DR. BRIAN WILLIAM PATTERSON, M.D.

Plain English
Researchers used a structured survey process with emergency department clinicians to rate how strongly 32 workplace factors influence the decision about where to send a patient after their ED visit. About 59% of factors were rated similarly regardless of how busy the department was, but the rest were rated differently under low versus high demand. Knowing which factors change with demand helps prioritize what to redesign to make patient discharge decisions safer and more reliable.

Plain English
Researchers randomly assigned adults with overweight or obesity and short sleep to either extend their sleep by about one hour per night or keep their usual schedule for six weeks. The group that extended sleep slept more consistently and reported better sleep quality, but their insulin sensitivity and blood sugar control did not improve compared to the group that changed nothing. Getting more sleep helps sleep health but does not fix the metabolic problems linked to short sleep in people with obesity.

Plain English
Researchers observed 20 emergency department visits and interviewed 18 clinicians to map the factors that shape how doctors decide where to send patients after evaluation. They found that physical environment had not been previously recognized as an influence, and that some factors behave differently depending on how busy the department is. Understanding these variations is a necessary step toward redesigning the decision-making process to be safer and more consistent.

Plain English
This study tracked how the body clears blood sugar after a meal in people with obesity and type 2 diabetes who lost about 20% of their body weight — either through diet alone or gastric bypass surgery. Diet-induced weight loss doubled the body's insulin-driven sugar uptake, producing better blood sugar control than surgery even when weight loss was equal. Surgery still powerfully improves diabetes, but the mechanism differs from diet-driven weight loss.

Plain English
A pilot study across 10 hospitals tested a wearable optical sensor that estimates cardiac troponin levels — a protein that rises in the blood when the heart is damaged — without drawing blood, completing the measurement in three minutes. In 840 patients, the device correctly identified elevated troponin levels with 90% sensitivity and achieved a C-statistic of 0.90. The device could speed up evaluation of patients with chest pain and reduce the need for blood draws in emergency settings.

Plain English
This is an earlier version of the same ambient AI implementation study published as PMID 40959192, describing a framework for deploying AI-generated clinical notes, monitoring their accuracy, and running an embedded randomized trial. A brief template change sharply reduced billing code accuracy until it was corrected. The publicly available protocols are intended to help other health systems implement ambient AI without compromising documentation quality.

Plain English
Researchers built a system called DR.KNOWS that connects AI language models to a structured medical knowledge database to improve diagnostic predictions from patient records. When tested on real electronic health record data from two geographic regions, DR.KNOWS produced more accurate diagnostic concepts than standard language model approaches alone. Linking structured medical knowledge to AI reasoning reduces diagnostic errors and brings AI clinical decision support closer to being trustworthy enough for routine use.

Plain English
Researchers redesigned the emergency department discharge process for older adults using a human-centered approach, introducing a new discharge instructions template, a redesigned patient handout, structured nurse training, and embedded call-follow-up prompts. They evaluated the redesign from the perspectives of doctors, nurses, and patients, finding broad acceptance and specific usability strengths in the new patient handout. The study identifies remaining design gaps — such as the need for pre-populated information — and provides a model for iterative improvement in ED discharge safety.

Plain English
Researchers trained a machine learning model on nearly one million hospital admissions and tested it prospectively across 21 hospitals to see how well it could predict which ward patients would deteriorate and need an ICU transfer or die within 24 hours. The model outperformed existing bedside warning scores across a wide range of patient types and remained accurate during real-time prospective testing, leading to FDA clearance for clinical use. Gradient-boosted machine learning substantially improves early detection of deteriorating patients compared to simpler scoring systems.

Plain English
Researchers embedded an AI screening tool into hospital electronic health records to flag patients at risk for opioid use disorder and prompt clinicians to refer them for addiction medicine consultations. The AI-assisted approach produced referral rates equivalent to the prior standard of care, while also reducing 30-day hospital readmissions by nearly half among referred patients. The tool offers a scalable, cost-effective way to systematically identify patients who would benefit from addiction medicine support.

Plain English
This is a preprint version of the same study published as PMID 41193667, validating an AI-as-a-judge framework for evaluating AI-generated clinical summaries from electronic health records. The best-performing model matched human reviewers closely and finished evaluations in seconds. Automating quality checks on AI-generated summaries makes large-scale safety monitoring of these tools feasible.

Plain English
A natural language processing algorithm was applied to emergency department provider notes to find older adult patients who came in because of a fall, then compared to the standard approach of using billing diagnosis codes. The algorithm found nearly twice as many fall-related visits as the code-based method, and the extra patients it identified were sicker overall, often coming in for serious underlying illnesses where a fall was a consequence rather than the main problem. Standard coding significantly undercounts fall visits, and the specific definition used can change who gets identified and how their outcomes look.

Plain English
Researchers interviewed and observed emergency department clinicians and patients to identify what goes wrong during care transitions for older adults, then used that information to define what a patient-carried safety document should contain. They identified 21 categories of barriers and facilitators and translated them into 5 design guidelines and 5 design requirements for such a document. The work provides a human factors foundation for creating a practical tool that patients can carry across care settings to reduce errors during transitions.

Plain English
Researchers built a governance and monitoring system for deploying an AI tool that listens to clinic visits and writes clinical notes, then embedded a randomized trial within routine care at a health system. They tracked how often clinicians used the tool, caught a workflow change that briefly cut billing code accuracy in half, and fixed it through retraining and template redesign. The published framework and monitoring protocols give other health systems a practical blueprint for rolling out ambient AI safely.

Plain English
Many hospitals, especially in rural and underserved areas, lack the staff and infrastructure to safely adopt AI tools, creating a widening gap between resource-rich and resource-poor health systems. This paper argues that a hub-and-spoke support network, connecting lower-resource hospitals to technical and legal expertise from better-equipped institutions, could close that gap. The authors point to their Health AI Partnership pilot as a proof of concept and call for federal investment to scale the model nationally.

Plain English
Researchers tested whether an AI system could evaluate the quality of AI-generated patient record summaries as reliably as human clinicians. The best-performing AI evaluator matched human ratings closely, completing each assessment in 22 seconds with agreement comparable to that between human reviewers. Automating this evaluation process makes it practical to continuously monitor AI-generated summaries for accuracy and safety at scale.

Plain English
This entry is a published correction to a previously published article and contains no study findings to summarize.

Plain English
This study developed a new method to separately measure insulin-driven and non-insulin-driven sugar disposal after a glucose drink across a range of metabolic states — from lean and healthy to obese with type 2 diabetes. As insulin action worsens, the body increasingly relies on non-insulin pathways to clear blood sugar, but these backup mechanisms are not sufficient to prevent high glucose in people with diabetes. Weight loss therapy restored the insulin-driven component.

Plain English
This study examined three people who naturally lack the protein ANGPTL3 — a regulator of fat metabolism — and compared them to matched controls. People without ANGPTL3 had triglyceride levels more than 10 times lower than normal, due to both reduced liver output of fat particles and faster clearance from blood. The findings show ANGPTL3 does more than just block fat-breakdown enzymes in the bloodstream; it also controls how much fat the liver releases.

Plain English
Emergency departments are increasingly expected to screen older patients for conditions like fall risk, cognitive impairment, and end-of-life care needs, but the added workload makes this difficult to implement. This paper proposes using automated risk scoring built into electronic health records to triage which patients need hands-on assessment, with different levels of readiness for different clinical conditions. Automated screening could extend geriatric emergency care to more patients without requiring clinicians to screen everyone manually.

Plain English
Researchers built a monitoring dashboard for a machine-learning clinical decision support tool that tracks not just model accuracy but the full chain from algorithm output to clinical action and patient outcomes. Using alluvial charts and control charts, the dashboard helped the development team and hospital stakeholders spot and fix problems in how the tool was being used. Monitoring ML-based clinical tools requires watching the entire care system, not just the algorithm's predictions.

Plain English
This is a conference proceeding version of the same eCARTv5 early warning score study published as PMID 40138535, describing development and validation of a gradient-boosted machine learning model to detect deteriorating hospital ward patients. The model outperformed standard scoring tools across nearly two million admissions and held up during prospective real-world testing. It received FDA clearance based on these results.

Plain English
Researchers compared detailed metabolic measurements in adults with obesity who were metabolically healthy versus metabolically unhealthy, including tests of muscle, liver, and fat tissue function. People with metabolically healthy obesity had more favorable muscle and fat tissue biology, lower circulating fats and sugars throughout the day, and less oxidative stress than those with metabolically unhealthy obesity. These findings outline specific biological mechanisms that may explain why some people with obesity avoid the metabolic complications that others develop.

Plain English
Researchers conducted brief structured visits with 16 emergency medicine physicians to educate them about a clinical decision support tool that identifies older adults at high fall risk and suggest referrals to a falls clinic. The conversations revealed five categories of factors shaping whether clinicians used the tool, including design features, misconceptions about the referral process, and ED busyness. Academic detailing — brief expert visits to clinicians — is a practical way to both teach clinicians how a tool works and simultaneously gather feedback to improve it.

Plain English
Researchers studied 576 heart attack patients across 10 emergency departments to understand how quickly getting an electrocardiogram affects the time to emergency artery-opening treatment. Patients who got their ECG within 10 minutes had shorter wait times for the procedure, and delays occurring during the triage window (11-30 minutes after arrival) had the strongest association with downstream treatment delays. Focusing efforts on getting ECGs done before or during triage, rather than after, would help the most patients reach timely treatment.

Plain English
Researchers analyzed electronic health records from over 730,000 emergency department visits spanning 17 years to measure how much written clinical information has accumulated in patients' records over time. The median number of notes available to an ED physician grew from 5 in 2006 to 359 by 2022, representing a more than 30-fold increase in text volume. This growth makes it nearly impossible for clinicians to review a patient's full record during a visit, making AI-based summarization tools increasingly necessary.

Plain English
Researchers analyzed Medicare claims for over 5,000 autistic older adults who died between 2013 and 2021 and compared their use of hospice care to matched non-autistic peers. Autistic older adults used hospice at similar rates and with similar timing as non-autistic adults, and no additional disparities based on sex, race, or rural location were found. Universal Medicare hospice coverage may be equalizing end-of-life care access for autistic older adults in ways that are not seen in other areas of healthcare.

Plain English
Researchers interviewed 15 older adults in the emergency department about a clinical decision support tool designed to identify their fall risk and refer them to a falls prevention clinic. Patients viewed the tool as simply another resource their doctor used and did not express concern about automated screening; the main barriers to attending the falls clinic were practical ones like transportation and distance. Future fall prevention tools need to address access barriers to follow-up care, not just the screening step.

Plain English
This study compared several widely used formulas for assessing how well the pancreas secretes insulin (beta-cell function) across groups ranging from lean and healthy to obese with type 2 diabetes. Different formulas produced contradictory rankings of beta-cell function among the same groups, and none accurately reflected the actual measured relationship between insulin secretion and blood glucose. Standard beta-cell function indices can be misleading and should be interpreted with caution.

Plain English
Researchers measured how the body produces and clears glucose during fasting in 239 people with varying blood sugar levels, from lean and normal to obese with diabetes. In people with mildly elevated fasting blood sugar, glucose production and clearance were both elevated but matched each other; in those with diabetic-range fasting glucose, clearance was severely impaired. Poor glucose clearance — not just excess production — is a central driver of fasting high blood sugar in obesity.

Plain English
This study measured fat clearance rates in women with obesity who had either high or low amounts of calorie-burning brown fat. Women with more brown fat had lower blood triglyceride and free fatty acid levels not because they produced less fat, but because they cleared it from the blood faster. Brown fat appears to protect against high blood fat levels by accelerating the removal of circulating fats.

Plain English
Researchers developed a method to measure how quickly different blood fat proteins (apolipoproteins) are made and cleared in mice using tiny blood samples and mass spectrometry. The technique works in standard lab mice and in two common mouse models of heart disease, revealing distinct differences in how these proteins are processed in each model. This tool enables detailed fat metabolism studies in mice that were previously only feasible in humans.

Plain English
Researchers trained machine learning models to predict emergency department crowding — specifically, how many patients with respiratory illness were waiting longer than four hours — at two hospitals with different characteristics and tested them across time periods including before and during the COVID-19 pandemic. Models held up well when applied to data from the same hospital during a different time period, but performed poorly when transferred directly to a different hospital. ED crowding can be predicted, but models need to be trained on local data rather than transferred directly from other institutions.

Plain English
Researchers administered satisfaction surveys to older adults who visited the emergency department after a fall and cross-referenced their responses with electronic health record data to evaluate the reliability and structure of a standard ED satisfaction scale. They found structural and scoring problems with the existing scale and proposed revisions to improve how it measures overall satisfaction. Refining how patient satisfaction is measured in the ED is necessary to make comparisons across studies and track quality improvement efforts meaningfully.

Plain English
Researchers evaluated the after-visit summary — the discharge instructions document given to patients leaving the emergency department — using a three-stage process that combined human factors experts, clinicians from multiple specialties, a patient representative, and an IT assessor. The process identified 78 usability problems, with the patient representative consistently rating issues as more serious than clinical experts. A staged heuristic evaluation that incorporates patient and diverse clinical perspectives is more thorough than standard expert review and identifies which problems can actually be fixed.

Plain English
Researchers built a real-time pipeline that reads clinical notes from the hospital's electronic health record, runs them through an AI model trained to detect signs of opioid misuse, and delivers the result as an alert to the treating clinician within minutes. Silent testing before launch showed the model retained accuracy similar to published studies, and clinician interviews led to design refinements before go-live. The detailed pipeline and pseudocode are shared openly so other hospitals can replicate the implementation.

Plain English
Researchers used stable isotope labeling to track how quickly a protein called SOD1 is made and cleared in spinal fluid of people with ALS caused by SOD1 gene mutations, people with the sporadic form of ALS, and healthy controls. In people with disease-causing SOD1 mutations, the mutant protein turned over twice as fast and was present at sixteen times lower concentration than the normal version. In sporadic ALS, SOD1 dynamics were indistinguishable from controls, suggesting the protein does not play the same central role in that form of the disease.

Plain English
Researchers compared the metabolic effects of losing 10% of body weight through diet alone versus diet plus structured exercise in people with obesity and prediabetes. Those who exercised in addition to dieting had twice the improvement in whole-body insulin sensitivity and showed stronger changes in muscle genes related to energy metabolism and blood vessel growth. Adding regular exercise to a calorie-restricted diet produces substantially greater metabolic benefits than dieting alone in people with obesity and prediabetes.

Plain English
This paper describes the study protocol for testing whether an automated machine-learning tool in the emergency department that identifies older adults at high fall risk and prompts referrals to a falls clinic actually reduces falls after discharge. The trial uses a regression discontinuity design to draw causal inferences without a traditional randomized assignment, comparing patients just above and below the risk score threshold. As of writing, the tool had flagged over 1,300 patients across two sites, with 15% of referred patients completing a clinic appointment.

Plain English
Researchers interviewed 12 emergency physicians about how easy a clinical decision support tool for pulmonary embolism diagnosis was to use in daily practice, focusing on a tool that was designed using human factors principles from the start. About two-thirds of the usability barriers and facilitators they described could be tied back to decisions made during the human-centered design process, while the remaining third pointed to gaps in that process. Applying human factors methods during design substantially improves usability, but physical environment and team workflow factors remain hard to anticipate in advance.

Plain English
Researchers systematically reviewed the documentation from the design process of a clinical decision support tool for diagnosing pulmonary embolism in the emergency department, comparing it against real-world usability problems identified in practice. Seven types of design decisions emerged, including workarounds and iterative changes, and some usability problems traced back to gaps in the original design process. Human-centered design reduces but does not eliminate usability problems, and continuous redesign after deployment is needed to address what initial design processes miss.

Plain English
Among 43 women with obesity who lost about 20% of their body weight, about a quarter showed dramatic improvements in insulin sensitivity while another quarter showed almost none. Those who responded best started out with the worst insulin sensitivity and reached normal levels after weight loss, suggesting a ceiling effect — once normal function is restored, further improvement is not possible. Baseline metabolic status, not the type of intervention, determines who benefits most from weight loss.

Plain English
This study compared the effects of gastric bypass surgery and diet-only weight loss on the pancreas's ability to secrete insulin (beta-cell function) in people with and without type 2 diabetes. In people without diabetes, weight loss reduced beta-cell function regardless of how it was achieved; in people with diabetes, marked weight loss strongly improved beta-cell function regardless of whether it came from surgery or diet. Surgery offers no unique advantage over matched diet-induced weight loss for improving pancreatic function.

Plain English
Women with overweight or obesity who developed gestational diabetes (GDM) already had worse insulin sensitivity and beta-cell function at 15 weeks of pregnancy — well before standard screening occurs — compared with women who did not develop GDM. As pregnancy progressed to 35 weeks, insulin resistance worsened more and beta-cell compensation was weaker in the GDM group. Early pregnancy blood sugar measurements show some predictive value for GDM but are not yet accurate enough to replace standard screening.

Plain English
University of Wisconsin Health developed a formal governance process for reviewing and approving AI tools before and after clinical deployment, using a multidisciplinary committee that evaluates accuracy, fairness, and ongoing performance. The structure provides a clear institutional pathway from initial review through live monitoring and has enabled safer adoption of multiple AI applications. A defined governance structure with ongoing oversight is essential for health systems to deploy AI tools responsibly and maintain clinician trust.

Plain English
This case study describes the practical challenges of taking a machine-learning fall-risk prediction model built on historical data and making it run automatically in real time within a hospital's electronic health record system. The authors encountered issues with data availability, infrastructure, and governance, and share the principles they used to resolve them. Moving predictive models from research databases to live clinical environments requires solving a distinct set of operational problems that are not addressed in model development literature.

Plain English
Researchers used four years of patient data from a large hospital to examine how emergency department physician workload affects not just in-ED outcomes but the amount of follow-up care patients use after discharge. As physician busyness increased up to a point, patients discharged home generated about 5% more follow-up care events, driven by physicians ordering more tests for less severe patients when under pressure. The true cost of ED crowding is higher than previously estimated because it generates additional healthcare utilization after patients leave.

Plain English
A randomized trial enrolled 585 older adults with self-reported memory concerns and assigned them to mindfulness meditation, structured exercise, both, or a health education control group for 18 months to test whether these interventions improve memory and thinking. Neither mindfulness nor exercise, alone or combined, produced meaningful improvements in memory, executive function, brain structure, or self-reported cognition at 6 or 18 months. Mindfulness training and exercise do not improve cognitive performance in older adults who have concerns about their memory but have not been diagnosed with dementia.

Plain English
Researchers tracked insulin sensitivity and beta-cell function across pregnancy in 184 socioeconomically disadvantaged African American women with overweight or obesity. Women who gained more weight during pregnancy than guidelines recommend showed a greater decline in insulin sensitivity by the third trimester, though their pancreatic function was not affected. Excess gestational weight gain has measurable negative effects on glucose metabolism even in the absence of gestational diabetes.