Source: Fast Company, Jesus Diaz
Photo: FC

People can now trick insurance with AI-generated X-rays. It’s just the latest example of how people can scam everyone from food delivery services to the IRS.

It’s so easy to cheat now. Using generative AI, anyone can get a free meal or product. They can even get free money by scamming the government itself. And, like radiologists have just discovered, they can even cheat doctors and insurance companies by using AI-generated X-rays.

According to a new study published by the Radiological Society of North America, most experts can’t distinguish fake fractures from the real thing now. Undetectable insurance fraud is one click away. It’s just the last of a growing list of low-hanging fruit, zero-cost scams made possible with the power of AI. And it’s only going to get worse.

Fake x-rays

The Radiological Society of North America’s study subjected 17 global medical specialists from six different countries, some boasting up to 40 years of field experience, to a visual test involving 264 X-rays—half authentic and half synthetic creations created by AI tools like ChatGPT and Stanford’s open-source RoentGen model. When left entirely in the dark about the presence of these artificial images, the physicians only managed to correctly identify the synthetic X-rays 41% of the time.

Even after receiving explicit warnings that fakes were hidden in the batch, their average success rate limped up to 75%, ranging between a dismal 58% and a respectable but imperfect 92%. A doctor’s decades of hands-on experience offered little statistical advantage in catching the deception, the study says, though musculoskeletal experts performed marginally better than their peers.

To make matters worse, the large language models responsible for birthing this digital chicanery—including GPT-4o, GPT-5, Gemini 2.5 Pro, and Meta’s Llama 4 Maverick—fared no better as automated detectives, scoring accuracy rates between 57% and 85%.

“Our study demonstrates that these deepfake X-rays are realistic enough to deceive radiologists, the most highly trained medical image specialists, even when they were aware that AI-generated images were present,” noted lead author Dr. Mickael Tordjman. “This creates a high-stakes vulnerability for fraudulent litigation if, for example, a fabricated fracture could be indistinguishable from a real one. There is also a significant cybersecurity risk if hackers were to gain access to a hospital’s network and inject synthetic images to manipulate patient diagnoses or cause widespread clinical chaos by undermining the fundamental reliability of the digital medical record.”

According to Tordjman, AI-generated medical images often look too perfect, with bones that are “overly smooth, spines unnaturally straight, lungs overly symmetrical, blood vessel patterns excessively uniform, and fractures that appear unusually clean and consistent, often limited to one side of the bone.” But that’s just yesterday’s crop of tools. Like AI-generated video, AI will make these X-rays absolutely perfect and undetectable soon. It’s the nature of the ever evolving AI beast.

To fight this, experts are demanding invisible watermarks and cryptographic signatures directly linked to the technician capturing the scan, effectively acting like a mathematical seal of authenticity that proves a human body was actually in the room.

Shallowfakes and raw deals

Fraudulent x-rays are a serious example of the more quotidian truth-bending that’s already happening. Take the rise of shallowfakes, which are surface-level digital illusions that require minimal effort to produce maximum financial deceit.

Ordinary consumers are using generative AI to visually alter their food deliveries into unappetizing disasters. It takes one click to digitally manipulate the interior of a hamburger or a piece of chicken so it appears raw, tricking algorithms and customer service reps into approving instantaneous refunds.

“The trend is real and growing,” observed generative AI fraud specialist Alberto Palomar to Spanish newspaper El Confidencial. “AI is putting it within the reach of anyone who has no idea about technology to take this trickery to all levels.”

While Uber Eats passes these fraudulent financial hits directly onto the unsuspecting restaurants, DoorDash maintains a strict corporate line, warning users that “trying to game the system with a fake image might seem clever at the moment, but it’s not worth a permanent ban over a $20 order.”

The human collateral damage in this digital swindle lands squarely on the gig workers delivering the food. When a customer successfully fakes a damaged or undercooked meal, the driver is penalized with bad ratings or permanent deactivation, says Ligia Guallpa, executive director of the Worker’s Justice Project. “The biggest complaint that deliveristas have is how the apps are aggressively punishing them for things that are out of their control,” she points out. Her organization was tracking roughly 1,500 active deactivation cases.

But it’s more complicated than that. Drivers are also weaponizing the technology to fake deliveries they simply steal. Austin-based DoorDash customer Byrne Hobart watched his Dasher accept the order, instantly mark it as complete, and upload an AI-generated porch photograph with the driver there. The company refunded his poke bowl and noted, “After quickly investigating this incident, our team permanently removed the Dasher’s account and ensured the customer was made whole.”

The million dollar paper trail

Meanwhile, the epidemic of micro-fraud is morphing into a macroeconomic catastrophe for the global insurance sector, mutating minor vehicular scrapes and broken smartphones into massive corporate liabilities.

In the United States, “20-30% of insurance claims may now include altered images, fabricated documents, or synthetic medical reports”, claims Shift Technology, a technology company that provides AI agents to automate claims.

In the UK, insurance company Allianz reported a 300% spike in the use of AI to alter documents, photos, and videos in customers claims from 2022 to 2023. It will only get worse, says global insurance data analytics company Verisk: “One in three consumers would consider digitally altering an insurance claim image or document to strengthen their case—and that number rises to 55% of Generation Z.”

In Spain, insurer AXA says it processes up to 30,000 claim-related documents a day, making it harder to spot synthetic tampering at scale. Arturo López-Linares, Claims Director at AXA Spain, outlined the terrifying breadth of the efforts. “It is an alarming trend. Documents have always been falsified, all our lives. The problem now is the ease with which you can do it and that these tools are within everyone’s reach,” he warned. “You can ask the AI to put a scratch on your car or modify a repairman’s invoice. It is impossible to catch it with the naked eye, so you also need to use technology to identify it.”

While acknowledging the digital cheating pool sits at just over 2% of the population, the math is unforgiving, says López-Linares: “We have gone from identifying only 3% of fraud cases with digital methods a few years ago to 30% currently… but it already accounts for millions of euros, and AI is playing a fundamental role.”

The problem with all this is that it is impossible to catch. Sure, you can analyze a digital photograph’s metadata—the hidden strings of code functioning like algorithmic fingerprints that log geolocation, device specs, and timestamp data—but since metadata can be effortlessly spoofed, that barrier is gone.

Some say the ultimate defense relies on advanced image analysis software, but as the X-ray study has demonstrated, that’s also hard and will soon be impossible. Future generations of AI will trump any forensic countermeasures we develop. Plus, the cost of these measures, which will be expensive to implement and run in server farms, effectively prices small and medium enterprises out of their own survival.

“This is what happens to many companies, processing returns or investing to catch fraud costs them more than assuming the cost of it,” Palomar concluded.

Perhaps now that AI is starting to dent the economy, the corporations and governments’ bottom line, the pressure will be high enough to push for mandatory truth-certification solutions that will benefit all of us.

The early-rate deadline for Fast Company’s Brands That Matter Awards is this Friday, April 10, at 11:59 p.m. PT. Apply today.

ABOUT THE AUTHOR

Jesus Diaz is a screenwriter and producer whose latest work includes the mini-documentary series Control Z: The Future to Undo, the futurist daily Novaceno, and the book The Secrets of Lego House.

https://www.fastcompany.com/91516743/ai-fraud-x-ray

Source: Politico, Yasmin Khorram and Ruth Reader
Photo: Courtesy of Doctronic

Pilot program will test how far patients and regulators are willing to trust AI in medicine.

In a first for the U.S., Utah is letting artificial intelligence — not a doctor — renew certain medical prescriptions. No human involved.

The state has launched a pilot program with health-tech startup Doctronic that allows an AI system to handle routine prescription renewals for patients with chronic conditions. The initiative, which kicked off quietly last month, is a high-stakes test of whether AI can safely take on one of health care’s most sensitive tasks and how far that could spread beyond one AI-friendly red state.

It also serves as an early check on how far policymakers and patients are willing to trust AI over trained doctors in decision-making. By inserting algorithms into one of medicine’s most fundamental relationships, Utah’s initiative could represent the first step in upending how care is delivered in the U.S.

That raises new questions about the safety of automating prescription refills, including how they should be regulated. So far, the Food and Drug Administration has not weighed in on Doctronic’s program. If the agency determines it has authority to regulate this use of AI, it could complicate or slow its expansion.

State officials and industry backers say relying more on artificial intelligence lowers costs, reduces medication lapses and improves access to care — while generating data that could shape AI policy beyond Utah.

Health care expenses keep climbing and clinicians — especially in rural areas — are stretched thin, said Margaret Busse, executive director of the Utah Department of Commerce. The state sees automating routine prescription renewals as a way to ease pressure on providers while lowering costs for patients, she said.

It’s also a way to “provide a pathway to innovation for entrepreneurs who are using AI in creative ways that may be bumping up against regulation,” she said.

But doctors’ groups warn that delegating some aspects of prescribing medication to AI could present new hazards.

In a statement, Dr. John Whyte, CEO and executive vice president at the American Medical Association, said: “While AI has limitless opportunity to transform medicine for the better, without physician input it also poses serious risks to patients and physicians alike.”

One concern is misuse or abuse, including the possibility that people struggling with addiction could try to game automated systems to obtain drugs inappropriately. Another concern is missing subtle clinical red flags or drug interactions that a doctor would catch.

“The company has to do that kind of trust building with their patients,” said Busse. “We want it to be done in such a way that people will trust that Utah is looking at this carefully and is not being cavalier about how we granted this regulatory mitigation. In a way it’s a risk for us as we do this.”

Al Carter, CEO and executive director at the National Association of Boards of Pharmacy, said pharmacists already use AI in their prescription fulfillment process and for patient consultations.

However, he cautioned, “the one challenge from a board of pharmacy standpoint is how do you regulate all this technology, and is this technology good for health care?”

In data shared with Utah regulators, Doctronic compared its AI system with human clinicians across 500 urgent care cases. The results showed the AI’s treatment plan matched the physicians’ 99.2 percent of the time, according to the company.

“The AI is actually better than doctors at doing this,” said Dr. Adam Oskowitz, Doctronic co-founder and an associate professor of surgery at the University of California San Francisco. “When you go see a doctor, it’s not going to do all the checks that the AI is doing.”

Oskowitz said the AI is designed to err on the side of safety, automatically escalating cases to a physician if there’s any uncertainty. Human doctors will also review the first 250 prescriptions issued in each medication class to validate the AI’s performance. Once that threshold is met, subsequent renewals in that class will be handled autonomously.

The company has also secured a one-of-a-kind malpractice insurance policy covering an AI system, which means the system is insured and held to the same level of responsibility as a doctor would be.

“In medicine, there’s always going to be potential issues that patients have,” said Oskowitz. “Whether it’s caused by the AI or not — we will take the risk. I think this is going to be infinitely safer than a human doctor.”

Doctronic also runs a nationwide telehealth practice that directs patients to doctors after an AI consultation.

In Utah, patients who use the system will visit a webpage that verifies they are physically in the state. Then the system will pull the patient’s prescription history and offer a list of medications eligible for renewal.

The AI walks the patient through the same clinical questions a physician would ask to determine whether a refill is appropriate. If the system clears the renewal, the prescription is sent directly to a pharmacy.

The program is limited to 190 commonly prescribed medications. Some medications — including pain management and ADHD drugs as well as injectables — are excluded for safety reasons.

Matt Pavelle, the company’s co-founder and co-CEO, said the system would give people easier access to their medications. “It’s hard to get a renewal — if you have a chronic condition and you can’t get your medication, terrible things happen.”

The company will charge $4 per prescription renewal, a price it says is temporary. Pavelle said the cost could drop quickly as the system scales up with renewals ultimately covered by insurance or bundled into a low annual fee.

Pavelle and Oskowitz are in discussion with other states such as Texas, Arizona and Missouri. They’re also weighing a national approval pathway, rather than navigating a patchwork of state-by-state rules.

Getting there will depend on the FDA.

Usually, states are responsible for writing their own rules governing how medicine is practiced. Since Doctronic’s AI is designed to renew patient prescriptions, it’s essentially practicing medicine and could fall under state regulations, according to Lowell Schiller, a former chief counsel for the FDA.

On the other hand, the agency has said it believes it has the authority to regulate AI as a medical device if it is used to diagnose, treat, or prevent disease.

Schiller says the FDA could hold off on taking action against Doctronic, citing medical marijuana as an example of where the agency has deferred to state laws instead of enforcing federal regulation.

But if the FDA does find that Doctronic’s AI is being marketed without appropriate authorization, it may try to bring the technology into compliance. Under President Donald Trump, the agency has shown an appetite for oversight. Earlier this year, the FDA sent a letter to health wearable company Whoop, saying it could not market its blood-pressure estimation technology without FDA approval. The FDA tries to make a decision on a marketing authorization application within 150 days for low to moderately risky devices that have no predicates, but it can take longer.

The FDA declined to comment saying the issue falls outside the agency’s regulatory purview.

In the past, the line has been clear, said Zach Boyd, director of Utah’s artificial intelligence policy office: States regulate the practice of medicine and the FDA regulates devices.

“Now we’re in this weird place where there are devices — maybe you could call them devices — that are purporting to practice medicine,” said Boyd.“Our philosophy has been to just take care of our side — of the state’s authority — and the FDA is going to figure out what it’s going to figure out.”

https://www.politico.com/news/2026/01/06/artificial-intelligence-prescribing-medications-utah

Source: ExtremeTech, Adrianna Nine
Photo: Wladimir Bulgar/Science Photo Library via Getty Images

The model is the first of its kind to use multiple types of imaging and language-based data to assess a cancer patient’s health.

Stanford Medicine has developed an artificial intelligence model that can accurately predict cancer patients’ prognoses and responses to treatment. The first of its kind to leverage multiple types of imaging and language-based data, the model has already shown promise with several forms of cancer, including lung cancer, gastroesophageal cancer, and melanoma.

Over the last few years, researchers have created a range of experimental AI models that examine imaging data for tiny signs of cancer that doctors and radiologists might easily miss. Early tests show that these models are highly effective. Sybil, a model developed by MIT and the Massachusetts General Cancer Center, can predict patients’ one-year lung cancer development with an 86% to 94% accuracy rate, while Harvard Medical School’s pancreatic cancer prediction model can map a patient’s three-year prognosis with 88% accuracy. Another MIT model even spots signs of the riskiest forms of breast cancer to shield patients from overtreatment.

Impressive as these models are, they share in common one essential shortcoming: They’re only capable of analyzing one form of data at a given time. Each model looks at MRI scans or CT scans or X-ray images or microscopy slides, then identifies areas of concern within that dataset. Even Microsoft’s multi-diagnostic AI model, which accepts a whopping nine forms of imaging data, must examine those types of imaging separately.

Stanford Medicine’s model, MUSK (short for multimodal transformer with unified mask modeling), looks at several types of data at once. In a paper for Nature, the researchers write that MUSK was trained on 50 million pathology images and 1 billion “text tokens” from more than 11,500 patients. Although the images depict different forms of cancer across X-rays, microscopy, and CT and MRI scans, the text tokens represent language-based medical data—exam notes, communications between specialists, and so on—associated with various cancer diagnoses.

MUSK’s ability to analyze multiple types of data simultaneously mimics how doctors assess a person’s imaging results and health records. It also allows MUSK to assist doctors in predicting prognoses, not making diagnoses, the latter of which most medical AI models are focused on.

Across the 16 major types of cancer on which MUSK was trained, the model is capable of accurately predicting a patient’s disease-specific survival 75% of the time, according to a Stanford Medicine release. That’s an 11% improvement over doctors’ average accuracy rate, which hovers around 64%. MUSK has also correctly identified which non-small cell lung cancer patients would benefit from immunotherapy 77% of the time (beating doctors’ 61% accuracy rate) and predicted which melanoma patients were most likely to relapse within 5 years of initial treatment with 83% accuracy.

“The biggest unmet clinical need is for models that physicians can use to guide patient treatment,” said senior study author and radiation oncologist Ruijiang Li. “Does this patient need this drug? Or should we instead focus on another type of therapy? If we can use artificial intelligence to assess hundreds or thousands of bits of many types of data, including tissue imaging, as well as patient demographics, medical history, past treatments, and laboratory tests gathered from clinical notes, we can much more accurately determine who might benefit.”

https://www.extremetech.com/science/stanford-medicines-ai-model-accurately-predicts-cancer-prognoses-treatment

Source: Knowledge@Wharton, Angie Basiouny
Photo: Knowledge@Wharton

Penn’s Michael Platt and Peter Sterling say the global fertility rate is falling because of broken social bonds and a sense of collective sadness.

Rising inequality and social isolation have led to an “epidemic of despair” that is driving down fertility rates worldwide, according to a new paper by Penn neuroscientists Michael Platt and Peter Sterling.

This acute sense of loneliness and anxiety is contributing to more physical and mental ailments, particularly in high-income countries, and dampening the most basic of human desires — procreation. The U.S. birth rate has declined by an average 2% a year for the last decade. The global fertility rate has plunged to 2.3 (live births per person) and is expected to continue decreasing below the 2.1 rate needed for population replacement.

“My co-author and I are just kind of gobsmacked by this because this is not what species do. They don’t decline in numbers, because reproduction is the driver of evolution,” Platt said during an interview with Wharton Business Daily on SiriusXM. (Listen to the podcast.)

The paper, “Declining Human Fertility and the Epidemic of Despair,” appears in the journal Nature Mental Health.

Declining Fertility Rates Are Bad for Business

A declining population may be beneficial for a planet already scarred by the effects of climate change and resource scarcity. But it could have profound effects on economies and labor markets, the professors said. Without enough young people, it will be difficult to staff work that requires “young muscle,” such as construction and the military, or find new recruits for fields such as medicine and engineering. There will be fewer consumers overall, and an overall reduction in the wages that generate taxes for programs like Social Security.

“Knowledge work may be a little bit less [affected] because of the rise of AI and tech. But that in and of itself is probably accelerating the conditions that we think are actually driving part of this fertility decline,” Platt said.

The professor pointed to a catastrophic rise in anxiety, depression, and obesity-related diseases such as diabetes that corresponds with the rise of digital culture, where people are interacting with screens more than with each other. The problems are worst among teen girls, who are reporting record high rates of sadness and suicidal thoughts.

Government Interventions Have Yet to Slow Fertility Rate Decline

All these factors coalesce to create a “negative momentum,” which the professors explain as a drop in the dopamine-inducing rewards that usually come from material gains and deep social bonds.

“If you’re spending more time on your phone or in front of a screen, you’re not out experiencing real life and making real connections, making real friends,” Platt said. “And you’re potentially limiting your ability to find the person you’re going to fall in love with and start a family.”

The professors argue that government interventions to encourage having babies, like subsidizing child care, have had little effect on the downward fertility trend. More foundational changes are needed, they said. Countries that have banned smartphones at school, for example, have reported improved mental health and less bullying among students.

“I think, unfortunately, what we’re looking at is something more like a restructuring of our economic and social lives,” Platt said. “That’s a big task, but we can start small.”

https://knowledge.wharton.upenn.edu/article/fertility-rates-are-declining-is-tech-to-blame

Source: Smithsonian Magazine, Will Sullivan
Photo: Doctors perform a pig kidney transplant on April 12. So far, the patient is recovering well, but doctors will need to continue to monitor her in the hospital and make sure her immune system doesn’t reject the transplanted organ. (Joe Carrotta for NYU Langone Health)

The woman, 54-year-old Lisa Pisano, also received a mechanical heart pump implant days earlier, making her the first person to undergo both procedures

Lisa Pisano, a 54-year-old woman from New Jersey, has become the second living person to receive a transplant of a gene-edited pig kidney. Doctors also implanted a mechanical heart pump, making Pisano the first person to receive both a heart pump and an organ transplant, according to a statement from NYU Langone Health, where the procedure was performed.

Pisano had been suffering from heart failure and end-stage kidney disease and was on dialysis before the surgery. Because of her medical conditions, she was ineligible for heart and kidney transplants from human donors. The Food and Drug Administration approved the experimental procedure through a “compassionate use” program for patients with serious or immediately life-threatening conditions.

The NYU team said Wednesday that Pisano is recovering well, according to Lauran Neergaard of the Associated Press (AP).

“I feel fantastic,” Pisano said from the hospital during a Wednesday press conference over Zoom, per Wired’s Emily Mullin. “When this opportunity came, I said, ‘I’m gonna take advantage of it.’”

While Pisano hasn’t shown signs of rejecting the transplant so far, the critical point for that may not come until a month after surgery, Robert Montgomery, who led the procedure and is the director of NYU Langone’s transplant institute, tells Scientific American’s Tanya Lewis.

Montgomery tells NPR’s Rob Stein that Pisano will probably need to recover in the hospital for several months. He can’t predict how much time the procedure will buy for her.

“We’re optimistic that she’ll be able to go home and spend time with her children and grandchildren and live a comfortable life,” he says to NPR.

Patients with conditions like Pisano’s aren’t candidates for typical transplants, in part because their odds of survival are low, and there are not currently enough organs available to meet the vast need. More than 100,000 people are on the national transplant waiting list—with the majority in need of a kidney—and 17 people die every day waiting for a transplant in the U.S. Of the nearly 808,000 people in the country with end-stage kidney disease, only about 27,000 received a transplant last year, per NYU’s statement.

Scientists hope pig organs could combat this shortage and help more people get transplants. The NYU team has previously experimented with transplanting gene-edited pig organs into brain-dead patients with the consent of their families. They have also twice transplanted pig hearts into deceased human patients.

Two living patients who were experiencing heart failure have received pig heart transplants. The first lived for two months following the procedure, and the second died after six weeks.

Last month, a patient at Massachusetts General Hospital received the first pig kidney transplant. He was discharged from the hospital in early April and continued to recover at home.

“I do think it’s incredibly exciting we now have a second genetically edited pig kidney that’s been put into a living person,” Jayme Locke, a transplantation surgeon at the University of Alabama at Birmingham, who previously transplanted pig kidneys into brain-dead patients, says to Scientific American.

Pisano received the heart pump implant first, on April 4. She would have lived for only days or weeks without the device, according to NYU. She then received the gene-edited pig kidney in a second procedure on April 12.

“Without the possibility of a kidney transplant, she would not have been eligible as a candidate for [a heart pump] due to the high mortality in patients on dialysis with heart pumps,” Nader Moazami, an NYU cardiac surgeon who performed the heart pump surgery, says in the statement.

The pig kidney was genetically altered to disrupt the production of a sugar called alpha-gal. In previous studies, the researchers had shown that making this edit prevented the recipient’s immune system from rejecting the transplant. The team also attached a pig thymus gland to the tissue to reduce the likelihood of rejection.

Some experts question the ethics of such experiments.

“I think there are worries about conducting these experiments in this way, where we are finding the most desperate patients who have no other options,” L. Syd Johnson, a bioethicist at SUNY Upstate Medical University, tells NPR. “Maybe those patients will benefit. Maybe they believe they will benefit and that the risks are worthwhile for them. But I do worry about whether or not we are taking advantage of particularly vulnerable and desperate patients in conducting these experiments.”

Doctors say this transplant, as well as the one last month, will provide key information about the safety and effectiveness of the procedure.

“Like with any other transplant, in the early days, we want to make sure that we have control of the immune system and that there aren’t early rejections,” Locke tells Scientific American. “If this works, that’s incredibly positive.”

Will Sullivan is a science writer based in Washington, D.C. His work has appeared in Inside Science and NOVA Next.

https://www.smithsonianmag.com/smart-news/second-patient-receives-gene-edited-pig-kidney-transplant-in-breakthrough-surgery

Source: Smithsonian Magazine, Marla Broadfoot, Knowable Magazine
Photo: By recording activity of specific neurons in the brain, researchers aim to pick up signals of intended speech. (Knowable Magazine)

Implanted devices record the brain waves associated with speech and then use computer algorithms to translate the intended messages

In Alexandre Dumas’s classic novel The Count of Monte Cristo, a character named Monsieur Noirtier de Villefort suffers a terrible stroke that leaves him paralyzed. Though he remains awake and aware, he is no longer able to move or speak, relying on his granddaughter Valentine to recite the alphabet and flip through a dictionary to find the letters and words he requires. With this rudimentary form of communication, the determined old man manages to save Valentine from being poisoned by her stepmother and thwart his son’s attempts to marry her off against her will.

Dumas’s portrayal of this catastrophic condition—where, as he puts it, “the soul is trapped in a body that no longer obeys its commands”—is one of the earliest descriptions of locked-in syndrome. This form of profound paralysis occurs when the brain stem is damaged, usually because of a stroke but also as the result of tumors, traumatic brain injury, snakebite, substance abuse, infection or neurodegenerative diseases like amyotrophic lateral sclerosis (ALS).

In The Count of Monte-Cristo, Monsieur Noirtier de Villefort cannot move or speak but communicates with his granddaughter through words and letters in a book. Project Gutenberg

The condition is thought to be rare, though just how rare is hard to say. Many locked-in patients can communicate through purposeful eye movements and blinking, but others can become completely immobile, losing their ability even to move their eyeballs or eyelids, rendering the command “blink twice if you understand me” moot. As a result, patients can spend an average of 79 days imprisoned in a motionless body, conscious but unable to communicate, before they are properly diagnosed.

The advent of brain-machine interfaces has fostered hopes of restoring communication to people in this locked-in state, enabling them to reconnect with the outside world. These technologies typically use an implanted device to record the brain waves associated with speech and then use computer algorithms to translate the intended messages. The most exciting advances require no blinking, eye tracking or attempted vocalizations, but instead capture and convey the letters or words a person says silently in their head.

“I feel like this technology really has the potential to help the people who have lost the most, people who are really locked down and cannot communicate at all anymore,” says Sarah Wandelt, a graduate student in computation and neural systems at the California Institute of Technology in Pasadena. Recent studies by Wandelt and others have provided the first evidence that brain-machine interfaces can decode internal speech. These approaches, while promising, are often invasive, laborious and expensive, and experts agree they will require considerably more development before they can give locked-in patients a voice.

Engaging the brain—but where?

The first step of building a brain-machine interface is deciding which part of the brain to tap. Back when Dumas was young, many believed the contours of a person’s skull provided an atlas for understanding the inner workings of the mind. Colorful phrenology charts—with tracts blocked off for human faculties like benevolence, appetite and language—can still be found in antiquated medical texts and the home decor sections of department stores. “We, of course, know that’s nonsense now,” says David Bjånes, a neuroscientist and postdoctoral researcher at Caltech. In fact, it’s now clear that our faculties and functions emerge from a web of interactions among various brain areas, with each area acting as a node in the neural network. This complexity presents both a challenge and an opportunity: With no one brain region yet found that’s responsible for internal language, a number of different regions could be viable targets.

For example, Wandelt, Bjånes and their colleagues found that a part of the parietal lobe called the supramarginal gyrus (SMG), which is typically associated with grasping objects, is also strongly activated during speech. They made the surprising discovery while observing a tetraplegic study participant who has had a microelectrode array—a device smaller than the head of a pushpin covered in scads of scaled-down metal spikes—implanted in his SMG. The array can record the firing of individual neurons and transmit the data through a tangle of wires to a computer to process them.

Bjånes likens the setup of their brain-machine interface to a football game. Imagine that your brain is the football stadium, and each of the neurons is a person in that stadium. The electrodes are the microphones you lower into the stadium to listen in. “We hope that we place those near the coach, or maybe an announcer, or near some person in the audience that really knows what’s going on,” he explains. “And then we’re trying to understand what’s happening on the field. When we hear a roar of the crowd, is that a touchdown? Was that a pass play? Was that the quarterback getting sacked? We’re trying to understand the rules of the game, and the more information we can get, the better our device will be.”

In the brain, the implanted devices sit in the extracellular space between neurons, where they monitor the electrochemical signals that move across synapses every time a neuron fires. If the implant picks up on the relevant neurons, the signals that the electrodes record look like audio files, reflecting a different pattern of peaks and valleys for different actions or intentions.

The Caltech team trained their brain-machine interface to recognize the brain patterns produced when a tetraplegic study participant internally “spoke” six words (battlefield, cowboy, python, spoon, swimming, telephone) and two pseudowords (nifzig, bindip). They found that after only 15 minutes of training, and by using a relatively simple decoding algorithm, the device could identify the words with over 90 percent accuracy.

The Brain-Computer Interfaces That Could Give Locked-In Patients a Voice

Researchers carried out brain-computer interface experiments with the help of a volunteer with tetraplegia who agreed to have electrodes implanted in their brain. This figure shows readings from neurons in a brain region called the supramarginal gyrus, which is strongly activated during speech. The volunteer was tested with a number of words, as well as some nonsense words. The population of neurons triggering the electrode array behaved similarly to each other between tests (ITI, or inter-trial interval), but displayed a lot of variability when the subject was shown words (CUE), spoke the words internally (INTERNAL SPEECH) or spoke the words out loud (VOCALIZED SPEECH). Each colored line indicates a different word. The pattern of variance, as indicated by the ups and downs on the vertical axis, enabled words to be distinguished from each other by the computer software.

Wandelt presented the study, which is not yet published in a peer-reviewed scientific journal, at the 2022 Society for Neuroscience conference in San Diego. She thinks the findings signify an important proof of concept, though the vocabulary would need to be expanded before a locked-in patient could foil an evil stepmother or procure a glass of water. “Obviously, the words we chose were not the most informative ones, but if you replace them with yes, no, certain words that are really informative, that would be helpful,” Wandelt said at the meeting.

Thoughts into letters into words

Another approach circumvents the need to build up a big vocabulary by designing a brain-machine interface that recognizes letters instead of words. By trying to mouth out the words that code for each letter of the Roman alphabet, a paralyzed patient could spell out any word that popped into their head, stringing those words together to communicate in full sentences.

“Spelling things out loud with speech is something that we do pretty commonly, like when you’re on the phone with a customer service rep,” says Sean Metzger, a graduate student in bioengineering at the University of California, San Francisco and the University of California, Berkeley. Just like static on a phone line, brain signals can be noisy. Using NATO code words—like Alpha for A, Bravo for B and Charlie for C—makes it easier to discern what someone is saying.

Metzger and his colleagues tested this idea in a participant who was unable to move or speak as the result of a stroke. The study participant had a larger array of electrodes—about the size of a credit card—implanted over a broad swath of his motor cortex. Rather than eavesdropping on individual neurons, this array records the synchronized activity of tens of thousands of neurons, like hearing an entire section in a football stadium groan or cheer at the same time.

Using this technology, the researchers recorded hours of data and fed it into sophisticated machine learning algorithms. They were able to decode 92 percent of the study subject’s silently spelled-out sentences—such as “That is all right” or “What time is it?”—on at least one of two tries. A next step, Metzger says, could be combining this spelling-based approach with a words-based approach they developed previously to enable users to communicate more quickly and with less effort.

“Still in the early stage”

Today, close to 40 people worldwide have been implanted with microelectrode arrays, with more coming online. Many of these volunteers—people paralyzed by strokes, spinal cord injuries or ALS—spend hours hooked up to computers helping researchers develop new brain-machine interfaces to allow others, one day, to regain functions they have lost. Jun Wang, a computer and speech scientist at the University of Texas at Austin, says he is excited about recent progress in creating devices to restore speech, but he cautions there is a long way to go before practical application. “At this moment, the whole field is still in the early stage.”

Wang and other experts would like to see upgrades to hardware and software that make the devices less cumbersome, more accurate and faster. For example, the device pioneered by the University of California, San Francisco lab worked at a pace of about seven words per minute, whereas natural speech in American English moves at about 150 words a minute. And even if the technology evolves to mimic human speech, it is unclear whether approaches developed in patients with some ability to move or speak will work in those who are completely locked in. “My intuition is it would scale, but I can’t say that for sure,” says Metzger. “We would have to verify that.”

Another open question is whether it is possible to design brain-machine interfaces that do not require brain surgery. Attempts to create noninvasive approaches have faltered because such devices have tried to make sense of signals that have traveled through layers of tissue and bone, like trying to follow a football game from the parking lot.

Wang has made headway using an advanced imaging technique called magnetoencephalography (MEG), which records magnetic fields on the outside of the skull that are generated by the electric currents in the brain, and then translating those signals into text. Right now, he is trying to build a device that uses MEG to recognize the 44 phonemes, or speech sounds, in the English language—like ph or oo—which could be used to construct syllables, then words, then sentences.

Ultimately, the biggest challenge to restoring speech in locked-in patients may have more to do with biology than with technology. The way speech is encoded, particularly internal speech, could vary depending on the individual or the situation. One person might imagine scrawling a word on a sheet of paper in their mind’s eye; another might hear the word, still unspoken, echoing in their ears; yet another might associate a word with its meaning, evoking a particular feeling-state. Because different brain waves could be associated with different words in different people, different techniques might have to be adapted to each person’s individual nature.

“I think this multipronged approach by the different groups is our best way to cover all of our bases,” says Bjånes, “and have approaches that work in a bunch of different contexts.”

https://www.smithsonianmag.com/innovation/brain-computer-interfaces-that-could-give-locked-in-patients-voice

Source: Crunchbase, Keerthi Vedantam
Photo: Dom Guzman

Editor’s note: This story is Part Three of our series on artificial intelligence startups and their impact on multiple sectors. In Part One, we analyze VC investment in AI over the last decade. Part Two looks at the billions of dollars rolling into AI-enhanced cybersecurity. — Special Projects Editor Christine Kilpatrick

There are simply too few doctors in the U.S., and too many patients who need them.

Amid doctor burnout and long waitlists to see specialists, a niche in technology that saw slow adoption rates was suddenly in high demand: medical imaging AI used to aid in diagnostics. Such technology could help prescreen patients or work alongside physicians to scan images and help find problems that may have gone unnoticed by the tired, overworked human eye.

Funding for startups with this technology jumped from $348 million to over $1 billion between 2020 and 2021, per Crunchbase data. Though that number has dipped to $883 million so far in 2022, it’s still the second-largest year of funding for AI in diagnostics to date.

And yet, despite its inherent advantages, doctors are still apprehensive about this new technology.

“It really requires a quite nuanced understanding of: How is this going to fit into a doctor’s workflow?” said Jacob Effron, a health care-focused investor at Redpoint Ventures. “How does it fit into the incentives of different people in the system?”

An answer to doctor burnout

America’s doctor shortage has effectively turned every clinic, doctor’s office and care organization into need-based systems where only the most urgent patients could see a physician in a timely manner.

“Humans are missing a lot of diseases because there is an inherent mindset where they’re thinking, ‘Can I treat this patient tomorrow?’” said Kaushal Solanki, CEO and founder of medical imaging AI startup Eyenuk. “And that’s not the preferred threshold.”

Doctors are juggling large patient panels every day, unable to spend enough time with each patient to better treat them. This leads to burnout, which in turn translates into low-quality care. Most importantly, patients who can’t see a doctor frequently are treated for problems that could have been avoided if caught early.

That’s where companies like Eyenuk come in. The 12-year-old California-based company’s platform can autonomously diagnose diabetic retinopathy, a disease that quietly grows behind the eyes and can worsen without immediate medical intervention. It was granted FDA clearance in 2020 and has processed around 2 million images to date. Eyenuk raised $26 million in Series A funding in October led by AXA Investment Managers, according to Crunchbase data.

Eyenuk’s platform prescreens patients and allows ophthalmologists to prioritize who to see based on need. But the goal is to one day put the device in hospitals and primary care offices so doctors can check patients’ eyes instead of referring them to a hard-to-find specialist.

“This could be operated by anybody with a high school diploma and produce an actionable report that can define next steps for the patient, whether they are referred to an ophthalmologist or an eye care specialist, or they would be seen next year for repeat screening,” Solanki said.

This type of medical imaging technology is used in a slew of other sectors as well. Pearl, a venture-backed dentistry startup based in California that has raised $11 million, has a platform called Second Opinion (can you guess why?) that scans teeth imaging to point out a variety of tooth ailments to doctors. Israel-based Aidoc offers radiology-focused artificial intelligence tools to customers that scan radiology images to look for potential issues. The company has raised over $237 million.

Adoption still lags

While adoption spiked during the pandemic, rollout of medical imaging AI is yet to be as widespread as its visionaries would have hoped.

“It’s actually not a problem of the technology not being sophisticated enough,” said Choi. “It’s an adoption problem, and really proving out the use cases to convince providers that there’s business value as well as clinical value to these solutions.”

Provider buy-in is paramount for almost any kind of health care offering, but running through a packed schedule of patients makes it difficult to learn and embrace a new technology that may impede workflow, especially if they don’t think it will add much value.

There’s good reason for that skepticism. The American College of Radiology found that most AI platforms aren’t independently validated, calling into question the accuracy of these platforms. The Food and Drug Administration doesn’t have consistent qualifications for how big or diverse the training data set ought to be.

“We need these models to work transparently and be explainable. And that’s the difference that clinicians are looking for — because a doctor deserves to know how these machine-learning models are reading their patients,” said William Padula, an assistant professor of pharmaceutical and health economics at the University of Southern California. “The fear here is that while the programmer knows what they’ve done to create the model, it’s unclear how exactly it’s looking at the patient.”

But the promise of AI can’t be understated. In a post-pandemic health system, easy-to-access diagnostic resources are going to be important. Public health officials are pushing for more at-home or accessible diagnostic tests for all sorts of illnesses. And diagnostics accounts for 70% of all health care decision-making.

“We believe that the technology has to be good enough to work on its own. And that actually creates value for the system,” Solanki said. “Now there’s one less thing that the specialists have to worry about, which is routine screening.”

Check back for Part Four in our series, which spotlights some creative ways startups apply AI to their sectors.

https://news.crunchbase.com/ai-robotics/venture-funding-ai-health-care-series

Source: Knowledge@Wharton, Katherine Milkman
Photo: Vaccines

Do Vaccine Lotteries Work? Maybe

Wharton’s Katy Milkman shares the lessons learned from last year’s Philly Vax Sweepstakes, a Penn-funded project designed to evaluate ways of increasing COVID-19 vaccines in the city.

There were high hopes last year when Penn researchers and Philadelphia policymakers launched the Philly Vax Sweepstakes, a project designed to increase COVID-19 vaccinations by randomly awarding cash prizes to some residents who got the shot.

But the carefully designed, six-week program yielded mixed results. While the overall sweepstakes may have generated extra vaccinations in Philadelphia, its test of targeting specific zip codes with higher odds of a win definitely didn’t work.

Katy Milkman, a Wharton professor and co-director of the Behavior Change For Good Initiative (BCFG) at Penn, is the lead author of the study published this month in the journal Nature Human Behavior. A team of scientists at BCFG partnered with city officials to design and fund the sweepstakes, which was unique from other vaccine incentive programs that were rolling out across the country in 2021. Instead of winning free food, scholarships, gift cards, and other items, Philadelphia residents were treated to a “regret lottery” in which they were automatically entered if they lived in the city (whether or not they’d been vaccinated). They were then contacted if their name was selected and could win up to $50,000, but only if they could prove that they’d taken the shot before the drawing.

The university and the city worked to spread the news about the Philly Vax Sweepstakes through local media and social networks, with a clear message that residents of three particular randomly selected ZIP codes (one picked every two weeks) would have a 50 to 100 times greater chance than other Philadelphians of winning cash. The researchers randomly chose these three ZIP codes from among 20 with the lowest vaccination rates in the city, hoping the boost would incentivize more residents in those ZIP codes to get vaccinated. But the bigger grab at cash did nothing.

“We got an incredibly clear answer: no. It was totally useless to multiply the chances by 50 to 100 of residents in certain ZIP codes winning. We do not see a benefit of that,” Milkman told Wharton Business Radio on SiriusXM.

“Our best guess is that we generated about one extra vaccine per $15 spent on this regret lottery, so not a bad return on investment at all.”
— Katy Milkman

The controlled part of the experiment didn’t produce even one extra vaccine for every $5,000 spent. The researchers confidently ruled out any benefit. Milkman said the result is still useful because it shows that geo-targeting, which seems like a sound idea, should not be considered as a way to motivate people.

The professor said there may still be overall value that can be extracted from vaccine regret lotteries, and she believes more research is warranted.

“My takeaway is, this tactic needs more testing,” Milkman said. “We aren’t sure, but we probably increased vaccinations overall in the city of Philadelphia. Our best guess is that we generated about one extra vaccine per $15 spent on this regret lottery, so not a bad return on investment at all.”

The experiment’s design of a regret lottery was based on prior research indicating that a feeling of regret can be a powerful motivator. Milkman said although the Philly Vax Sweepstakes yielded equivocal results, it was generally more promising than the other types of incentive programs that were unveiled around the country last year with a lot of fanfare but little uptake. Another evaluation she co-authored (led by Penn Medicine professor Harsha Thirumurthy) looked at state vaccine lotteries and other incentives, and the results were much more discouraging.

“Maybe there would be value in more cities trying out regret lotteries,” she said. “It could have been a fluke; our results were equivocal. But it was encouraging enough that its worth more testing at that local, community level using this kind of regret-designed lottery.”

One indisputable finding from the study is the importance of mass communication, Milkman said. Spreading the news about the sweepstakes boosted interest in a very measurable way. Although every adult was eligible and registration wasn’t necessary, residents were told they could go to a website to verify that their names were on the sweepstakes list. In the few first weeks of radio, television, newspaper, and online coverage, more than 1 in 20 Philadelphians visited the site to confirm their inclusion.

“You do need to get the word out,” she said. “To get that kind of visibility and that kind of free marketing probably does require something like a city-wide opportunity.”

https://knowledge.wharton.upenn.edu/article/do-vaccine-lotteries-work-maybe

Source: Smithsonian Magazine, Margaret Osborne
Photo: Joy Milne can smell Parkinson’s Disease. (The University of Manchester)

The test, which would be the first of its kind for this disease, can make a diagnosis in about three minutes

Ten years after marrying her husband, Joy Milne noticed something odd: Her husband had taken on an “unpleasant” and “musty” odor, primarily around his shoulders and the back of his neck, she tells BBC News’s Elizabeth Quigley.

“I kept saying to him, ‘You’re not showering properly,’” Milne, a retired nurse from Perth, Scotland, says to Sky News. “And he became quite angry about it at first.”

As the years passed, the smell didn’t go away, and Milne began to notice her husband had gotten more moody. She took him to the doctor, and he was diagnosed with Parkinson’s disease.

Several years later, the couple joined a support group for Parkinson’s patients and families. When they entered the room, Milne noticed instantly the others had the same musty smell as her husband.

Milne, who has hereditary hyperosmia, or a heightened sense of smell, has worked with Parkinson’s disease doctors and researchers since her husband’s death in 2015. Now, a team of scientists from the University of Manchester in England, in tandem with Milne, say they’ve developed a simple skin swab test to detect the disease within about three minutes.

“This test has the potential to massively improve the diagnosis and management of people with Parkinson’s disease,” neurologist Monty Silverdale, the study’s clinical lead, tells EuroNews’ Natalie Huet.

The findings, published last week in the Journal of the American Chemical Society, dig deeper into exactly what Milne can smell. The scientists found that people with Parkinson’s have certain lipids of high molecular weight in their sebum—an oily substance found on the skin—that are more active, per a statement. Sebum tends to collect in the upper back region, the same area where Milne noticed a strong change in her husband’s scent.

The new skin swab test collects this sebum from patients’ backs. Using mass spectrometry, a tool for identifying compounds using their weight, researchers say this test can detect the disease with 95 percent accuracy under laboratory conditions, according to the BBC.

The team sampled 79 people with Parkinson’s and 71 healthy individuals. They identified 500 compounds that are different between people with the disease and those without it, per the statement.

Parkinson’s disease is a nervous system disorder that can cause shaking, trouble sleeping, limb stiffness and mental and behavioral changes. The disease starts slowly and worsens over time. Around 10 million people worldwide are living with Parkinson’s, and about 60,000 Americans are diagnosed with the disease every year.

Currently, there is no cure for the disease, and doctors can only diagnose it by monitoring patients’ symptoms, per the BBC. But a test could make a big difference for those with Parkinson’s, James Jopling, the Scotland director of Parkinson’s UK, tells the publication.

“People have to wait months or years to be diagnosed, so the fact that you could get the treatment and support you need and that researchers could begin new treatments is incredibly important,” Jopling says to the BBC.

The swab test is still in the early phases of research and has yet to be tested outside of laboratory conditions, per PA Media.

“We are hoping within two years to be able to start to test people in the Manchester area,” says Perdita Barran, a chemist at the University of Manchester who led the research, to the BBC.

Margaret Osborne is a freelance journalist based in the southwestern U.S. Her work has appeared in the Sag Harbor Express and has aired on WSHU Public Radio.

https://www.smithsonianmag.com/smart-news/woman-who-can-smell-parkinsons-helps-develop-a-skin-swab-diagnostic-test

Source: Crunchbase, Keerthi Vedantam
Photo: Dom Guzman

Dr. Audrey Wystrach spent 25 years treating dogs, cats and livestock in foul-smelling veterinary hospitals still operating on faxes and post-it notes. Antiquated booking processes stuck patients in the waiting room—sometimes for an hour—as vets scrambled to see them promptly.

It was a situation howling for change. Wystrach turned to her brother Michael (formerly the CEO and co-founder of meal-kit service Freshly), and in 2020 the Wystrachs launched Petfolk, a tech-enabled veterinary practice. Investors took note, and on Wednesday, Petfolk announced a $40 million Series A round led by White Star Capital to bridge the communication gap between pet owners and their vets.

“Honestly, I had better communication with my Starbucks than I did with my vet,” Michael Wystrach said.

The North Carolina startup combines telehealth with brick-and-mortar and mobile clinics to make it easier for pet owners to book appointments and access follow-up care instructions without dealing with long wait times. The ability to properly follow up on care, whether scheduling medicines at a certain time, exercising pets, or portioning out food, is key to making sure they don’t need to revisit the animal clinic.

“As you go to your human doctor, sometimes you leave scratching your head saying, ‘OK, what did they say? What do I do? How do I do this?’ And so we really wanted to create a support network of virtual care on top of physical space,” Audrey Wystrach said. “Empowering the customer with a ton of information will ultimately lead to better outcomes for our pets.”

Right now, Petfolk has four locations and mobile clinics in North Carolina, Florida and Georgia. It will soon offer 24/7 virtual care.

The rise of pet tech

The American Society for the Prevention of Cruelty to Animals found roughly one in five households adopted an animal during the pandemic, accelerating the need for veterinary care.

While venture funding toward the pet-related startup world is growing, it’s not exactly raining cats and dogs. Per Crunchbase data, 2021 saw a mere $578.9 million in the space for startups focused on pet care, mail-in pet diagnostics services, and pharmaceuticals. BondVet, billed as an urgent care for animals, raised $170 million in private equity money last year. The Vets, a mobile clinic and telehealth platform for pets, raised $40 million in seed funding in January.

But pandemic-era changes to human health care is slowly making its way to the pet industry, with telehealth and tech-enabled care much like OneMedical becoming more popular among a largely millennial crowd of pet owners.

“It was funny because a lot of Audrey’s original thesis and mission was really built around this idea of connected care, really being connected with virtual care and telemedicine …” Michael said. “And COVID only accelerated both of those things. So if anything, it just made our mission and our product that much more relevant for customers because it was that much more attuned to telemedicine.”

https://news.crunchbase.com/fintech-ecommerce/veterinary-care-telehealth-white-star