David Jh Wu

Clinical artificial intelligence (AI) has advanced rapidly, with frontier large language models now matching or exceeding physician performance on simulated diagnostic reasoning and clinical decision-support tasks. Yet adoption has outpaced the evidence base: fewer than 5% of cleared U.S. Food and Drug Administration (FDA) AI/machine learning (ML) devices have undergone peer-reviewed evaluation and prospective randomised trials remain scarce. This narrative review synthesises evidence from ARISE’s State of Clinical AI Report 2026 across six domains: model performance and human benchmarking, evaluation methodology and benchmark validity, foundational methods including multimodal and multiagent systems, clinical workflow integration, patient-facing applications and applied domain-specific AI. We conducted a targeted literature search of PubMed, medRxiv and arXiv supplemented by expert nomination, including empirical studies of AI in clinical contexts. Across domains, frontier models demonstrate strong clinical reasoning capabilities but exhibit persistent failures in uncertainty calibration, metacognition and robustness to distributional shift. Standard benchmarks have saturated, prompting development of multidimensional evaluation frameworks that assess safety, real-world workflows and agentic capabilities. New trends in foundational methods include converting medical data into tokens and developing multiagent, multimodal systems. In clinical workflows, early prospective trials in clinical decision support and diagnostic imaging show promising results, though human-AI collaboration remains suboptimal and risks of automation bias and clinician deskilling are emerging. In tandem, patient-facing AI is progressing with more personalised health assistance; with this comes the risk of patient overtrust, raising the bar for guardrails. The most credible translational progress occurs on narrowly defined tasks with clear endpoints, while broader clinical autonomy awaits advances in evaluation and human-AI interaction.

Doctors increasingly rely on AI in the clinic, yet which report features make AI-generated responses useful and trustworthy remains unclear. In this randomized mixed-methods study, 34 oncology physicians provided 294 ratings of four blinded AI systems across five vignettes, alongside 20 semi-structured interviews analyzed with a prespecified LLM-assisted qualitative pipeline. Despite similar references, an evidence-graded report adapted from OpenEvidence was rated significantly lower in overall utility than standard OpenEvidence (mean difference, −0.96; 95% CI, −1.26 to −0.66; P<.001). Qualitative analysis identified six themes and seven design requirements. Oncologists valued rapid orientation, evidence retrieval, and verification, preferring concise, scannable reports with quantitative outcomes, recognizable bolded guidelines, explicit uncertainty, and verifiable citations. Trust deteriorated with citation mismatch, buried provenance, evidence misclassification, overconfident recommendations, and poor organization. Evidence presented differently can alter perceptions of clinical utility and trust; accuracy alone is insufficient, and report design must also be empirically evaluated.

' Context engineering is the delicate art and science of filling the context window with just the right information for the next step.” Andrej Karpathy

Specialist consults in primary care and inpatient settings typically address complex clinical questions beyond standard guidelines. eConsults have been developed as a way for specialist physicians to review cases asynchronously and provide clinical answers without a formal patient encounter. Meanwhile, large language models (LLMs) have approached human-level performance on structured clinical tasks, but their real-world effectiveness requires evaluation, which is bottlenecked by time-intensive manual physician review. To address this, we evaluate two automated methods: LLM-as-judge and a decompose-then-verify framework that breaks down AI answers into verifiable claims against human eConsult responses. Using 40 real-world physician-to-physician eConsults, we compared AI-generated responses to human answers using both physician raters and automated tools. LLM-as-judge outperformed decompose-then-verify, achieving human-level concordance assessment with F1-score of 0.89 (95% CI: 0.750, 0.960) and Cohen’s kappa of 0.75 (95% CI 0.47,0.90) —comparable to physician inter-rater agreement κ = 0.69-0.90 (95% CI 0.43-1.0).

Purpose: Colorectal cancer (CRC) is the third–most frequently diagnosed cancer globally. Studies have linked low serum albumin with increased risk of CRC, but the causal nature of the association remains unclear. In the present study, we explored the potential causal relationship using bidirectional Mendelian randomization (MR). Methods: Instrumental variants for albumin were obtained from a genome-wide association study (GWAS) on 102,223 Eastern Asian participants to investigate the effect of albumin on CRC. Summary statistics of CRC were obtained from a GWAS on 7,062 CRC cases and 195,745 controls of Eastern Asian ancestry. Bidirectional MR analysis was performed using inverse variance weighting (IVW) for primary analysis, supplemented with a maximum likelihood–based method, MR-PRESSO test, leave-one-out analysis, and MR-Egger regression. Stratification analyses were further performed. Results: We found that genetically predicted serum albumin per unit was associated with a lower risk of CRC (OR 0.75, 95% CI 0.59– 0.95 with IVW). No evidence of pleiotropy was observed. Sex-stratified MR analysis showed that serum albumin was inversely associated with risk of CRC in men (OR 0.71, 95% CI 0.53– 0.96), but not in women (OR 0.81, 95% CI 0.55– 1.19) using IVW. Reverse MR analysis suggested a genetic predisposition toward CRC was not associated with serum albumin. Conclusion: Our study revealed a suggestive sex disparity in the effect of albumin, which deserves further exploration of the potential biological mechanism.

Background: Although observational epidemiological studies have found that smoking is positively associated with risk of rheumatoid arthritis (RA), assessing the causality of this relationship has remained elusive because conventional observational studies are susceptible to bias such as confounding and reverse causation. Here, we applied the Mendelian randomization (MR) approach to examine the potential causal relationship between smoking and risk of RA. Methods: Summary statistics data for RA were obtained from a meta-analysis of genome-wide association studies (GWAS), including 14,361 RA cases and 43,923 controls of European ancestry. The instrumental variables (IV) and the genetic association estimates for smoking initiation and lifetime smoking were obtained from a GWAS meta-analysis including 1,232,091 individuals and a GWAS of 462,690 individuals of European ancestry, respectively. MR analyses were performed using the inverse-variance weighted (IVW) method and supplemented with the weighted-median method. Potential pleiotropy was assessed using the MR-Pleiotropy RESidual Sum and Outlier (MR-PRESSO) test and MR-Egger regression. Sensitivity analyses were further performed to test the robustness of the association. Results: We found that compared with never smokers, genetic predisposition to smoking initiation was positively associated with risk of RA (odds ratio (OR) = 1.32, 95% confidence interval (CI) = 1.15–1.52, P = 9.17 × 10⁻⁵ using the IVW method). Similarly, genetically predicted lifetime smoking was associated with an increased risk of RA (OR = 1.55, 95% CI = 1.13–2.14, P = 0.007). Sensitivity analyses using alternative MR methods and different sets of IVs produced similar results, suggesting the robustness of our findings. Conclusions: These results provide support for a causal association between smoking and increased risk of RA. Further studies are warranted to explain the underlying mechanisms of smoking in the development of RA.

Objective: Observational epidemiological studies have reported a relationship between coffee intake and risk of stroke. However, evidence for this association is inconsistent, and it remains uncertain whether the association is causal or due to confounding or reverse causality. To clarify this relationship, we adopted a Mendelian randomization (MR) approach to evaluate the effects of coffee consumption on the risk of stroke and its subtypes.Methods: A meta‐analysis of genome‐wide association studies (GWASs) including 91,462 coffee consumers was used to identify instruments for coffee consumption. Summary‐level data for stroke, intracerebral hemorrhage, ischemic stroke (IS), and IS subtypes were obtained from GWAS meta‐analyses conducted by the MEGASTROKE consortium. MR analyses were performed using the inverse‐variance‐weighted, weighted‐median, MR‐PRESSO (Pleiotropy RESidual Sum and Outlier) test and MR‐Egger regression. Sensitivity analyses were further performed using alternative instruments to test the robustness of our findings.Results: Genetically predicted coffee consumption (high vs infrequent/no) was not associated with risk of stroke. Similarly, among coffee consumers, MR analysis did not indicate causal associations between coffee consumption (cups/day) and risk of stroke. However, in the subgroup analysis, we found weak suggestive evidence for a potential protective effect of coffee consumption on risk of small vessel (SV)‐IS, although the association did not reach statistical significance after correction for multiple comparisons.Interpretation: This study suggests that coffee consumption is not causally associated with risk of stroke or its subtypes. Further studies are warranted to elucidate the possible association between coffee intake and risk of SV‐IS, as well as its potential underlying mechanisms. ANN NEUROL 2020;87:525–532

Objective: To evaluate the telomere length (TL) in patients with RA relative to that in controls and to test whether TL is causally associated with risk of RA. Methods: Systematic review and meta-analysis of relevant literature was conducted to evaluate the association between TL and RA. Standardized mean differences with 95% CIs of TL in RA patients relative to controls were pooled using fixed or random-effects models. TL-related single-nucleotide polymorphisms were selected from a genome-wide association study of 37 684 individuals, and summary statistics of RA were obtained from a genome-wide association study meta-analysis including 14 361 RA patients and 43 923 controls. Mendelian randomization was performed using the inverse-variance weighted, weighted-median and likelihood-based methods. Sensitivity analyses were performed to test the robustness of the association. Results: In the meta-analysis of 911 RA patients and 2498 controls, we found that patients with RA had a significantly shorter TL compared with controls (standardized mean differences = −0.50; 95% CI −0.88, −0.11; P = 0.012). In the Mendelian randomization analysis, we found that genetically predicted longer TL was associated with a reduced risk of RA [odds ratio = 0.68; 95% CI 0.54, 0.86; P = 0.002 using the inverse-variance weighted method]. Sensitivity analyses using alternative Mendelian randomization approaches yielded similar findings, suggesting the robustness of the causal association. Conclusion: Our study provides evidence for a negative causal association of TL with risk of RA. Further studies are warranted to elucidate the underlying mechanism for the role of telomeres in the development of RA.