Research Blog

Following the people and events that make up the research community at Duke

Students exploring the Innovation Co-Lab

Author: Victoria Wilson

My Face Belongs to The Hive (and Yours Does Too)

By

On October 4, 2023

In Artificial Intelligence, Business/Economics, Computers/Technology, Policy, Responsible Conduct

Imagine having an app that could identify almost anyone using only a photograph of their face. For example, you could take a photograph of a stranger in a dimly lit restaurant and know within seconds who they are.

This technology exists, and Kashmir Hill has reported on several companies that offer these services.

An investigative journalist with the New York Times, Hill visited Duke Law Sept. 27 to talk about her new book, Your Face Belongs To Us.

The book is about a company that developed powerful facial recognition technology based on images harnessed from our social media profiles. To learn more about Clearview AI, the unlikely duo who were behind it, and how they sold it to law enforcement, I highly recommend reading this book.

Hill demonstrated for me a facial recognition app that provides subscribers with up to 25 face searches a day. She offered to let me see how well it worked.

Screen shot of the search app with Hill’s quick photo of me.

She snapped a quick photo of my face in dim lighting. Within seconds (3.07 to be exact), several photos of my face appeared on her phone.

The first result (top left) is unsurprising. It’s the headshot I use for the articles I write on the Duke Research Blog. The second result (top right) is a photo of me at my alma mater in 2017, where I presented at a research conference. The school published an article about the event, and I remember the photographer coming around to take photos. I was able to easily figure out exactly where on the internet both results had been pulled from.

The third result (second row, left) unsettled me. I had never seen this photo before.

A photo of me sitting between friends. Their faces have been blurred out.

After a quick search of the watermark on the photo (which has been blurred for safety), I discovered that the photograph was from an event I attended several years ago. Apparently, the venue had used the image for marketing on their website. Using these facial recognition results, I was able to easily find out the exact location of the event, its date, and who I had gone with.

What is Facial Recognition Technology?

Researchers have been trying for decades to produce a technology that could accurately identify human faces. The invention of neural network artificial intelligence has made it possible for computer algorithms to do this with increasing accuracy and speed. However, this technology requires large sets of data, in this case, hundreds of thousands of examples of human faces, to work.

Just think about how many photos of you exist online. There are the photos that you have taken and shared or that your friends and family have taken of you. Then there are photos that you’re unaware that you’re in – perhaps you walked by as someone snapped a picture and accidentally ended up in the frame. I don’t consider myself a heavy user of social media, but I am sure there are thousands of pictures of my face out there. I’ve uploaded and classified hundreds of photos of myself across platforms like Facebook, Instagram, LinkedIn, and even Venmo.

The developers behind Clearview AI recognized the potential in all these publicly accessible photographs and compiled them to create a massive training dataset for their facial recognition AI. They did this by scraping the social media profiles of hundreds of thousands of people. In fact, they got something like 2.1 million images of faces from Venmo and Tinder (a dating app) alone.

Why does this matter?

Clearly, there are major privacy concerns for this kind of technology. Clearview AI was marketed as being only available to law enforcement. In her book, Hill gives several examples of why this is problematic. People have been wrongfully accused, arrested, detained, and even jailed for the crime of looking (to this technology) like someone else.

We also know that AI has problems with bias. Facial recognition technology was first developed by mostly white, mostly male researchers, using photographs of mostly white, mostly male faces. The result of this has had a lasting effect. Marginalized communities targeted by policing are at increased risk, leading many to call for limits on the use of facial recognition by police.

It’s not just government agencies who have access to facial recognition. Other companies have developed off-the-shelf products that anyone can buy, like the app Hill demonstrated to me. This technology is now available to anyone willing to pay for a subscription. My own facial recognition results show how easy it is to find out a lot about a person (like their location, acquaintances, and more) using these apps. It’s easy to imagine how this could be dangerous.

There remain reasons to be optimistic about the future of privacy, however. Hill closed her talk by reminding everyone that with every technological breakthrough, there is opportunity for ethical advancement reflected by public policy. With facial recognition, policy makers have previously relied on private companies to make socially responsible decisions. As we face the results of a few radical actors using the technology maliciously, we can (and should) respond by developing legal restraints that safeguard our privacy.

On this front, Europe is leading by example. It’s likely that the actions of Clearview AI are already illegal in Europe, and they are expanding privacy rights with the European Commission’s (EC) proposed Artificial Intelligence (AI) regulation. These rules include requirements for technology developers to certify the quality of their processes, rather than algorithm performance, which would mitigate some of these harms. This regulation aims to take a technology-neutral approach and stratifies facial recognition technology by it’s potential for risk to people’s safety, livelihoods, and rights.

Post by Victoria Wilson, MA Bioethics and Science Policy, 2023

Neuroscience Shows Why Sex Assault Victims “Freeze.” It’s Not Consent.

By

On July 6, 2023

In Artificial Intelligence, Behavior/Psychology, Neuroscience

Warning: the following article discusses rape and sexual assault. If you or someone you know has been sexually assaulted, help is available.

Image: DreamStudio AI, with prompt “Woman, screaming, sitting on the witness stand in a U.S. court of law, in the style of Edvard Munch’s ‘The Scream’”

“You never screamed for help?”

“Why didn’t you fight back?”

These are questions that lawyers asked E. Jean Carroll in her rape case against former president Donald J. Trump this spring. These kinds of questions reflect a myth about rape: that it’s only rape if the victim puts up a fight.

A recent review of the research, “Neuroscience Evidence Counters a Rape Myth,” aims to set the record straight. It serves as a call to action for those in the scientific and legal professions. Ebani Dhawan completed this work at the University College London with Professor Patrick Haggard. She is now my classmate at Duke University, where she is pursuing an MA in Bioethics & Science Policy.

Ebani Dhawan

Commonly accepted beliefs and myths about rape are a persistent problem in defining and prosecuting sexual assault. The intentions of all actors are examined in the courtroom. If a victim freezes or does not attempt to resist during a sexual assault, perpetrators may claim there was passive acquiescence; that consent was assumed from an absence of resistance.

From the moment a victim reports an assault, the legal process poses “why” questions about the survivor’s behavior. This is problematic because it upholds the idea that survivors can (and should) choose to scream or fight back during an assault.

This new paper presents neuroscientific evidence which counters that misconception. Many survivors of sexual assault report ‘freezing’ during an assault. The researchers argue that this is an involuntary response to a threat which can prevent a victim from actively resisting, and that it occurs throughout biology.

Animal studies have demonstrated that severe, urgent threats, like assault or physical restraint, can trigger a freeze response involving fixed posture (tonic immobility) or loss of muscle tone (collapsed immobility). Self-reports of these states in humans shed light on an important insight into immobility. Namely, that we are unable to make voluntary actions during this freezing response.

An example of this is the “lockup” state displayed by pilots during an aviation emergency. After a plane crash, it’s hard to imagine anyone asking a pilot if they froze because they really wanted to crash the plane.

Yet, quite frequently victims of sexual assault are asked to explain the freeze response, something which is further made difficult by the impaired memory and loss of sense of agency which often accompanies trauma.

The legal process around sexual assault should be updated to reflect this neuroscientific evidence.

THIS MYTH HAS REAL CONSEQUENCES.

The vast majority of sexual assault cases do not result in a conviction. It is estimated that out of every 1,000 sexual assaults in the U.S., only 310 are reported to the police and only 28 lead to felony conviction. That is a conviction rate of less than 3%.

In England and Wales, just 3% of rapes recorded in the previous year resulted in charges. According to RAINN, one of the leading anti-sexual assault organizations, many victims don’t report because they believe the justice system would not do anything to help — a belief that these conviction rates support.

E. Jean Carroll named this in her trial. She said, “Women don’t come forward. One of the reasons they don’t come forward is because they’re always asked, why didn’t you scream? You better have a good excuse if you didn’t scream.”

This research serves as a much-needed call-to-action. By revisiting processes steeped in myth, justice can be better served.

I asked Ebani what she thinks must be done. Here are her recommendations:

  1. The neuroscience community should pursue greater mechanistic understanding of threat processing and involuntary action processes and the interaction between them. 
  2. Activists and legal scholars should advocate for processes reflective of the science behind involuntary responses like freezing, and the inability of victims to explain that behavior.
  3. Neuroscientists should contribute to Police officers’ education regarding involuntary responses to rape and sexual assault.

“I’m telling you: He raped me whether I screamed or not.” – E. Jean Carroll

Post by Victoria Wilson, Class of 2023

Only Mostly Dead? The Evolving Ethics of Evaluating Death

By

On March 31, 2023

In Behavior/Psychology, Biology, Biomedical Engineering, Cardiology, Lecture, Medicine, Religion, Responsible Conduct

I recently had the pleasure of attending Professor Janet Malek’s lecture: Only Mostly Dead? The Evolving Ethical Evaluation of Death by Neurologic Criteria, a lecture sponsored by the Trent Center for Bioethics, Humanities & History of Medicine.

Dr. Malek is an associate professor in the Duke Initiative for Science & Society, and at the Baylor College of Medicine Center for Medical Ethics and Health Policy.

Janet Malek Ph.D.

We don’t often talk about death. On the surface, it seems like it would be a straight-forward concept. You’re either dead, or you’re not dead. Right? It turns out that clinically defining death is not so simple.

Popular media has some grasp on the ambiguity of the definition of death. Remember this scene from the popular movie, The Princess Bride? Suspecting that the protagonist is dead, his friends bring him to a miracle-worker and have the following conversation. 

Miracle Max: “Whoo-hoo-hoo, look who knows so much. It just so happens that your friend here is only MOSTLY dead. There’s a big difference between mostly dead and all dead. Mostly dead is slightly alive. With all dead, well, with all dead there’s usually only one thing you can do.

Inigo Montoya: What’s that?

Miracle Max: Go through his clothes and look for loose change.

In real life, death used to be determined by cardiopulmonary criteria – when the heart and lungs stop working.  In recent decades the idea that death can be determined using neurologic criteria – when the brain stops working – has gained acceptance. As neuroscience and technology has evolved, so too have our definitions. Now that we know more about how the brain works, we know that there may be some brain activity even after a person has met the criteria for death by neurologic criteria (DNC). This leads to philosophically rich and practically relevant questions of ethics – for example, when do we stop providing life-sustaining care? In the field of bioethics and beyond, there is high demand for discussion on this topic.

There has been controversy over defining death since the 1650’s — when a woman named Anne Greene woke up after being hanged. It wasn’t until the 1980’s that a consensus definition of death was first identified. Here is a brief history:

1950s

  • Widespread availability of ventilators led to the identification of a state described as death of the neurological system.

1960s

  • Advances in organ transplantation foster discussion on the ethics of defining death.
  • A committee at Harvard Medical School examined the definition of Brain Death. They created a definition of “Irreversible Coma,” which focused on loss of neurological function.

1980s

  • The 1980 Uniform Determination of Death Act (UDDA) provided a legal basis for clinically determining death as: an individual who has sustained either 1) irreversible cessation of circulatory and respiratory functions OR 2) irreversible cessation of functions of the entire brain.
  • 1981: President’s Commission for the Study of Ethical Problems in Medicine and Biomedical and Behavioral Research report. Findings are centered on questions of functioning of the organism as a whole and the brain’s role in coordinating it.

1990s-2000s

  • Clinicians arrive at general agreement that a patient in a state of coma or unresponsiveness, without brainstem reflexes and who fails an apnea test is dead by neurologic criteria. Largely it is accepted that “brain death is death” but there is not complete consensus.

2010-late

  • 2013: Case of Jahi McMath. A 13-year old girl was declared “brain dead” in California, and a death certificate was issued. However, the family fought to have her maintained on life support. They moved to New Jersey, the only state which recognized objections to brain death, and the “brain dead” declaration was reversed. Jahi lived there for 4 years before passing away. This famous case caused people to reconsider the concept of brain death.

2020s:

  • Recent innovations in heart transplantation technology will likely challenge the acceptance of the Dead Donor Rule (DDR) which requires that an individual is clinically declared dead before vital organs are removed for transplantation.
  • 2021: Assembly of the Determination of Death Committee, tasked with updating the Uniform Determination of Death Act (UDDA). Duke faculty (and founding director of Science & Society) Nita Farahany, is involved with this process.

What ethical issues and practical questions challenging Death by Neurologic Criteria (DNC) today? Dr. Malek shared the following case.

Following a tragic car accident, Ms. Jones, a 20-year-old college student, was brought to the hospital, having suffered significant anoxic brain injury. The medical team determined that she met criteria for DNC. However, her family refused to allow for further testing. Several days passed. Ms. Jones was maintained on life support, during which she did not show signs of improvement. After several difficult conversations, the family consented for assessment and Ms. Jones was declared dead — using the criteria associated with DNC.

What is the proper amount of time to continue life-sustaining treatment if a physician suspects the patient will never recover?

Although this may sound like an uncommon occurrence, nearly half of neurologists have been asked to continue neurologic support for patients that may meet criteria for DNC.

Obligating life support for patients suspected of meeting DNC, either through the family’s refusal for testing or by direct request, would likely result in ethical harms such as violation of the dignity of decedent, unjustly using scarce resources, or causing moral distress in caregivers.

However, it may be permissible to maintain life support in these situations. Dr. Malek says that we do not yet have a good ethical framework for this. Reasonable accommodations that are in line with professional guidelines probably have minimal impact, and might provide some psychosocial benefits to families.

Is consent required to test for DNC? Should it be?

Legal and professional standards favor the idea that testing for DNC likely falls under the category of implied consent, which assumes that a person would want reasonable medical care in the event of unconsciousness. In fact, 80% of neurologists think that getting consent for these evaluations is unnecessary.

These are extremely difficult questions, and there is continuing controversy over what the correct answers should be. Dr. Malek advises medical experts to work with healthcare administrators to develop clear institutional policies.

Post by Victoria Wilson, 2023 MA student in Bioethics & Science Policy

Student Researchers Share What They Know About AI and Health

By

On December 8, 2022

In Artificial Intelligence, Data, Medicine, Students

The healthcare industry and academic medicine are excited about the potential for artificial intelligence — really clever computers — to make our care better and more efficient.

The students from Duke’s Health Data Science (HDS) and AI Health Data Science Fellowship who presented their work at the 2022 Duke AI Health Poster Showcase on Dec. 6 did an excellent job explaining their research findings to someone like me, who knows very little about artificial intelligence and how it works. Here’s what I learned:

Artificial intelligence is a way of training computer systems to complete complex tasks that ordinarily require human thinking, like visual categorization, language translation, and decision-making. Several different forms of artificial intelligence were presented that do healthcare-related things like sorting images of kidney cells, measuring the angles of a joint, or classifying brain injury in CT scans.

Talking to the researchers made it clear that this technology is mainly intended to be supplemental to experts by saving them time or providing clinical decision support.

Meet Researcher Akhil Ambekar

Akhil standing next to his poster “Glomerular Segmentation and Classification Pipeline Using NEPTUNE Whole Slide Images”

Akhil Ambekar and team developed a pipeline to automate the classification of glomerulosclerosis, or scarring of the filtering part of the kidneys, using microscopic biopsy images. Conventionally, this kind of classification is done by a pathologist. It is time-consuming and limited in terms of accuracy and reproducibility of observations. This AI model was trained by providing it with many questions and corresponding answers so that it could learn how to correctly answer questions. A real pathologist oversaw this work, ensuring that the computer’s training was accurate.

Akil’s findings suggest that this is a feasible approach for machine classification of glomerulosclerosis. I asked him how this research might be used in medicine and learned that a program like this could save expert pathologists a lot of time.

What was Akhil’s favorite part of this project? Engaging in research, experimenting with Python and running different models, trying to find what works best.

Meet Researcher Irene Tanner

Irene Tanner and her poster, “Developing a Deep Learning Pipeline to Measure the Hip-Knee-Ankle Angle in Full Leg Radiographs”

The research Irene Tanner and her team have done aims to develop a deep learning-based pipeline to calculate hip-knee-ankle angles from full leg x-rays. This work is currently in progress, but preliminary results suggest the model can precisely identify points needed to calculate the angles of hip to knee to ankle. In the future, this algorithm could be applied to predict outcomes like pain and physical function after a patient has a joint replacement surgery.

What was Irene’s favorite part of this project? Developing a relationship with mentor, Dr. Maggie Horn, who she said provided endless support whenever help was needed.

Meet Researcher Brian Lerner

Brian Lerner and his poster, “Using Deep Learning to Classify Traumatic Brain Injury in CT Scans”

Brian Lerner and his team investigated the application of deep learning to standardize and sharpen diagnoses of traumatic brain injury (TBI) from Computerized Tomography (CT) scans of the brain. Preliminary findings suggest that the model used (simple slice) is likely not sufficient to capture the patterns in the data. However, future directions for this work might examine how the model could be improved. Through this project, Brian had the opportunity to shadow a neurologist in the ER and speculated upon many possibilities for the use of this research in the field.

What was Brian’s favorite part of this project? Shadowing neurosurgeon Dr. Syed Adil at Duke Hospital and learning what the real-world needs for this science are.

Many congratulations to all who presented at this year’s AI Health Poster Showcase, including the many not featured in this article. A big thanks for helping me to learn about how AI Health research might be transformative in answering difficult problems in medicine and population health.

By Victoria Wilson, Class of 2023

Cancer Stigma, Contraceptives, Covid-19: 2022 Global Health Research Showcase

By

On October 21, 2022

In Cancer, Field Research, Global Health, Medicine, Policy, Science Communication & Education

Last Monday, Oct. 17, Duke University students who had conducted global health research had the opportunity to present their work. From North Carolina to Sub-Saharan Africa, the 2022 Global Health Research Showcase featured works that tackle some of the world’s most pressing health issues. Over 40 undergraduate, Masters, and PhD student projects examined a broad range of issues, determinants, and phenomena in countries from almost every continent. Here’s a few project highlights, in case you missed it:

Maeve Salm, pursuing her Master of Science in Global Health, went to Tanzania to study contraceptive use. Tanzania’s youth are highly impacted by teen pregnancy, and Salm wanted to understand desires for contraceptive use among adolescents affected by HIV. She learned that, much like in the U.S., stigma influences access to sexual healthcare for adolescents. This qualitative study aimed to support young people in achieving their desired health outcomes and reducing HIV transmission by examining barriers and facilitators to family planning. Findings indicate that youth agency in reproductive health is of utmost importance.

Maeve Salm presenting her poster at the 2022 Global Health Research Symposium.

Wondering about the Covid-19 response in other countries? Master of Science in Global Health Candidate Stephanie Stan explored the barriers and enablers to the pandemic response in Peru. Per capita, Peru experienced the highest mortality rate form the disease compared to any other country. Due to several challenging factors, they were slow to receive COVID-19 vaccines. However, they implemented highly successful vaccination campaigns once vaccines were obtained. What can be learned from Peru’s pandemic response? Prolonged and proactive collaborations between sectors (healthcare, academics, and government) enable swift public health responses in a crisis. It’s important to have elected officials who are empowered to make decisions promoting science.

“Definitely meeting all the incredible people that I interviewed and learning about their work and involvement in Peru’s pandemic response. Learning about what happens moving forward from their point of view.”

Stephanie Stan, when asked about her global health research experience

Winning the first-place Graduate Student Research Award, Judith Mwobobia’s project examined the stigma of cancer in sub-Saharan Africa. Stigma is a huge barrier to receiving treatment, which is a problem considering that 70% of global cancer deaths originate from Africa. Perceptions of financial stress, misconceptions about cancer, and fear of death were common attitudes driving cancer stigma. Proposed interventions included education and policy recommendations for low-resourced communities. Mwobobia is pursuing her Master of Science in Global Health. Clearly a supportive group, her classmates erupted in cheers when the award was announced.

By Victoria Wilson, Class of 2023

Powered by WordPress & Theme by Anders Norén