Weekly Rundown: Doximity launches DoseCheck; AI recommends costlier care than doctors, study finds

Doximity launches safety feature for AI-based medication dosing

Doximity launched DoseCheck, a new safety feature within Doximity Ask designed to catch unsafe or incorrect medication dosing recommendations generated by artificial intelligence. 

Doximity Ask is the company's clinical decision support tool.

Rather than relying on an LLM to double-check another LLM, DoseCheck adds a deterministic, rule-based verification layer, checking AI-generated dosing claims against structured information including FDA labeling and professional guidelines, the company said. By checking against FDA labeling and professional guidelines, DoseCheck increases accuracy to more than 99%, the company claims.

Doximity tested nine widely used AI models with more than 8,000 dosing questions based on real physician queries. At least 2.3% of dosing claims made by these models were wrong, and 43.8% of detected errors were serious or potentially lethal.

According to the analysis, common sources of errors include dosing newer drugs, failing to account for patient-specific factors such as kidney function, age, weight and medical history, and mistakes involving units, decimal points or drugs with similar names.

"AI is probabilistic. Drug dosing can’t be," said Louis-Antoine Mullie, M.D., chief AI researcher at Doximity. "DoseCheck doesn’t ask a language model to grade its own homework. It checks dosing claims against a structured, authoritative database, so physicians get the same verified dosing answer every time."

DoseCheck is free and has been rolled out to all Doximity Ask users across mobile and desktop.

Hims & Hers preps for peptides launch

Health and wellness platform Hims & Hers plans to launch a peptides specialty after the federal government signaled it will allow compounding pharmacies to manufacture six peptides temporarily.

The U.S. Department of Health and Human Services will not take action against licensed pharmacists who dispense certain unapproved peptides, according to a new policy published this week in the Federal Register.

Stopping short of a formal declaration allowing their use, HHS Secretary Robert F. Kennedy Jr. has cleared the way for the compounding of six popular wellness peptides, Fierce Pharma reported. Under Kennedy’s "enforcement discretion policy" issued Wednesday, HHS will not take enforcement action against pharmacies, physicians and federal facilities compounding six specific peptides, assuming those parties meet certain specific qualifications. 

The specific compounds included (PDF) in the list are BPC-157, epitalon, KPV, MOTS-c, semax and TB-500, all of which were discussed in a meeting of the FDA’s Pharmacy Compounding Advisory Committee (PCAC) in July, with the panel ultimately voting to recommend including the treatments onto the 503A Bulk Drug Substances list. 

Through its new Peptides+ specialty, Hims & Hers plans to offer access to Sermorelin, Glutathione, B12, NAD+, and L-carnitine, with BPC-157 and TB-500 to follow when allowable. "If a customer is eligible, a personalized treatment plan will be prescribed and delivered by independent providers, treatments supplied by licensed pharmacies, with ongoing clinical support our customers can trust," CEO Andrew Dudum wrote in a blog post.

"We’re focused on entering this specialty the right way, not the fastest. Most of what we know about these compounds is still early — with limited published studies, we draw primarily from clinical experience from the last decade. We’ve consistently acknowledged that while these compounds hold promise, there is still information we do not yet know. Provider-supervised use is how we plan to build on this foundation and make sure that access and innovation can move at the same pace," Dudum said.

Google rolls out educational resource on healthcare AI

The tech giant launched Google AI for Healthcare Professionals Certificate, a new program designed to teach healthcare professionals responsible, practical ways to use AI in their day-to-day work. Clinicians at Google developed the virtual course, and the program is accredited for up to five American Medical Association (AMA) continuing education credits.

NYU Grossman School of Medicine, the University of Alabama at Birmingham and Touro University are offering the program to their faculty, residents and medical students. Leading nursing institutions, including Purdue University School of Nursing, are offering the certificate as part of their healthcare professional training, according to Google.

"As I’ve traveled around the world this year, I’ve heard loud and clear from healthcare professionals how much need there is for a thoughtful but practical educational resource on AI, made specifically for healthcare," Michael Howell, M.D., Google’s chief health officer, wrote in a blog post.

The program offers flexible, self-paced lessons that focus on non-clinical administrative tasks. It is designed to be completed in less than five hours and to keep patients’ safety and privacy front and center, according to Google. Participants will learn how to integrate AI into their daily routine, including optimizing daily administrative tasks, strengthening patient communication and digesting complex clinical studies.

AI tools recommend more tests, referrals than doctors, study finds

Large language models recommend more up-front testing, referrals and treatment than the physicians who actually saw the patients and at more than three times the cost, according to a study published in NEJM AI.

Researchers from the Technion, Stanford University and collaborating institutions tested 20 general-purpose and four specialized medical AI systems on records from 200 real outpatient and emergency visits. Each received the clinical presentation without the physician’s assessment and plan. The team priced the recommended tests and referrals in both the AI and physicians’ plans using the same 2026 Medicare rates.

Across the 20 general-purpose systems, AI-suggested care cost a mean of $314 per visit, compared with $97 for care recommended by treating physicians. Suggested care was more costly in 180 of 200 cases. The average cost of care per AI system ranged from $209 to $599.

Diagnostic testing rose from $71 per visit in the physicians’ plans to $176 across the systems, about $105 more per visit. The referral rates for the physician’s plans were 0.15 per visit compared to 0.87 for the LLM’s plans, which at the corresponding Medicare rates is $111 more per visit, according to the study.

"The four specialized medical systems showed comparable excess (2.7× to 3.5×), so medical fine-tuning did not curb — and may reinforce — cost escalation," the study authors wrote.

The researchers used MTSamples, a publicly available collection of 4,999 real de-identified medical transcription sample reports contributed by practicing clinicians as reference material for medical transcriptionists. The notes are documentation from real patient visits, but they are a convenience sample assembled for transcription training rather than a representative draw from U.S. ambulatory practice, the study authors wrote. 

"Under a single standard prompt, every AI system we evaluated, general purpose and medically specialized alike, suggested more diagnostic tests, medications, and specialist referrals than the treating physician on the same cases," the study authors wrote.

LLMs appear optimized to be comprehensive and avoid missing diagnoses, and may show an analogous tendency toward exhaustive workup, the study authors noted.

The researchers argued that much of the excess can be reduced through better prompting, suggesting cost inflation is partly a deployment and design problem rather than an unavoidable feature of AI.

Researchers reevaluated 19 models using different instructions. A simple cost-aware prompt telling models to recommend only clinically necessary, cost-effective items reduced average recommended costs from 3.2 times physician costs to 2.5 times.

A stronger "parsimonious-but-safe" prompt told models to focus on the three most likely diagnoses, explicitly rule out dangerous "cannot-miss" conditions, and recommend only tests whose results would change management. That cut costs to just 1.4 times physician costs. But diagnostic agreement with physicians fell in 17 of the 19 systems, illustrating why cost reductions need to be evaluated alongside other changes in the recommendations, the study authors noted.

"Getting the diagnosis right matters, but so does the path you take to get there," said Prof. Dvir Aran of the Technion, the study’s corresponding author, in a statement. "Accuracy and safety have dominated how clinical AI is trained and evaluated. Efficiency needs the same attention: whether each test or referral is likely to help the patient."

Eko Health offers new stethoscope that blends analog listening with AI

Digital health company Eko Health launched a new stethoscope, the CORE 300, that integrates traditional and digital listening modes directly into a single chestpiece.

The AI-enabled stethoscope, which received FDA clearance, is also in line with premium traditional stethoscopes, making the digital technology more accessible to a broader group of healthcare professionals, the company said.

With a twist of the chestpiece, clinicians can switch between familiar traditional listening and Eko’s advanced digital listening technology, with up to 35x amplification, active noise cancellation and real-time heart rate displayed directly on the device. When connected to the Eko App and Eko's FDA-cleared AI for structural murmur detection, CORE 300 can also bring AI disease detection into the physical exam.

"Millions of healthcare professionals around the world use a stethoscope to examine patients every day," said Connor Landgraf, co-founder and CEO of Eko Health, in a statement. "Our goal is to bring the most ubiquitous clinical exam into the precise digital era. Once the physical exam becomes digital, we combine better listening with objective measurements and AI-powered clinical decision support right at the point of care."

Clinical technologies are evolving quickly, but adoption depends on making new capabilities fit naturally into the way clinicians already practice, said Steven R. Steinhubl, M.D., chief medical officer of Eko Health and a board-certified cardiologist. 

"CORE 300 bridges that gap. It preserves the traditional experience of auscultation while putting enhanced digital listening, objective measurements, and access to clinical decision support a twist of the chestpiece away. That combination can help make a smarter, more connected physical exam a routine part of care rather than the exception," Steinhubl said.

With Eko's digital stethoscope, clinicians can record, visualize and save heart, lung and body sounds and interpret the data with Eko’s FDA-cleared AI and software. According to the company, the device could expand screening capabilities in primary care offices, busy clinics and rural healthcare settings.

Eko Health launched its first Eko CORE digital stethoscope attachment in 2025, and, to date, has delivered more than 750,000 digital stethoscopes and AI-enabled devices worldwide.

RevSpring rolls out conversational AI agent for healthcare payments 

Healthcare engagement and payment solutions provider RevSpring debuted a conversational intelligent virtual agent that helps patients manage payment plans over the phone. 

Let’s Talk expands on RevSpring’s voice platform and is integrated with its payment portal. It gives patients the ability to make payments, create payment plans, ask questions and respond to outbound voice reminders. 

Initial results of the tool show a 36% increase in patient payment rates and a 27% decrease in transfer to live agents, the company says. Moreover, average call durations decreased 25% with Let’s Talk. 

“Healthcare organizations are looking for ways to automate more patient interactions and improve financial outcomes without losing sight of the need for patient empathy,” said Scott MacKenzie, RevSpring CEO, in a statement. “Adding payment plans to Let’s Talk expands what patients can accomplish through conversational self-service, helping more callers resolve their needs without transferring to a representative. We’re building on years of proven payment automation to help providers improve efficiency and scale self-service across the patient financial experience.” 

Symplr launches AI healthcare staff scheduling tool 

Enterprise healthcare operations software provider symplr launched Monday an integrated workforce management solution combining AI-driven scheduling and timekeeping. 

The tool—dubbed symplr Time and Scheduling—combines symplr’s predictive nurse and staffing scheduling tool with its timekeeping solution. Nurse managers can identify and address staffing gaps "hundreds of days in advance," ultimately reducing time spent on staffing and timecard tasks by approximately 50%, executives say. 

"Nurse leaders shouldn’t have to spend hours building schedules and scrambling to fill open shifts when they could be supporting their teams and patients," said Ali Morin, symplr chief nursing informatics officer, in a statement. "Nurses want more flexibility and predictability, while health systems need the right people in the right place at the right time. We designed symplr Time and Scheduling to deliver both, giving nurses and nurse leaders more time for patient care and less time managing the work around it." 

Additional AI capabilities are expected to roll out in 2027, the company says. 

Cleveland Clinic Cancer Institute establishes AI task force 

Cleveland Clinic Cancer Institute launched an AI task force Monday in a bid to evaluate the technologies and streamline adoption in a "safe and deliberate manner."

The task force is dubbed the Cancer Institute Office of Technology and Digital Strategy. It is led by Paul Thambi, M.D., and Louis Williams, M.D. AI is currently being used by the institute to personalize radiation doses, reduce notetaking time and streamline patient referral processes, according to the Oct. 27 announcement.  

Williams said in a statement it’s "crucial to have a system to continuously re-evaluate the effectiveness of each tool," adding that outputs of models should be analyzed on "an ongoing basis."

"There can be drift in a model that may influence its accuracy over time or there may be bias in the model that isn’t apparent upfront," Williams said. "That’s why it’s necessary to have a unified strategy that includes training models to identify mistakes before a model goes into production as well as longitudinal risk assessment to ensure safety."