Adolescent major depressive disorder is a highly prevalent and debilitating condition, yet objective and generalizable neurofunctional signatures remain elusive. Existing neuroimaging approaches largely rely on resting-state fMRI, or controlled tasks often lack ecological validity and generalizable signatures of depression-related neural dysfunctions. This study proposes a Visual–Affective Decoding (VAD) framework integrating cross-subject visual and affective decoding from whole-brain fMRI acquired during naturalistic movie viewing. A dual-pathway anomaly detection model trained exclusively on healthy controls establishes normative brain function and identifies MDD through deviations from this norm. In 200 adolescents (112 MDD, 88 HC), VAD achieved an AUC of 0.87, 95.56% sensitivity, 83.89% accuracy, and 72.22% specificity. Using the same anomaly threshold, it generalized to an independent external adolescent cohort (70 MDD, 68 HC), achieving an AUC of 0.81 with an accuracy of 75.36%, supporting its cross-site generalizability. VAD further revealed interpretable neurofunctional signatures, including blunted emotional reactivity, attenuated visual representation, and affective patterns consistent with emotional rumination. Region-level contribution analyses identified dissociable visual- and affective-sensitive cortical regions, consistent with dysfunctions in visual, default-mode, salience, and higher-order associative networks. These findings demonstrate that VAD provides objective, interpretable, and generalizable neurofunctional signatures for adolescent depression diagnosis and clinically relevant biomarker development.
Ambient artificial intelligence (AI) scribes may reduce documentation burden, but little is known about their adoption during enterprise-scale ambulatory rollout. We conducted a retrospective observational cohort study of 2901 ambulatory clinicians at an integrated academic health system using linked electronic health record and vendor usage data from December 2024 through May 2025. Clinician adoption was defined as any use during the study period, and monthly visit-level uptake as the proportion of eligible ambulatory visits in which the ambient AI scribe was used. Adoption and uptake varied substantially across practice settings, specialties, and clinician types. Clinician-level adoption was highest in community internal medicine (67%), community general pediatrics (69%), and academic neurology (53%), and lowest in academic general surgery (16%), obstetrics and gynecology (18%), and psychiatry (19%). Community attending physicians and advanced practice providers showed marked increases in median visit-level uptake, reaching approximately 80%-90% in some groups by May 2025. Distinct longitudinal patterns included slow and steady adoption, early use followed by discontinuation, transient use, and late adoption. These findings suggest that implementation support and evaluation strategies should be tailored to local practice context rather than applied uniformly across a health system.
Artificial intelligence (AI) increasingly supports clinical decisions, but human oversight is protective only if clinicians recognize and resist erroneous output. We systematically reviewed controlled and real-world studies of clinician responses to incorrect automated advice. Forty-six reports represented 41 independent cohort clusters, including 40 core cohorts. Across three adjusted cohorts, clinician correctness was substantially higher with correct than incorrect automated advice (pooled OR 10.19, 95% CI 7.43–13.98; I² = 0%); a ≥ 2020 sensitivity analysis was similar but imprecise (OR 9.22, 95% CI 2.06–41.13). In four heterogeneous real-world computerized ECG studies, 26.6% of erroneous interpretations were not corrected or recognized (95% CI 6.4–65.6%; I² = 95.8%). Contemporary AI and large language model studies also documented harmful reliance but were too heterogeneous for pooling. Patient-level adverse events were reported in one retrospective cohort. Human oversight should be evaluated as an error-interception capability rather than assumed to ensure safety.
Standardized patients (SPs) are central to clinical communication training but remain constrained by cost, scalability, and reliance on trained actors. We present AI standardized patients (AI-SPs), large language model–driven patient simulators built on a three-layer information architecture that adaptively modulates information disclosure according to learner skill. In Phase 1, blinded expert assessment of 350 simulated consultations generated by five frontier LLMs showed that, under a common architecture, learner skill level accounted for more of the observed performance variation than model choice (η2 = 0.31 vs. 0.06); whether this reflects the architecture itself cannot be determined without comparing models with and without it. In Phase 1b, analysis of 155 live student consultations identified systematic failures to elicit safety-critical information, enabling automated curriculum diagnostics without expert observation. In a Phase 2 three-arm pilot randomized study (n = 58), OSCE performance gains with AI-SP training did not differ significantly from those with human SP training; the trial was designed to estimate effect sizes for a future confirmatory study, not to demonstrate equivalence. Self-efficacy gains were numerically largest with AI-SP training, but this exploratory difference did not survive correction for multiple comparisons. These findings suggest that architecture-driven AI-SPs could provide a scalable, model-portable approach to clinical communication training.
We tested whether technology-based structural racism mechanisms (digital redlining, path dependency, technical difficulties) were associated with lower telehealth video visit rates (versus audio-only; in-person visits excluded) among Black compared to White patients. Using structured and unstructured electronic health record data from 137,297 patients with completed telehealth visits over 26.5 months linked with broadband market data, we examined correlates of racial differences in video versus audio-only visits. Patient race had no significant unadjusted effect, but adjustment for individual and zip code level confounders revealed a 6% reduction in video visit rate ratios for Black patients, widening to 7–9% with interaction terms. For digital redlining, broadband market conditions were worse in areas with high neighborhood disadvantage and Black residential composition. However, broadband conditions were not associated with video visit rates. For path dependency, portal nonenrollment was more common among Black patients but had a stronger negative association with video visit rates among White patients. For technical difficulties, Black patients experienced more difficulties, which were associated with greater odds of a subsequent audio-only visit—particularly among Black patients. Findings suggest technology-based structural racism may operate through conditional and sequential mechanisms across telehealth stages. Health systems should monitor multilevel telehealth inequities across care stages.
