搜索结果：NEJM AI

More than 50 years after Hart proposed the inverse care law, artificial intelligence (AI) in health care risks repeating the same pattern: Those who could benefit most have the least access. Using publicly available U.S. data, we show that rural areas face greater health burdens but have fewer health care resources and lower capacity to implement AI solutions. These disparities most likely extend beyond the United States. Age-adjusted mortality rates, chronic disease prevalence, and socioeconomic challenges increase with rurality, while the health care workforce and infrastructure decline. This mismatch creates contexts where AI could be most beneficial. However, AI implementation capacity declines from metropolitan to rural areas across key indicators, including interoperability infrastructure, AI adoption, and large language model readiness. In addition, clinical AI systems trained predominantly on urban populations raise concerns about distribution shift and transportability when applied to rural populations. Without deliberate intervention, AI risks amplifying rather than addressing existing disparities. Addressing this misalignment requires coordinated policy, research, and regulatory efforts that explicitly account for geography and equity. Policy should support foundational infrastructure in underserved systems, research should evaluate AI against current care alternatives rather than ideal standards, and regulation should address disparities in access and diffusion to ensure AI benefits reach areas of greatest clinical need.

Sleep underpins cognition, disease prevention, and overall brain health, yet objective, integrative biomarkers of brain health remain lacking. We hypothesized that overnight sleep electroencephalography (EEG) could provide a substrate for such a biomarker. We asked whether a newly developed, end-to-end, data-driven deep learning framework for sleep EEG can learn a latent representation of brain health and distill it into a single score relevant to cognition, disease status, and mortality. We analyzed 36,000 polysomnography recordings from 27,000 subjects from six cohorts. EEG data were represented as one-dimensional time series or a two-dimensional time-frequency spectrogram. A multitask deep neural network, trained end-to-end without expert-defined features, learned a 1024-dimensional brain health latent space and jointly predicted cognitive performance, disease status, and sleep metrics. The latent representation was additionally distilled into a single brain health score. We compared performance with demographic baselines, conventional EEG metrics (e.g., rapid eye movement fraction, spindle density), and classic multivariate machine learning approaches. The deep learning-derived brain health scores consistently surpassed demographic and expert-defined EEG feature models. For cognitive outcomes, correlations (r) rose from small (demographic-only) to moderate (up to r=0.40), while disease classification areas under the receiver operator curve improved from 0.50-0.55 at baseline to 0.65-0.75. In age-adjusted Cox models, a one-standard-deviation increase in the brain health score was associated with a 31%-35% reduced risk of mortality (hazard ratio 0.65 to 0.69; P<0.0001), topping conventional EEG metrics. Gains over classic machine learning, plus latent space visualization, indicated that both established physiological markers and novel EEG features drove enhanced performance. A multitask, end-to-end deep learning approach generated an interpretable, sleep-derived brain health biomarker. By modeling cognition, disease, and mortality, this framework provides a robust index of brain health and may be extended to additional modalities, further enhancing its clinical utility. (Funded by the National Institutes of Health and others.).

Patients navigating a fragmented health care system may feel increasingly tempted to turn to publicly available large language models for quick answers to clinical questions; however, these tools were not built with patient safety, risk stratification, or escalation pathways in mind. In this case study, the authors describe how Included Health designed, piloted, and clinically governed a risk-stratified artificial intelligence (AI) digital assistant that offered generalized health guidance while reliably routing higher-risk situations to human clinicians. Building on OpenAI's generative pretrained transformer 4 (GPT-4) model, the team created a multitier risk classification engine that separated emergency, high-risk, and standard-risk patient inquiries; developed conservative safety guardrails that blocked AI advice and triggered escalation for concerning symptoms; and ran a continuous human-in-the-loop audit program that reviewed 100% of clinical interactions during the pilot. Using a randomized rollout to half of the patient population, the authors found that the risk-stratified assistant maintained a high level of clinical safety (96% accurate guidance, 0% critical safety events, and no AI-generated diagnoses) while reducing standard-risk queries routed to human support by 65%, shortening average human response times from 9.6 to 3.6 minutes, and improving resolution of health inquiries without additional visits. This blueprint illustrates how health care organizations can pair proactive risk analysis, adversarial testing, and ongoing governance to deploy patient-facing generative AI that is explicitly designed to put safety ahead of convenience and still meet patients' expectations for timely, trustworthy guidance.

Multimodal large language models (LLMs) that can interpret clinical text and images are emerging as potential decision-support tools, yet their accuracy on standardized cases and how it compares with human performance across different difficulty levels remains largely unclear. This study aimed to rigorously evaluate the performance of four leading LLMs on the 200-item New England Journal of Medicine (NEJM) Image Challenge. We assessed OpenAI o4-mini-high, Claude 4 Opus, Gemini 2.5 Pro, and Qwen 3, and benchmarked the top model against three medical students (Years 5-7) and an internal-medicine attending physician under identical test conditions. Additionally, we characterized the dominant error types for OpenAI o4-mini-high and tested prompt engineering strategies for potential correction. Our results suggest that OpenAI o4-mini-high achieved the highest overall accuracy of 94%. Its performance remained consistently high across easy, moderate, and difficult cases. The human accuracies in this cohort ranged from 38.5% for three medical students to 70.5% for an attending physician-all significantly lower than OpenAI o4-mini-high. An analysis of OpenAI o4-mini-high's 12 errors revealed that most (83.3%) were outputs reflecting lapses in diagnostic logic rather than input processing. Notably, simple prompting techniques like chain-of-thought and few-shot learning corrected over half of these initial errors. Within the context of this standardized challenge, a leading multimodal LLM delivered high diagnostic accuracy that surpassed the scores of both peer models and the recruited human participants. However, these results should be interpreted as evidence of pattern recognition capabilities rather than human-like clinical understanding. While further validation on real-world data is warranted, these findings support the potential utility of LLMs in educational and standardized settings, highlighting that most residual errors are due to logic gaps that can be partly mitigated by refined user prompting, emphasizing the importance of human-AI interaction for maximizing reliability.

Postdeployment monitoring of artificial intelligence (AI) systems in health care is essential to ensure their safety, quality, and sustained benefit - and to support governance decisions about which systems to update, modify, or decommission. Motivated by these needs, the authors developed a framework for monitoring deployed AI systems organized around three complementary principles: system integrity, performance, and impact. System integrity monitoring focuses on maximizing system uptime, detecting runtime errors, and identifying when changes to the surrounding information technology ecosystem have unintended effects. Performance monitoring focuses on maintaining accurate and equitable system behavior in the face of changing health care practices (and thus input data) over time. Impact monitoring assesses whether a deployed system continues to have value in the form of benefit to clinicians, staff, and patients. Drawing on examples of deployed AI systems at their academic medical center, the authors provide practical guidance for creating monitoring plans based on these principles that specify which metrics to measure and at what cadence, who is responsible for acting when metrics change, and what concrete follow-up actions should be taken - for both traditional and generative AI. They also discuss challenges in implementing this framework, including the effort of monitoring for health systems with limited resources, and the difficulty of incorporating data-driven monitoring practices into complex organizations where conflicting priorities and definitions of success often coexist. This framework offers a starting point for health systems seeking to ensure that AI deployments remain safe and effective over time.

Current evaluations of Large Language Models (LLMs) in medicine primarily focus on text-based benchmarks, leaving their multimodal diagnostic capabilities in complex, real-world clinical scenarios largely undefined. Furthermore, comparisons against large-scale human benchmarks remain scarce. To address this gap, we conducted a comprehensive evaluation of state-of-the-art multimodal LLMs (GPT-4o, Claude 3.7, and Doubao) using 272 complex cases from the New England Journal of Medicine Image Challenge (2009&#x2013;2025). Uniquely, we benchmarked AI performance against a massive global dataset of 16,401,888 physician responses, representing the largest comparative study of human-AI diagnostic reasoning to date. Strikingly, all multimodal LLMs significantly outperformed the global physician collective (P&#x2009;<&#x2009;0.001). Claude 3.7 achieved a diagnostic accuracy of 89.0%, surpassing the physician majority vote (46.7%) by an absolute margin of over 40 percentage points. Even in challenging cases where human accuracy fell below 40%, Claude 3.7 maintained an accuracy of 86.5%. A novel finding of this study is the remarkably low concordance between high-performing models and physicians (Cohen&#x2019;s &#x3ba;: 0.08&#x2013;0.24). The ratio of model-advantage to physician-advantage cases reached 15.4:1, suggesting that MLLMs succeed in distinct areas where human cognition often falters. Our findings demonstrate that MLLMs have reached a superhuman tier in multimodal diagnostic accuracy. The substantial performance gap, coupled with low human-AI concordance, implies that MLLMs do not merely replicate human knowledge but utilize fundamentally distinct and complementary diagnostic reasoning pathways. These results position multimodal LLMs as critical, independent second readers capable of augmenting clinical decision-making in diagnostically difficult scenarios.

Patients with end-stage kidney disease (ESKD) have a high rate of hospitalizations related to fluid overload and infections. Artificial intelligence (AI)-driven models may improve patient care by predicting the risk of hospitalization. The authors conducted a retrospective, observational matched cohort study of adult patients with ESKD who were receiving value-based hemodialysis at integrated kidney care clinics across the United States in 2023. Two AI-powered machine learning models calculated risk scores (range: 0-1) and the models identified patients with a risk score of 0.64 or above who were at risk for hospitalization within 7 days in relation to infections or fluid status abnormalities. To prevent avoidable hospitalizations, case reviews and interventions were conducted for the patients identified by the models. The AI models generated scores for all patients, but only high-risk scores triggered case review and possible intervention. The authors linked electronic medical records and Medicare claims data and conducted multivariate logistic regression analyses to examine the impact of AI-driven interventions on the odds of all-cause hospitalization in patients with ESKD. A total of 10,294 patients representing 83,928 risk scores were included in the analysis. AI-driven intervention was associated with an 8% reduction in the odds of hospitalization within 7 days (odds ratio=0.92; P=0.025). These interventions were most effective for high-risk patients with scores between 0.64 and 0.85, but had no statistically significant effect for patients with scores above 0.85. Factors that were independently associated with higher rates of hospital admission included a higher risk score (>0.75), chronic high-risk scores, older age, and a higher number of hospital admissions in the year prior. AI-driven interventions were associated with a reduction in the odds of hospitalization among patients with ESKD receiving managed kidney care. These findings underscore AI's potential to assist health care providers with targeted risk interventions for patients with ESKD.

Despite rapid advances in artificial intelligence (AI) across sectors, health care remains one of the least transformed domains. This stagnation is not due to lack of data, clinical need, or innovation, but rather to persistent regulatory and economic misalignment. Even AI tools cleared by the U.S. Food and Drug Administration that meet clinical efficacy standards often face major barriers to adoption, largely driven by outdated reimbursement frameworks and fragmented incentives among stakeholders. The result is a systemic failure to deploy technologies that could meaningfully reduce clinician workload, shorten wait times, and improve patient lives. In this article, we examine the reimbursement landscape for health AI, focusing first on tools that fit existing regulatory pathways, outlining payment barriers and proposing policy reforms. These include resolving Current Procedural Terminology adoption bottlenecks, addressing integration overhead, and aligning pricing models with AI cost structures. We then extend the discussion to the emerging domain of generative AI in health care, highlighting the urgent need for prospective regulatory frameworks to ensure patient benefits. (Funded by the National Institutes of Health and the Leukemia and Lymphoma Society.).

The wide availability of labeled electrocardiogram (ECG) data has driven major advances in artificial intelligence (AI)-based detection of structural and functional cardiac abnormalities and thus ECG-based diagnosis. However, many critical, high value clinical diagnostic applications, such as assessing myocardial ischemia and coronary microvascular dysfunction, remain underserved due to the limited availability of labeled datasets. We developed a self-supervised ECG foundation model and demonstrate how this approach can overcome this limitation. A modified vision transformer model was pretrained using a large database of unlabeled ECG waveforms (MIMIC-IV-ECG, N=800,035). The model was then fine-tuned using smaller databases that included high-quality labels derived from positron emission tomography (N=3,126) and clinical reports (N=13,704) for 12 clinical, demographic, and traditional ECG prediction tasks. Diagnostic accuracy and model generalizability were evaluated across five additional cohorts including the publicly available PTB-XL and UK Biobank databases and labels from cardiac magnetic resonance imaging (MRI) and single photon emission computed tomography (SPECT). Diagnostic performance varied across tasks with area under the receiver operating characteristic curve (AUROC) ranging from 0.763 for detection of impaired myocardial flow reserve (MFR < 2) to 0.955 for impaired left ventricular ejection fraction (LVEF < 35%). Self-supervised learning (SSL) pretraining greatly improved diagnostic accuracy in 11 of the 12 prediction tasks compared to conventional de novo supervised training. The model retained strong performance across three external and two internal cross-modality databases, with AUROC ranging from 0.771 for impaired MFR to 0.949 for impaired LVEF. This versatile ECG foundation model demonstrates that SSL pretraining enhances diagnostic accuracy and generalizability across diverse cardiac diagnostic applications. By enabling effective learning from limited labeled data, this approach supports AI development for complex but clinically critical tasks, such as detecting myocardial ischemia and coronary microvascular dysfunction, where high-quality labels are costly and scarce.

Community-based internal medicine residencies train nearly half of all U.S. IM residents and are essential for staffing underserved areas. These programs often face educational gaps due to limited resources, faculty development, and procedural training opportunities. We designed a 12-month quality initiative for 27 residents at a small community hospital blending new high-fidelity simulation drills (e.g. in-situ rapid-response and procedural labs), restructured didactics (gamified quizzes, NEJM Resident 360&#x2122;, HumanDx cases), enhanced orientation ("intern survival guide"), research education, wellness retreats, community outreach and AI-driven, individualized case feedback. New interventions were implemented alongside existing activities, with targeted outcomes including clinical competency, board readiness, communication skills, and well-being. Over one year, the mean In-Training Examination (ITE) scores rose from 56.7 % to 59.9 % (p = 0.025) and percentile rank by 11.8 points, CAHPS patient communication scores improved from the 66th to 97th percentile (p < 0.05), and resident satisfaction (ACGME survey) increased from 70 % to 84 %. 78 % of residents found simulation "very useful," while 74 % felt supported by wellness initiatives. Gamified learning was highly relevant to 44 %. Sustained impact needs continued leadership and formal GME integration. Investing in resident education directly improves patient care, creating a continuous cycle of clinical excellence.

Hemorrhage is the leading cause of preventable death on the battlefield, yet combat medics lack clinical decision support systems to help stratify hemorrhage risk in trauma casualties. We previously trained the Automated Processing of the Physiological Registry for Assessment of Injury Severity - Hemorrhage Risk Index (APPRAISE-HRI) software to associate patterns in vital signs (heart rate and blood pressure) collected from trauma patients with three HRI levels: I (low), II (average), or III (high). To independently validate APPRAISE-HRI and obtain U.S. Food and Drug Administration (FDA) clearance, we collected trauma registry and continuous vital sign data from 5895 trauma patients (543 with hemorrhagic injuries and 5352 controls) in an emergency department or during prehospital transport to one of eight medical centers. The study outcome was hemorrhagic injury, defined by documented injuries and blood transfusion. Using the likelihood ratio to assess the ability of APPRAISE-HRI to stratify hemorrhage risk, we found that hemorrhagic patients were 6.88 times as likely as controls to be at level III, strongly suggesting the presence of hemorrhage at this level. Similarly, hemorrhagic patients were 0.18 times as likely as controls to be at level I, suggesting the absence of hemorrhage at this level. Hemorrhagic patients were almost as likely as controls to be at level II (0.70 times as likely). Subsequently, the U.S. Department of Defense obtained FDA 510(k) clearance for the artificial intelligence-enabled APPRAISE-HRI Class II device (K233249), the first software as a medical device approved for assessing hemorrhage risk in trauma patients, allowing for triage and identification of casualties who need immediate attention and evacuation. (Funded by the U.S. Army Medical Materiel Development Activity and the Combat Casualty Care Program Area Directorate (CCCPAD) of the U.S. Army Medical Research and Development Command (USAMRDC), Fort Detrick, MD and others.).

Artificial intelligence (AI) is poised to transform the infrastructure of health care. AI can now interpret clinical conversations and automate back-office operations, and will soon be able to deliver clinician-grade care under the direction of a clinician. This model holds particular promise for primary care, where workforce shortages and rising chronic disease burden demand scalable, integrated solutions. A key barrier to adoption is that U.S. reimbursement is not designed for clinical AI agents. Time-based billing structures penalize physicians for using AI tools that enhance productivity. Traditional transaction-based payment models risk misalignment with care delivery. And without guardrails, added AI workforce capacity can inflate utilization and cost. Current payment models risk bypassing physician oversight of AI services, fragmenting care, and undermining integration with value-based systems. The authors propose a payment framework that aligns incentives around clinical AI agents by reimbursing for care delivered through validated workflows rather than per software license or time spent. Payers would reimburse physicians for outputs of care, enabling them to invest in AI tools and, over time, build the foundation for linking payment to measurable health outcomes. This payment architecture keeps AI-delivered care anchored in physician responsibility, preserving accountability while enabling innovation. When combined with the traceability of digitized AI workflows, this approach lays the groundwork for a system that scales care while preventing fraud and misuse.

Missed opportunities for diagnosis are a critical subset of diagnostic errors that can lead to adverse patient outcomes. These errors frequently arise from failures in the diagnostic process, particularly in ensuring that recommended follow-ups are scheduled and completed. In large health systems, such as Parkland Health in Dallas, Texas, which conducts over 500,000 radiologist studies annually, the challenge of reliably identifying and managing follow-up recommendations is amplified by the reliance on structured note templates (macros) within electronic health records. Improper use or modification of these macros can result in missed notifications and suboptimal care. The authors developed and implemented a custom-built artificial intelligence (AI) agent that uses a pretrained large language model designed to act as an additional safety net for the identification and management of recommended follow-ups from radiologist notes. The AI agent reviews clinical impressions, extracts and standardizes key details for follow-up, and integrates these findings into the digital health workflow for patient outreach. Model performance was evaluated on a sample of 10,000 randomly selected radiologist notes and further assessed during 3 months of silent production mode, encompassing over 120,000 unique imaging studies. The AI agent achieved a balanced accuracy exceeding 97% for identifying radiologist notes requiring follow-up, correctly flagging 6.18 times more cases than the existing macro-based system (513 vs. 83 based on a sample of 10,000 studies). It also demonstrated over 94% accuracy in characterizing the timing of follow-up, the recommended procedure, and the underlying abnormality prompting the follow-up. This approach enabled the digital health team to more reliably identify patients in need of follow-up and improved the integration of actionable findings into patient outreach workflows. Implementation of an AI agent as an additional safety net significantly improved the identification of missed diagnostic opportunities in radiologist notes and accurately extracted key details that aid in patient outreach and scheduling. By enhancing the reliability of follow-up identification and standardizing key details, this approach increases the likelihood that patients receive appropriate care with the intention of optimizing health care outcomes in high-volume clinical settings.

Tecovirimat is approved for smallpox treatment under the Food and Drug Administration Animal Rule on the basis of efficacy in nonhuman primate models of mpox (previously known as monkeypox). However, the clinical efficacy of tecovirimat against human clade II mpox is unclear. In a phase 3, international, double-blind, randomized, placebo-controlled trial, we evaluated the efficacy of oral tecovirimat in adults with laboratory-confirmed clade II mpox. Participants were randomly assigned in a 2:1 ratio to receive tecovirimat or placebo for 14 days. The primary outcome was clinical resolution, assessed in a time-to-event analysis in participants with active skin or mucosal lesions. Secondary outcomes included reduction in pain, assessed in all participants with laboratory-confirmed mpox and in those with severe pain at baseline (pain score, 7 to 10; scale, 0 [no pain] to 10 [worst pain imaginable]); complete lesion healing (assessed in a time-to-event analysis); viral DNA clearance; and safety. Of 412 participants who underwent randomization (275 to tecovirimat and 137 to placebo), 344 had laboratory-confirmed mpox, and 336 had active skin or mucosal lesions and were included in the primary analysis. By day 29, the estimated cumulative incidence of clinical resolution was 83% with tecovirimat and 84% with placebo; the competing-risks hazard ratio for clinical resolution was 0.98 (95% confidence interval [CI], 0.74 to 1.31; P&#x2009;=&#x2009;0.89). No substantial between-group differences were seen in pain reduction among participants with severe pain (difference, 0.1 point; 95% CI, -0.8 to 1.0), in complete lesion healing (competing-risks hazard ratio, 0.97; 95% CI, 0.75 to 1.26), or in viral DNA clearance. The incidence of adverse events of grade 3 or higher was similar in the two groups (4% with tecovirimat and 3% with placebo). This trial showed no evidence that tecovirimat therapy shortened the time to clinical resolution, reduced pain, or increased viral clearance among adults with clade II mpox. (Funded by the National Institute of Allergy and Infectious Diseases of the National Institutes of Health; STOMP/A5418 ClinicalTrials.gov number, NCT05534984.).

Large language model (LLM)-based chatbots are increasingly used for behavioral health support. Few studies have rigorously evaluated their advice on alcohol misuse. We evaluated seven publicly available chatbots-including general-purpose and behavioral health-focused tools-in responding to alcohol misuse-related questions. Using a fictional case, we simulated longitudinal chatbot interactions over seven days, using 25 prompts derived from real-world Reddit posts. Using an evaluation framework specific to chatbots, four clinicians independently rated each chatbot's transcript along five domains: empathy, quality of information, usefulness, responsiveness, and scope awareness. Clinicians also assessed secondary dimensions, including stigmatizing language and challenging the user (vs. only validating feelings). We generated descriptive statistics on performance and identified examples of problematic output. Across all chatbots, empathy was the highest-rated domain (mean score 4.6/5) while quality of information was the lowest (mean 2.7/5). There was considerable variation in overall mean performance scores across the chatbots, ranging from 2.1 (SD 1.1) to 4.5 (SD 0.8). There were no significant differences in performance between behavioral health and general-purpose chatbots. All chatbots had one or more examples of guidance deemed inappropriate, over-stated, or inaccurate. All avoided stigmatizing or judgmental language and supported self-efficacy. Chatbots were perceived to vary widely in their ability to support individuals with alcohol misuse. While generally strong in empathy, there is room for improvement in response quality. As chatbot use expands, users and clinicians should be aware of the strengths and weaknesses of chatbots in providing advice on alcohol misuse.

Current preoperative blood ordering relies on outdated, procedure-based maximum surgical blood order schedule (MSBOS) that overlook individual patient risk. Smart Match, a machine-learning (ML) tool that predicts patient-specific transfusion needs, was developed and silently validated. Integrated into real-time workflows, Smart Match demonstrates the potential of ML to replace MSBOS with personalized, data-driven approaches to perioperative blood management. Using 82 variables and 1921 features from the electronic medical record of a tertiary academic medical center, a custom extreme gradient boosting (XGBoost) model was optimized to predict perioperative red blood cell (RBC) transfusions before performing adult elective surgeries, aligning sensitivity with native MSBOS thresholds. The retrospective cohort was split into training, validation, and test sets. Silent prospective validation was conducted in real-world elective cases, comparing Smart Match predictions with both MSBOS and clinician RBC preordering behaviors at multiple time points preceding surgery. Model discrimination, calibration, and clinical utility were assessed. In addition, a hybrid model to estimate daily blood bank needs was developed. Retrospective data consisted of 235,054 cases with a 3.04% (95% confidence interval [CI], 2.97 to 3.11) transfusion rate. The model predictors included medical history, laboratory results, demographics, surgery information, medications, transfusion history, and MSBOS recommendations. The test data achieved an area under the receiver-operating-characteristic curve (AUROC) of 0.94 (95% CI, 0.93 to 0.95) and an area under the precision-recall curve (AUPRC) of 0.57 (95% CI, 0.53 to 0.61). Silent prospective validation (n=24,003, transfusion rate 2.18%, 95% CI, 2.07 to 2.46) maintained a 0.94 (95% CI, 0.92 to 0.95) AUROC and a 0.55 (95% CI, 0.51 to 0.60) AUPRC. Prospectively, the model's sensitivity was 0.72 (95% CI, 0.67 to 0.75) and its positive predictive value was 0.34 (95% CI, 0.31 to 0.37), surpassing both MSBOS and clinician behavior. Our hybrid model outperformed MSBOS for daily RBC needs with a mean absolute error of 12.86 (95% CI, 11.86 to 13.84) versus 13.34 (95% CI, 12.14 to 14.51) on the test set. Smart Match reliably outperforms MSBOS and clinician orders, works in real time, and will next be assessed through a randomized control trial to gauge its usability and clinical impact. (Funded by the National Institute of General Medical Sciences and others.).

Mount Sinai Health System (MSHS), one of New York City's largest academic medical centers, faced a common patient-access challenge: individuals presenting with new symptoms often did not know where to turn, contributing to delayed care, unnecessary clinic and emergency department visits, and inefficient use of provider and facility resources. To improve care navigation, MSHS implemented a scalable, evidence-based, artificial intelligence (AI)-driven, digital self-triage solution. After a competitive market evaluation and request for proposal process beginning in 2021, MSHS selected Clearstep, a platform built in partnership with Dr. Barton Schmitt, the coauthor of the Schmitt-Thompson telephone triage protocols. Clearstep combined a probabilistic natural-language processing layer with a rules-based clinical expert system. The solution, branded "Check Symptoms & Get Care," was deployed across the MSHS public website and integrated within the MyMountSinai mobile app (powered by Epic) in early 2023. From over 60,000 visits, approximately 22,000 patients completed digital triage sessions (37% initiation and approximately 80% completion), with high satisfaction (a system usability scale score of 85.5 and approximately 75% of users rating greater than or equal to 8 out of 10), broad after-hours utilization (71% of use outside business hours), and zero reported safety incidents. Blinded clinician comparison and continuous post-deployment review demonstrated 88%-96% concordance with physician triage. The authors describe the team, hurdles, metrics, and a tiered road map so that organizations of varying technical readiness can adapt this model.

This issue of NEJM Catalyst Innovations in Care Delivery includes articles, case studies, and research reports on hospital transfers, outpatient antibiotic therapy, pediatric care quality, pediatric cardiac care, musculoskeletal care costs, frontline clinician support, patient portals, an AI digital assistant, and home-based care.

NEJM Catalyst Insights Council members are excited about artificial intelligence in health care delivery for a range of tasks, and they say it already aids patients and clinicians. But the greatest impacts lie in the future.