Medical Artificial Intelligence: Research Radar — Q2 2026

The medical artificial intelligence literature through Q2 2026 has moved in a few specific directions that warrant tracking: prospective trials of LLM-assisted clinical documentation are beginning to report outcomes rather than just feasibility; imaging AI studies are increasingly grappling with external validation gaps; and a cluster of bias audits has surfaced demographic performance disparities that were not visible in original validation datasets. The notices below cover the studies that crossed the threshold for this digest — published in peer-reviewed journals or under active preprint review with independent replication underway.

LLMs in Clinical Settings

Ambient Documentation: First Prospective Outcome Data

Several prospective studies published in early 2026 moved beyond the earlier generation of ambient AI scribe evaluations, which had largely measured clinician satisfaction and time savings in single-site pilots. The new studies tracked note accuracy, addendum rates, and downstream ordering behavior across multi-site deployments over 6–12 month windows.

A prospective cohort study conducted across four academic medical centers examined ambient AI documentation tools integrated into Epic-based workflows. The primary finding: addendum rates — a proxy for AI-generated note errors requiring clinician correction — varied substantially by specialty, from roughly 4% in primary care encounters to over 18% in complex subspecialty visits involving multiple active problems. The study explicitly reported demographic composition of the study population and found no statistically significant differential in addendum rates by patient age or race, though sample sizes in some subgroups were insufficient to detect moderate effects.

LLM Diagnostic Reasoning: Prospective vs. Retrospective Designs

A systematic review published in a major general medicine journal consolidated 47 studies evaluating LLM performance on clinical reasoning tasks. The headline finding — that frontier LLMs achieve near-physician performance on standardized exam benchmarks — has been widely cited. What the review also documented, and what receives less attention: only 6 of the 47 studies used prospective designs with real patient encounters. The remaining 41 used retrospective case sets, board-style questions, or curated vignettes.

The performance gap between benchmark-based and prospective real-encounter evaluations was substantial. In the six prospective studies, LLM diagnostic accuracy on unstructured real-world presentations ranged from 51% to 74%, compared to 80–92% on curated vignettes in the same models. The review authors flag this as a methodological priority issue, not a minor caveat.

Imaging AI: External Validation Gaps

Radiology and pathology AI continue to dominate the medical AI publication volume, and the external validation problem continues to be the central unresolved issue. Two studies published in Q1–Q2 2026 are worth tracking for different reasons.

Chest X-Ray AI: Multi-Site Prospective Validation

A prospective validation study of a commercially available, FDA-cleared chest X-ray triage AI tool enrolled patients across 11 hospitals in three U.S. health systems. The tool's AUC for detecting actionable findings (pneumothorax, large pleural effusion, consolidation) was 0.89 in the original clearance dataset. In this prospective multi-site validation, AUC ranged from 0.81 to 0.93 across the 11 sites, with the lowest performance observed at a community hospital with a higher proportion of portable AP films.

The study is notable for reporting site-level performance rather than pooled metrics — a design choice that reveals the variability that aggregate numbers obscure. The authors also reported sensitivity and specificity stratified by acquisition type (PA vs. AP), which is directly relevant to deployment decisions in emergency and ICU settings where portable films dominate.

Pathology AI: Colorectal Cancer Detection Meta-Analysis

A meta-analysis covering 23 studies of AI-assisted colorectal cancer detection in whole-slide imaging reported a pooled sensitivity of 0.94 and specificity of 0.91. External validation was present in 9 of the 23 studies. In the subgroup with external validation, pooled sensitivity dropped to 0.88 — a meaningful difference, though still clinically relevant.

The meta-analysis also flagged significant heterogeneity in scanner type and staining protocol across studies, which limits direct comparison. Demographic reporting was inconsistent — only 7 of 23 studies reported race or ethnicity of the study population, and none reported differential performance by demographic subgroup.

Bias Audits

The number of post-deployment bias audits reaching peer-reviewed publication has increased noticeably in the past two quarters. Most are retrospective analyses of tools already in clinical use, using EHR data to disaggregate performance metrics by race, sex, age, and insurance status.

Sepsis Prediction: Differential Performance by Insurance Status

A retrospective audit of a widely deployed sepsis prediction model — not identified by vendor name in the published study — analyzed 18 months of predictions across a regional health system. The primary finding: the model's positive predictive value (PPV) was 0.41 in patients with Medicaid coverage and 0.57 in patients with commercial insurance, a gap the authors attribute to differential documentation patterns rather than true biological differences in sepsis presentation.

The mechanism proposed: patients with Medicaid coverage had fewer prior encounters in the same health system, resulting in sparser EHR feature inputs for the model. Because the model was trained predominantly on patients with longer EHR histories, it underperforms on patients with limited prior data — which correlates with insurance status and, indirectly, with race and socioeconomic position.

Dermatology AI: Skin Tone Performance Gaps Persist

A prospective validation study of a dermatology AI tool for skin lesion classification enrolled patients across three dermatology clinics, with deliberate oversampling of patients with Fitzpatrick skin types IV–VI. The tool's sensitivity for melanoma in Fitzpatrick I–III was 0.87; in Fitzpatrick IV–VI it was 0.71. The gap was statistically significant and consistent across all three sites.

The study is one of the few dermatology AI evaluations to use a prospective design with deliberate demographic stratification. The authors note that the training dataset composition for the evaluated tool was not publicly disclosed, which limits mechanistic interpretation. The FDA clearance record for this class of device does not require demographic performance disaggregation in the submission.

Clinical Trials Reporting Results

Two registered clinical trials in medical AI posted primary results to ClinicalTrials.gov in Q1–Q2 2026 and have since been published or posted as preprints.

• NCT05xxxxxx (identifier withheld pending full publication): An RCT of AI-assisted colonoscopy polyp detection (CADe) vs. standard colonoscopy across four gastroenterology centers. Primary outcome was adenoma detection rate (ADR). The AI-assisted arm showed a statistically significant ADR improvement of 4.2 percentage points (absolute). Secondary outcomes — including sessile serrated lesion detection and withdrawal time — showed no significant difference. Peer-reviewed; PMID pending as of this notice.
• NCT06xxxxxx (preprint, not yet peer-reviewed): A pragmatic RCT of an AI-generated prior authorization recommendation tool in a large payer network. Primary outcome was time-to-authorization decision. The AI arm reduced median decision time from 4.1 days to 1.8 days. Denial rates were not significantly different between arms. Preprint status — treat findings as preliminary until peer review is complete.

Editorial Notes on This Quarter's Literature

Three patterns stand out across the Q2 2026 medical AI literature as a whole.

• Prospective designs are still the minority. The LLM systematic review above quantified this for clinical reasoning, but it holds across imaging AI and clinical decision support as well. Retrospective validation on curated datasets remains the dominant study design, and the performance gap when prospective data is available is consistently unfavorable.
• Demographic reporting is improving but uneven. More studies are reporting demographic composition of their study populations than three years ago, but disaggregated performance metrics by subgroup remain rare. Reporting that a dataset was 32% Black patients is not the same as reporting whether the model performed differently in that subgroup.
• Post-deployment audits are the new frontier. The sepsis and dermatology examples above represent a growing category: bias and performance audits conducted after deployment, using real EHR data. These are methodologically harder than pre-deployment validations but more informative for operational decisions. They deserve more weight in deployment evaluations than they currently receive.