AI Health
Friday Roundup
The AI Health Friday Roundup highlights the week’s news and publications related to artificial intelligence, data science, public health, and clinical research.
September 20, 2024
In this week’s Duke AI Health Friday Roundup: the physiology behind “choking” under pressure; LLM classifiers for social determinants of health; wearables log differences in long COVID patients; ORI finalizes updates to research integrity rules; conversations with LLMs reduce adherence to conspiracy theories; Kolmogorov-Arnold networks for explainable AI; LLMs for protein tinkering; the evolution of outboard memory for physicians; pros and cons of AI-powered genomic prediction; much more:
AI, STATISTICS & DATA SCIENCE
- “Chatterjee, an assistant professor of biomedical engineering at Duke, will use the MIRA funding to develop generative AI algorithms to create molecules that can bind to––and potentially alter––proteins associated with cancer and other diseases….Because they don’t need to decipher the structure of the target protein, this approach allows Chatterjee and his team to aim for targets that have previously been out of reach. They are currently developing a platform to study how they can target and degrade the proteins associated with Alexander disease, an extremely rare genetic disease.” An article by the Duke Pratt School of Engineering’s Michaela Martinez describes recent work by Duke professor Pranam Chatterjee that leverages large language models to identify potential therapeutic avenues for modifying proteins.
- “AI researchers have long wondered if it’s possible for a different kind of network to deliver similarly reliable results in a more transparent way….An April 2024 study introduced an alternative neural network design, called a Kolmogorov-Arnold network (KAN), that is more transparent yet can also do almost everything a regular neural network can for a certain class of problems. It’s based on a mathematical idea from the mid-20th century that has been rediscovered and reconfigured for deployment in the deep learning era.” Quanta’s Steve Nadis describes recent research highlighting the potential for Kolmogorov-Arnold Networks to provide a more transparent, explainable kind of neural network AI.
- “In recent years, hopes and predictions have proliferated about the potential of AI and genomics to transform the UK’s approach to medicine – with greater levels of efficiency, precision and personalisation held up as the prize for investment and adoption…This report examines a technology at the vanguard of this promised transformation: AI-powered genomic health prediction (or AIGHP). AIGHP refers to a set of AI-driven techniques that enable predictions about people’s future health and drug responses to be made from genomic data.” A report by the Ada Lovelace Institute examines the potential of AI-powered genomic prediction technologies for healthcare in the UK.
- “Recently, Zakka and a physician-led team compared Almanac with other popular language models. The panel preferred Almanac’s answers over the others, finding them to be significantly more factual and complete. He said Almanac also offers better security against cyber-attacks because it won’t reply to a doctor’s question if it detects its been corrupted to output information that’s potentially incorrect or harmful.” An article available from Stanford University’s Scope blog describes Almanac, a large-language model AI design specifically to provide clinical and reference support for physicians.
- “…results suggest that supplementing training data with synthetic data may optimize predictive performance for identifying SDoH from the unstructured data present in clinical notes. Degradation in performance that was based on the training set and validation set suggests that the generalization of these models would potentially require a combination of authentic and synthetic notes, rather than just one or the other.” In a research article published in PNAS, Gabriel and colleagues share findings from a study that examined the use of LLM classifiers to identify social determinants of health from patient electronic health records.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “The researchers found that, in jackpot scenarios, the activity of neurons associated with motor preparation decreased. Motor preparation is the brain’s way of making calculations about how to complete a movement — similar to lining up an arrow on a target before unleashing it. The drop in motor preparation meant that the monkey’s brains were underprepared, and so they underperformed.” A Nature news article by Jude Coleman reports on recent research that sheds light on why monkeys (and people) tend to “choke” in stressful, high-stakes scenarios.
- “We found that individuals with Long COVID symptoms experience a significantly different trajectory in changes from their pre-infection RHR [resting heart rate] norms compared to those without Long COVID. Understanding the natural history of Long COVID by characterizing acute-phase changes and symptoms can potentially help identify individuals early on who have a greater likelihood of developing long-term symptoms or complications and identifying individuals who would have the greatest benefit from early interventions…” A research article by Radin and colleagues, published in NPJ Digital Medicine, presents findings from a study that sought to shed light on the course of long COVID by examining patient data gathered from wearable sensors.
- “In this phase 3 trial, neoadjuvant pembrolizumab combined with chemotherapy followed by adjuvant pembrolizumab resulted in a significant improvement, as compared with neoadjuvant chemotherapy alone, in overall survival among patients with previously untreated, high-risk, early-stage triple-negative breast cancer. After a median follow-up of 75.1 months, the 5-year overall survival was 4.9 percentage points higher with the addition of pembrolizumab.” A research article published by Schmid and colleagues in the New England Journal of Medicine describes results from a randomized trial neoadjuvant prembrolizumab plus chemotherapy in early-stage “triple-negative” breast cancer.
- “The medical reference landscape will continue to change as reference tools are integrated with clinical trials, machine-learning algorithms, and electronic medical records. Choices regarding reference tools are deeply intertwined with clinical and even sartorial norms (what pocket guide could fit into today’s favored clinical uniform, the embroidered fleece?). The demands and constraints associated with search algorithms and ‘prompt engineering’ may engender new ways of approaching, framing, and engaging with medical data.” An essay by Lea and colleagues, published in the New England Journal of Medicine, traces the evolution of the “peripheral brain” – that is, the outboard aids to memory, reference, and recall – used by physicians over generations.
COMMUNICATION, Health Equity & Policy
- “Today, the U.S. Department of Health and Human Services, through the Office of Research Integrity (ORI), finalized the 2024 Public Health Service (PHS) Policies on Research Misconduct. This rule updates the 2005 regulation and clarifies requirements for addressing research misconduct in PHS-funded research….This updated regulation clarifies both ORI’s regulatory oversight responsibilities and the role of PHS-funded organizations in institutionalizing research integrity in addition to refining requirements for addressing research misconduct in PHS-funded research.” The Department of Health and Human Services’ Office of Research Integrity has finalized a set of updates to rules governing research integrity for projects that receive HHS grant funding.
- “Without directly affecting any patient-facing care, the Change Healthcare attack profoundly disrupted the US health care system. The byzantine complexity of the network of ancillary services, supply chains, systems, and vendors supporting every hospital in the country is a web instinctually appreciated by every physician, but it is only after such linchpins fail that we fully appreciate their importance….The Change Healthcare attack hints at the existence of a tremendously consolidated and, therefore, vulnerable market for key health care infrastructure services.” A perspective article published in JAMA Internal Medicine by Neprash and colleagues advocates for a sense of urgency among health systems regarding the need for robust cybersecurity measures.
- “We found that transgender youth with unsupportive families, often marked by silence, violence, expulsion, clothing restrictions, or conversion therapy, have a higher probability of attempting suicide and running away at younger ages than transgender youth with supportive families, who often affirm their identity, use preferred names and correct pronouns, provide financial aid for transition, facilitate legal changes, actively research support methods, and advocate for them.” A research article by Campbell and colleagues, published in July in JAMA Pediatrics, presents findings from a survey study of mental health among transgender youth.
- “Although conspiracy theories are widely seen as a paradigmatic example of beliefs that rarely change in response to evidence, we hypothesized that dialogues with LLMs—which can use facts and evidence to rebut the specific claims made by any given conspiracy believer—would be efficacious in debunking conspiracy beliefs. Our findings confirmed this prediction: A brief interaction with a pretrained LLM substantially reduced belief in a wide range of conspiracy theories. The robustness of this effect is particularly noteworthy…” A remarkable paper by Costello and colleagues, published in Science, presents findings from a study that examined whether conversations with large language models could reduce adherence to conspiracy-theory beliefs.
