Shaping Solutions for Remote Care and Communications

Investigating the origins of disease outbreaks and transmissions is key to advising urgent responsive actions in affected areas. To dramatically improve the efficiency of this critical task, APL researchers are applying an advanced tool called ChainChecker to help responders visualize data and make quick, decisive actions in the field. The APL-enhanced system combines epidemiological and genomic data to track disease transmission. Initially developed by the Centers for Disease Control and Prevention (CDC) and Imperial College London, ChainChecker, bolstered by four years of APL development, has become pivotal to determining outbreak responses.

In work highlighted in the journal The Lancet Microbe in February 2024, the APL team applied the tool to investigating the 2020 Ebola outbreak in the Democratic Republic of the Congo retrospectively to prove the ChainChecker’s power. The outbreak involved concurrent clusters of the virus on opposite sides of the country, a distance that complicated tracing the chains of transmission and muddied any potential connection between the clusters themselves. ChainChecker helped researchers visualize how the virus spread and understand the outbreak’s dynamics; scientists confirmed two Ebola subtypes, one likely originating from a survivor of a previous outbreak. This dual analysis of genetics and transmission patterns underscored the utility of the approach for understanding disease dynamics.

“ChainChecker is giving responders quick and critical information on novel transmission modes, infection prevention and control measures, and other aspects of an outbreak,” explained Miles Stewart, an APL project lead and software engineer. “Our advanced tools are improving the ability of responders to make better decisions in the field during an outbreak through rapid, more accurate interpretation of data.”

Beyond Ebola, ChainChecker is being applied to other viral hemorrhagic fevers, like Marburg virus disease. APL’s collaboration with the CDC and international health agencies is improving response strategies and ensuring faster, more accurate interpretations of outbreak data.

“From data analysis to interdisciplinary collaboration, APL’s capabilities are changing our approach to complex biological challenges,” Stewart said.

Among those challenges is giving medics an easier way to deliver lifesaving care to people injured in remote conflict zones or natural disasters. APL researchers leveraged the power of augmented reality, predictive anatomy visualization and artificial intelligence (AI) to develop a tool that makes it possible to “see” beneath the skin and predict where organs are situated, said Anna Knight, an APL biomedical engineer who leads medical image integration for the project.

The anatomy visualization aid relies on a statistical shape atlas — a detailed map of variations in human anatomy that the team built with hundreds of CT scans chosen to match the gender, ethnicity and body shape diversity of war­fighters in the U.S. military. A user can predict the probable location of an injury in specific organs through visible landmarks on the patient’s body. Then, with an augmented reality headset, medics can see a digital overlay of the patient’s predicted anatomy.

“This tool could help to address a major challenge in field medicine, making it so medics can envision what lies beneath the surface and guide emergency response,” said Bobby Armiger, the project’s principal investigator and head of exploratory science for APL’s Research and Exploratory Development Department. Preliminary results showed the models can predict the individual shape of 66 different anatomical structures within the thorax alone.

The entire project sprang from medics’ need to improve trauma care on the battlefield. Recent shifts in American military engagements, characterized by prolonged conflicts in remote regions (such as Afghanistan) necessitate prolonged care in the field, underscoring the need for advanced trauma care in austere environments. Where in the past airlifting individuals to safety was a key part of the strategy, Armiger said the goal is to enable more self-sufficient, immediate care on the battlefield and in remote areas. The effort fits squarely within the APL Global Health Mission Area’s focus on assured care, counteracting natural, deliberate and accidental health threats through groundbreaking science and engineering.

“We’re developing technology that could one day contribute toward making care accessible wherever it’s needed the most,” said Alan Ravitz, chief engineer of the mission area.

While augmented reality is still uncommon on the battlefield, applications like this are geared toward enhancing medic training in the near term and may one day enable any service member to deliver battlefield medicine.

APL researchers are also refining methods to predict how someone will react to a pathogen, disease or treatment. The Immunity Twin project — initially funded as part of APL’s Innovation Program as a Propulsion Grant — combines biological data with advanced computational tools; and with touches of machine learning and systems modeling, an APL team is building a dynamic model of how the human immune system might react to various conditions.

“Our objective is to develop both digital and in vitro models of a person’s immune system to predict responses to diseases or treatments,” said Sarah Grady, a molecular biologist and the project’s principal investigator. “We are targeting the unpredictability in immune system responses, aiming for a future where personalized medical interventions are commonplace.”

To address this challenge, APL pulled on the multidisciplinary collaboration few institutions can match, assembling a team of virologists, computational modelers and data scientists to create both physical and digital replicas of immune systems. They’re focused on the respiratory system, studying how different lungs respond to the flu — a common virus that can be especially harmful to vulnerable populations.

Our objective is to develop both digital and in vitro models of a person’s immune system to predict responses to diseases or treatments. We are … aiming for a future where personalized medical interventions are commonplace.

— Sarah Grady, molecular biologist and principal investigator for Immunity Twin

According to APL biologist and project co-investigator Molly Gallagher, the integration of lab work and computational modeling ensures the team is collecting data that directly informs the digital model — separating the APL effort from similar work at other institutions.

Another unique aspect of APL’s approach is the use of “cell painting,” in which, by staining cells with fluorescent dyes, researchers can image various cell structures under a microscope. The difference? In a first, APL has modified the cell-painting method to work with living lung organoid cells, allowing the observation of real-time responses to viral infections.

“This allows us to observe real-time responses to viral infections,” Gallagher explained. “We can see what’s happening inside the cells over time, giving us more accurate data.”

APL also uses advanced 3D tissue culture models that mimic human lungs, providing more sophisticated insights than traditional single-cell models. Coupled with cutting-edge imaging systems, this approach enables researchers to collect real-time data and account for genetic differences that affect immune responses.

A key element of the project is its iterative approach to data integration. The team uses biological data to build its digital models, which are continuously refined through further experiments. This feedback loop is critical for developing predictive models with real-world applications that extend beyond individual health care to include military and pandemic preparedness.

“Our models are built to predict how the immune system might respond to new challenges,” said Grady, referencing the team’s work related to influenza. “With accurate simulations, we can make more personalized decisions about treatments and vaccines.”

Reducing Risks in Evolving Environments

As sea levels rise and natural defenses like coral and oyster reefs decline, coastal communities face increased risks from storms, hurricanes and tsunamis. In response, APL scientists have been researching innovative solutions to enhance coastal resilience, taking a multidisciplinary approach that combines engineering, marine biology and AI to strengthen natural defenses and build long-term coastal durability.

“We can’t rely on a single solution to establish coastal resilience,” said Sarah Herman, who leads the biological and chemical sciences program at APL. “We need a suite of targeted solutions that can solve the entirety of the problem without creating new challenges.”

One APL initiative delivering valuable insight is CATNIP (Causal-Cascading AI Tipping-Point Neuro-symbolic Intervention and Prediction), led by senior AI research scientist Jennifer Sleeman. CATNIP is an AI tool designed to predict how tipping points — critical thresholds that could tip a natural system into an entirely different state once crossed — trigger cascading effects on ecosystems and coastal infrastructure. The model builds on previous work with the Defense Advanced Research Projects Agency’s AI-assisted Climate Tipping-point Modeling program, which was the first demonstration of AI as a climate modeling assistant, where researchers can run experiments by asking the AI climate modeling assistant natural language questions, similar to using a large language model.

CATNIP extends this capability to examine the tipping points of coral reefs, sea-ice melt and sea-level rise so it can study how tipping points are interconnected and determine interventions that could mitigate individual and cascading tips.

The tool’s new coral reef model looks at how factors such as temperature, carbon dioxide levels, wave energy and pollution affect reef populations and local flooding, which in turn can lead to food insecurity, infrastructure damage and population displacement.

Shifting from marine ecosystems to heavily developed regions, APL researchers are creating a tool that can map air temperatures at a granular level, deciphering the contributions of every building, street, park and body of water to an area’s climate.

Using multisource data collection, including open-source sensors, satellite data and advanced interpolation methods, the team developed an accurate and nuanced model that predicts air temperature variations at a better-than-city-block scale.

“We need to not only better understand the patterns of air quality and temperature but also use this knowledge to drive design decisions that improve environments with dense infrastructure,” said Krista Rand, a systems engineer and disaster researcher focused in this area.

The team tested their idea in Baltimore, where existing academic relationships with Johns Hopkins University made it possible for the team to access air temperature data from sensors near the university’s Homewood Campus, which improved data availability in the otherwise sparse open-source network.

Protecting People, Food and Infrastructure Against Cyberthreats

The cyberthreat landscape is constantly evolving, and with machine learning algorithms, generative AI and internet-dependent devices entering homes, businesses and infrastructure at a rapid pace, threat actors have many more opportunities to target organizations and consumers. APL continues to spearhead efforts to combat such vulnerabilities.

In 2024, the Laboratory took a leading role in addressing critical threats to the U.S. food supply, including natural disasters, cyberattacks and even espionage. With growing concerns over the vulnerability of food and agriculture systems, the Laboratory has emerged as a key player in strengthening the resilience of these essential sectors.

APL excels at and has a track record of analyzing, managing, deciphering and using data. We have seen an explosion of data in the food, agriculture and veterinary defense problem space, and we need to leverage that data to better understand the health and resilience of our agricultural infrastructure and how to protect it.

— Karen Meidenbauer, biological scientist, public health professional, veterinarian and assistant program manager for Homeland Chemical, Biological, Radiological, Nuclear, Explosives Defense

Following the release of the White House’s National Security Memorandum on Strengthening the Security and Resilience of U.S. Food and Agriculture in 2022, APL convened a workshop that brought together agencies such as the Department of Homeland Security, the Department of Agriculture, and the Federal Bureau of Investigation to coordinate strategies and address gaps in food security.

“We saw a need for people to communicate and collaborate,” said Karen Meidenbauer, a biological scientist, public health professional, veterinarian and assistant program manager for Homeland Chemical, Biological, Radiological, Nuclear, Explosives Defense. APL’s leadership helped clarify agency roles and facilitated discussions about threats to food systems. Through this new community of practice initiative, the Lab is driving collaboration and innovation to secure the nation’s food systems while earning a role as a trusted partner in this critical area.

Beyond fostering collaboration, APL’s expertise in data analysis and threat assessment has been crucial in developing early warning systems to protect agricultural infrastructure. Ongoing Lab research and development efforts focus on identifying biothreats, improving food security and enhancing disease surveillance in livestock and crops.

“APL excels at and has a track record of analyzing, managing, deciphering and using data,” Meidenbauer said. “We have seen an explosion of data in the food, agriculture and veterinary defense problem space, and we need to leverage that data to better understand the health and resilience of our agricultural infrastructure and how to protect it.”

Outside the agricultural sphere, APL cybersecurity researchers created a groundbreaking tool called ACES, short for Adaptive Compositional Exploration via Symbolic execution, that accelerates the discovery of cyber vulnerabilities within systems, laying the foundation for the automatic generation of ways to analyze systems for vulnerabilities and mitigate those weaknesses. Traditional methods to spot software vulnerabilities suffer from “path explosion” — the exponential growth in the number of paths a computer can take to execute even the simplest piece of software. It is a lot like navigating inside a maze, said Jen Spero, the ACES principal investigator. “If you can make only decisions based on what’s in front of you,” she said, “there’s no telling if you’re making your way toward a desirable outcome or a dead end.”

ACES instead uses a technique akin to navigating a maze from above, rather than turn by turn; it breaks the software into smaller chunks, analyzes them in isolation and then emulates the program mathematically so it can evaluate many inputs simultaneously. Additional techniques make it possible to analyze the software without the source code and hone in on the chunks that seem most likely to contain critical weaknesses.

Right away, ACES could find vulnerabilities in binary code designed to intentionally create path explosions in about a tenth of the time that traditional techniques take. In later tests with real-world binary systems, it identified vulnerabilities in less than an hour that a traditional tool failed to find in a week.

After four years of development, the program now has analysis capabilities that allow it to reason across multiple interconnected binaries, and a Python-based application programming interface to make it tailorable.

In 2024, Spero’s team continued to focus on a human–machine teaming approach, building a foundation that will allow a human analyst to see what ACES is finding in real time so they can speed up vulnerability discovery and develop patches and other mitigations. The team is also integrating ACES with an innovative APL reverse-engineering tool called Quarry. Many efforts have focused on custom solutions tailored to a specific use, but something more generalized and scalable is needed to keep up with the breakneck pace of software creation and updates.

“A lot of people in the reverse engineering community are doing similar kinds of work, but the domain-specific workflows have a lot of arbitrary differences because people are looking at different systems,” said Malcolm Taylor, the Quarry principal investigator. “You might build a custom solution for one system, but it won’t scale up if you then need to reverse engineer 100 different systems.”

Quarry is the best of all of these worlds. It merges existing and widely used data science tools into a state-of-the-art system explicitly for reverse engineering work, yet it is flexible enough to accommodate customizable workflows and handle new and emerging tools.

Quarry is a living, breathing thing. We’re over the Moon with where it is already, but we hope this is just the beginning.

— Malcolm Taylor, Quarry principal investigator

So far, Quarry has unlocked doors so engineers could tackle problems at scales far larger than were ever possible before. Within a year of Quarry’s creation, government sponsors provided funding for further development, and it has been used to modernize and automate a reverse engineering workflow that previously took weeks or more to perform manually.

“Quarry is a living, breathing thing,” Taylor said. “We’re over the Moon with where it is already, but we hope this is just the beginning.”

Enhancing Technologies for Efficiency and Assurance

Large language models (LLMs), like those used in OpenAI’s ChatGPT and Google’s Gemini, are used increasingly for everything from helping students write essays to assisting businesses in market research on consumer behaviors and trends. These models, however, sometimes generate false and even absurd information that they present as true. These hallucinations, as they are called, can be comical in some cases, but when accuracy really matters, they can be detrimental or, sometimes, dangerous.

One common and effective way to mitigate this problem is to manually import the necessary information to prevent models from making mistakes — for instance, you might have a database of valid responses on hand, or some high-quality source material to draw from — but that is not always possible. Now, APL researchers are working on a different, potentially complementary method: asking the models to think twice.

Known as Lie-DAR (Lie Detecting with Automated Requests), the APL-developed system breaks a model’s response into individual claims before asking it to validate each claim and consider other relevant information. Similar responses before and after these inquisitions likely mean the information was accurate the first time around. But if there is a lot of variance, the LLM was likely hallucinating.

“Lie-DAR sits between the human user and the large language model, without interfering with the original output,” explained Kyle Klarup, an APL data scientist and the principal investigator of Lie-DAR. “You get the same response you would otherwise, but Lie-DAR assesses the accuracy of that response.”

Klarup emphasized that Lie-DAR doesn’t solve hallucinations, but it does attenuate the problem significantly. One analysis found LLM accuracy improved overall by 19% after Lie-DAR, although at the cost of it being slower because it had to double-check every claim. The team sees that as a sign that Lie-DAR will be most applicable to situations where accuracy matters more than speed, such as providing parameters that disease forecasting models need to make accurate predictions or identifying critical weaknesses in supply chains. The program is also structured so it can be quickly reconfigured to incorporate new functionality as new hallucination mitigation techniques become available.

Meantime, another APL team tackled getting to sources of truth in a different way. AI-generated images and videos, also known as deepfakes, are becoming easier to produce, more realistic and more difficult to detect. The combination has led to a rise in fraud and identity theft, harassment, misinformation and political interference.

In 2024, APL researchers in APL’s Special Operations Mission Area refined an APL-developed tool to better detect these images.

“The need to develop technologies that help our nation identify and engage against misinformation, and provide a credible deterrent in the information environment, is urgent,” said Dan Silvera, deputy mission area executive.

Known as GOLDFINCH (Generating Organic Looking Diverse Faces/Investigating Novel Concepts for High throughput), the tool builds sets of fake images that train a machine learning model to better identify what qualities signal a real photo versus a fake photo.

GOLDFINCH demonstrates the rapid adaptability of AI and APL’s abilities to quickly follow open-source development with sponsor-specific use cases.

— Emily Brown, assistant supervisor of APL’s Analytics Capabilities Group

“Instead of needing dozens, if not hundreds, of fake images to train our data sets, we only need one,” said GOLDFINCH’s lead developer, Parker Jackim. “We’ve significantly reduced the time it takes to train automated detectors that can identify deepfakes, which can ultimately speed up the entire detection process.”

Not only is GOLDFINCH fast, it is efficient and adaptable. The team developed code that utilizes open-source AI tools. As industry drives optimization of open-source systems, GOLDFINCH benefits from the continuous improvements, bug fixes and added capabilities.

“GOLDFINCH demonstrates the rapid adaptability of AI and APL’s abilities to quickly follow open-source development with sponsor-specific use cases,” said Emily Brown, assistant supervisor of APL’s Analytics Capabilities Group.