The AI market is full of big promises of grand transformation. Health care is a prime target for those promises, beset as it is by financial pressures, labor shortages, and the growing burden of caring for an aging population. AI developers are targeting functions that vary widely, from curing cancer and performing surgery to streamlining routine administrative tasks.
The opportunity is genuine, but execution can be difficult. Numerous software vendors have tried to “fix” health care challenges but failed because they misunderstood the environment. “Health care is very complex,” says Steve Bethke, vice president of the solution developer market for Mayo Clinic Platform, which supports the buildout and deployment of digital solutions for health care companies through data-based insights and expert validation. “Solution developers must have a deep focus on clinical and technical capabilities, and then align their solutions to the relevant business impacts. If they miss any dimension, the solution will not be adopted or drive value.”
AI applications for health care are proliferating rapidly. The U.S. Food and Drug Administration has approved more than 1,300 AI-enabled medical devices, mostly for interpreting diagnostic images. More than half of these were approved in the past three years, with the earliest dating as far back as 1995. Non-radiological applications carry out tasks as diverse as tracking sleep apnea, analyzing heart rhythms, and planning orthopedic surgeries.
AI applications that do not count as medical devices— for example, those that handle scheduling and administrative tasks—are more difficult to track but are also rapidly increasing. AI can help coordinate complex tasks and workflows that are often conventionally managed by whiteboards and sticky notes. Such functions may well outstrip clinical uses in their impact on health systems. A recent survey of technology leaders found that 72% said their top priority for AI was reducing caregiver burden and improving caregiver satisfaction, while over half (53%) cited workflow efficiency and productivity.
Any health care-related application can potentially impact patient care, whether directly or indirectly, and AI apps that are poorly designed or inadequately trained and validated can put patients at risk. Providers recognize that risk: In the same survey, 77% said immature AI tools are a significant barrier to adoption. Regulators and lawmakers are also keeping an eye on the risks as development and adoption burgeon, though the U.S. regulatory picture is still in flux, as a 2024 report to Congress on AI in health care observes.
To tackle some of the technical challenges, many health care providers are partnering with application developers to build AI solutions. In a recent study, McKinsey found that 61% of health care organizations intend to pursue partnerships with third-party vendors to develop customized generative AI solutions as a primary strategy as opposed to building them in-house or buying off-the-shelf products.
But health care-specific AI applications must also be tailored to the nuanced clinical needs of medical providers as well as the complex business and regulatory considerations of the wider sector. This is where developers can benefit from working with a partner with a deep understanding of the health care environment to tailor applications to what providers want and need most. Doing so helps to position AI products for maximum impact and value, avoiding the pitfalls unique to the health care environment.
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
Ajay Gannerkote, president of Integrated DNA Technologies (IDT), says what’s most exciting about CRISPR is its potential to shift medicine from managing disease to directly correcting its root cause. “For patients with severe genetic conditions, especially those with no existing treatment options, that represents a fundamental change in what’s possible,” he said.
IDT played a pivotal role in manufacturing the personalized gene editing therapy given to baby KJ Muldoon to treat his rare metabolic disorder. Today, KJ is free from the toxic ammonia buildup that drives a 50% mortality rate for his condition in infancy. While his story highlights the life-changing potential of gene editing, the field now wrestles with the next challenge: expanding these therapies to benefit broader patient populations.
In contrast to KJ’s urea cycle disorder, which stemmed from a single disease-causing mutation that could be precisely targeted, many genetic disorders arise from numerous mutations scattered across a gene where individualized corrections are too resource-intensive to scale.
Gannerkote says turning powerful gene editing tools into broadly accessible clinical therapies requires progress across multiple fronts. Many CRISPR therapies are still bespoke, with manufacturing processes that are not yet standardized or easily repeatable, leading to long timelines and high costs. In regulation, therapy developers and government regulators face a learning curve when evaluating new modalities, particularly when speed is critical for patients with life-threatening conditions.
Today’s gene editing companies reflect on what’s required to scale personalized CRISPR therapies for maximized impact in the clinic.
End-to-end
Sadik Kassim, PhD, CTO of Genomic Medicines at Danaher, explains that personalized therapies do not naturally lend themselves to traditional drug-development models. Gene editing companies are now seeking “platformization,” where common manufacturing processes are standardized, and limited elements, such as guide RNAs, are customized for each patient to reduce costs and speed timelines.
“Baby KJ’s treatment succeeded because multiple elements aligned simultaneously,” explained Kassim. The foundational science, which achieved successful gene corrections in animal models of phenylketonuria (PKU), an inherited metabolic disorder caused by mutations in the PAH gene that impair the enzyme responsible for breaking down phenylalanine, had already been developed in the academic labs led by Children’s Hospital of Philadelphia (CHOP) physician scientists, Rebecca Ahrens-Nicklas, MD, PhD, and Kiran Musunuru, MD, PhD. Teams were then able to move quickly when the clinical need became clear.
Regulatory engagement was also critical. Danaher teams worked directly with the FDA to streamline the treatment approval process without compromising patient safety. That collaboration compressed a timeline that would normally take 18–24 months down to roughly six months.
“Replicating this for future patients will require moving away from one‑off efforts and toward repeatable platforms with established processes, validated assays, and clearer regulatory precedents, so that speed becomes the norm rather than the exception,” Kassim said.
Amy Pooler, PhD, CSO of ElevateBio, agrees that the transition steps between therapy design and manufacturing are often where the greatest delays occur. ElevateBio seeks to address this bottleneck by building an end-to-end genetic medicine platform.
“A critical driver for the company is making sure we have a clear line of sight into manufacturing from the very beginning,” Pooler said. “One reason Baby KJ’s case was successful is that Danaher managed the handoffs smoothly.”
Pooler also describes developing genetic medicines as “building the plane while you’re flying it.” The field still lacks enough data to reliably predict patient outcomes. Every clinical trial readout provides a valuable lesson for the field.
“I’m excited about the clinical evidence that’s starting to accumulate, showing gene editing can be transformative for patients, which we didn’t have five to ten years ago,” she said.
Large gene, generalizable therapy
ElevateBio’s expanding CRISPR toolbox includes base, prime, and epigenetic editing. Notably, the Durham-based company’s AI platform generates novel recombinases for targeted gene insertion, an approach that holds promise as a generalizable medicine that could treat patients regardless of their underlying disease-causing mutation.
Using AI-guided design, ElevateBio explores entirely new regions of protein space to discover potent and highly specific recombinases that expand the range of diseases amenable to gene editing. These engineered enzymes, which possess 50% or less homology to known proteins, can access novel genomic regions that remain difficult to target with existing CRISPR technologies.
Ben Kleinstiver, PhD, associate investigator at Massachusetts General Hospital (MGH) and co-author of the NEJM study describing KJ’s case, says the FDA’s Plausible Mechanisms Pathway has helped address some of the regulatory challenges to streamline the path to the clinic. Yet, there remains a major motivation for pan-mutation approaches that are more widely applied across patients.
Kleinstiver’s research group, in collaboration with Full Circles Therapeutics, recently developed a circular single-stranded DNA donor (ssDNA) that enables safer kilobase-scale integration for human cells.1 The technology provides an alternative to double-stranded DNA (dsDNA) donors that evoke harmful immune responses yet are required for recognition by the diverse suite of genome editing enzymes. Notably, the new circular donor maintains recombinase compatibility by attaching a short region of dsDNA that can go undetected by the cytosolic DNA sensor and immune system activator, cGAS.
Patients now
While the gene editing field often concentrates on large indications driven by a single common mutation, Edward Kaye, MD, CEO and director of Aurora Therapeutics, aims to extend these technologies beyond the “lucky few” who share the same mutation.
Aurora’s leadership team, from left: Morgan Maeder, PhD, Edward Kaye, MD, and David Litvak, MBA [Aurora]
Co-founded by Jennifer Doudna, PhD, CRISPR Nobel Laureate, and Fyodor Urnov, PhD, scientific director of the Innovative Genomics Institute, Aurora launched in January to build a sustainable pipeline to scale rare disease treatments. Traditionally, developing therapies for these ultra-rare or N-of-1 conditions can require several million dollars for a single patient.
Aurora is pursuing an “umbrella IND” strategy that allows multiple guide RNAs to be evaluated within a single clinical trial. The company’s initial focus is on PKU.
PKU offers several advantages for early clinical development. Patients are routinely identified through newborn screening programs shortly after birth, which facilitates trial participant identification and enrollment. The condition also benefits from a clear regulatory precedent: reductions in phenylalanine levels are an established clinical endpoint used to move therapies toward approval.
“What we learn from PKU will be used for many other diseases because we have the systems in place,” said Kaye. “It expands gene editing into many more patients, by going after one disease first.”
Kaye also stresses the importance of engaging patient communities, whose input can ensure studies and regulatory processes are not overly burdensome for patients and families.
Maher Masoud, CEO of MaxCyte, emphasizes putting patients at the forefront. He adds that most gene-editing therapies in the clinic require significant patient conditioning, which can lead to lengthy treatment cycles and clinical trial timelines. Yet he sees these barriers to scale being eroded over the near term. As an example, modalities, such as allogeneic cell therapies, require far less patient conditioning and easier dosing regimens to support cheaper therapies.
In 2013, MaxCyte partnered with CRISPR Therapeutics on early work that led to the first FDA-approved therapy based on CRISPR-Cas9, Casgevy, with MaxCyte’s ExPERT electroporation platform enabling the efficient delivery of gene editing machinery into cells.
More than a decade later, the company has developed more than 1,000 applications and protocols. The broad engineering platform can repeatedly engineer batches of at least 20 billion cells using CRISPR-Cas9 in addition to base and prime editing.
Masoud says low-significant gene editing commercial success has been a bottleneck to scaling personalized therapies. Yet, he reiterates that CRISPR and other gene editing technologies were discovered a short 12 years ago.
“With CRISPR, we are finally seeing cures, Casgevy, LYFGENIA, and baby KJ are proof of that,” he says. “This is just the beginning.”
References
Tou, C.J., Xie, K., Ferreira da Silva, J., et al. Invasive DNA donors and recombinases license kilobase-scale writing. Nature. 2026. DOI: 10.1038/s41586-026-10241-z.
Interventions: Behavioral: AGENCIA Digital Self-Guided; Behavioral: AGENCIA In-person With Digital Assistant
Sponsors: Fundación Pública Andaluza para la gestión de la Investigación en Sevilla; Hospital Universitario Virgen del Rocio; Instituto de Salud Carlos III
An artificial intelligence (AI) model developed at the Mayo Clinic can detect very early signs of pancreatic cancer from CT scans of the abdomen that are normally invisible to the human eye.
Researchers tested the Radiomics-based Early Detection MODel (REDMOD) and found it was able to identify 73% of very early pancreatic ductal adenocarcinoma. In contrast, only 39% of these cases were identified by radiologists.
Around 67,530 Americans are expected to be diagnosed with pancreatic cancer in 2026. It has a very poor prognosis, around 13% survival at five years, largely because it is often diagnosed at a late stage.
“Early detection of pancreatic ductal adenocarcinoma is the most powerful approach to improve survival,” write lead author Ajit Harishkumar Goenka, MD, a researcher and clinician at the Mayo clinic, and colleagues in the journal Gut.
“However, this objective is fundamentally impeded by the morphologically normal appearance of the pancreas on conventional imaging during its curable pre-clinical phase.”
To try and tackle the issue of early diagnosis, Goenka and team developed REDMOD to analyze very early signs of pancreatic cancer on computed tomography (CT) scans that are normally very hard to detect with the human eye.
As part of the study the researchers evaluated 219 routine CT scans of the abdomen from patients who later developed pancreatic cancer. They also included 1243 matched controls scans of people who had not developed cancer.
After the AI model was developed and trained the researchers tested it in 63 pre-diagnostic cases and 430 controls. These scans were taken around 16 months before the cancer cases were diagnosed.
REDMOD correctly identified 73% people who would later develop pancreatic cancer, while radiologists looking at the same scans only picked up about 39% of cases. The ability of the model to identify people without the disease was also good and it correctly identified 88% of people who did not have pancreatic cancer. When patients had repeat scans, the AI’s risk score was very consistent, agreeing 90–92% of the time.
“The demonstrated ability of the framework to consistently detect these occult signals on a large clinically oriented dataset, combined with its high longitudinal stability and validated specificity, establishes a robust foundation for AI-augmented early detection,” write Goenka and colleagues.
“While prospective validation is paramount to confirm clinical utility, the REDMOD framework represents a significant advance towards shifting the paradigm for sporadic pancreatic ductal adenocarcinoma from a late-stage symptomatic diagnosis to proactive pre-clinical interception, offering tangible hope for improving outcomes in this challenging disease.”
The four health system CEOs summoned before a Congressional committee Tuesday likely breathed sighs of relief early in the hearing, when it became clear they had friends in the audience.
Instead, committee members largely blamed the other party’s health care policies for driving U.S. health care prices to levels inaccessible to many Americans.
The hearing was part of the House Ways and Means Committee’s effort to understand the root causes of rising health care costs in the U.S. It comes three months after the committee heard from the CEOs of the country’s largest health insurers, who largely deflected blame onto hospitals and drugmakers.
In attendance were the CEOs of some of the country’s largest health systems: HCA Healthcare, a for-profit system of 190 hospitals, and CommonSpirit Health, a nonprofit system of 158 hospitals. The CEOs of New York-Presbyterian and North Carolina’s ECU Health were also there.
BackgroundAlthough generative artificial intelligence offers substantial potential benefits in healthcare, negative attitudes and elevated anxiety among nurses may hinder its effective integration into clinical practice. Evidence regarding the psychological impact of generative artificial intelligence on nurses remains limited.ObjectiveThis study examined the relationships among sociodemographic characteristics, anxiety, and attitudes toward generative artificial intelligence among nurses.MethodsA cross-sectional correlational design was employed. Data were collected from 312 hospital nurses using online questionnaires assessing sociodemographic characteristics, attitudes toward artificial intelligence, and artificial intelligence-related anxiety. Data were analyzed using IBM Statistical Package for the Social Sciences (SPSS) Statistics software version 28.ResultsHigher levels of artificial intelligence-related anxiety were associated with less favorable attitudes toward artificial intelligence. Sociodemographic characteristics and anxiety scores collectively explained 49.4% of the total variance in attitudes toward artificial intelligence. Gender, experience with artificial intelligence, use of artificial intelligence in nursing care, awareness of artificial intelligence applications in healthcare, hours spent on the internet, age, and professional experience accounted for 24.7% of the variance in negative attitudes toward generative artificial intelligence.ConclusionAnxiety and experiential factors play a central role in shaping nurses’ attitudes toward generative artificial intelligence. Increasing nurses’ exposure to and awareness of artificial intelligence in nursing practice may reduce anxiety and support its acceptance and appropriate use.
