewed study in Academic Radiology evaluated 25 musculoskeletal imaging cases where three AI systems generated visual summaries from report text alone.
Multimodal artificial intelligence, albeit showing great potential in computational pathology, remains limited to isolated patch-level interpretation and often fails to analyze gigapixel-scale…
Researchers introduced EndoVLM, a vision-language assistant tailored for gastrointestinal endoscopy, using ConvNeXt as a hierarchical visual encoder instead of Vision Transformers and a three-stage…
Researchers retrospectively tested ChatGPT on 300 histopathologically confirmed oral lichen planus cases with at least 24 months of follow-up, using serial clinical records, intraoral photographs, and…
Researchers tested Gemini 3.1 Pro in a zero-shot setting to detect lumbar disc herniations on sagittal MRI from 119 SPIDER cases (26% prevalence).
ace