In today’s Duke AI Health Friday Roundup: probing the limits of transformer models; the merits of visual explanations in healthcare; potential for racial and ethnic bias in algorithmic healthcare tools; gratitude ceremonies for donated bodies; AI targets antibiotic candidate for resistant pathogen; postdoc pipeline for life sciences in danger of drying up; climate change may narrow range of livable area on Earth for millions; much more:
AI, STATISTICS & DATA SCIENCE
- “Our empirical findings suggest that Transformers solve compositional tasks by reducing multi-step compositional reasoning into linearized subgraph matching, without necessarily developing systematic problem-solving skills. To round off our empirical study, we provide theoretical arguments on abstract multi-step reasoning problems that highlight how Transformers’ performance will rapidly decay with increased task complexity.” A preprint by Dziri and colleagues, available from arXiv, presents both findings and conjecture regarding the potential limitations of transformer/LLM AIs, and addresses why the models perform so impressively in some areas, but flounder at some seemingly simple tasks.
- “The development of a useful medical AI tool requires a high-quality and diverse set of data, a multi-disciplinary approach, and a rigorous validation process to demonstrate its performance and value in real-world settings. The AI needs to be designed with human factors in mind and requires continuous monitoring and improvement to ensure high performance and patient impact.” A commentary article published in Nature Medicine by Widner and colleagues describes the experience (and resulting lessons) of developing and deploying a medical AI application.
- “In short, ChatGPT and related tools based on large language models (LLMs), which include Microsoft Bing and GitHub Copilot, are incredibly powerful programming aids, but must be used with caution.” A feature article in Nature by Jeffrey M. Perkel presents advice for coding with ChatGPT.
- “…participants perceived widespread and increasing use of algorithms in health care and lack of oversight, potentially exacerbating racial and ethnic inequities. Increasing awareness for clinicians and patients and standardized, transparent approaches for algorithm development and implementation may be needed to address racial and ethnic biases related to algorithms.” A paper published in JAMA Health Forum by Jain and colleagues present findings from a qualitative survey study of the use of healthcare algorithms based on patient race/ethnicity.
- “…we screened ~7,500 molecules for those that inhibited the growth of baumannii in vitro. We trained a neural network with this growth inhibition dataset and performed in silico predictions for structurally new molecules with activity against A. baumannii. Through this approach, we discovered abaucin, an antibacterial compound with narrow-spectrum activity against A. baumannii.” A research article published in Nature Chemical Biology by Liu and colleagues describes the use of a deep learning model to screen candidate antibiotics for the treatment of a drug-resistant pathogen.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “As optimum conditions shift away from the equator and toward the poles, more than 600 million people have already been stranded outside of a crucial environmental niche that scientists say best supports life. By late this century, according to a study published last month in the journal Nature Sustainability, 3 to 6 billion people, or between a third and a half of humanity, could be trapped outside of that zone, facing extreme heat, food scarcity and higher death rates…” An ominous article by ProPublica’s Abraham Lustgarten reports on a recent study that projects that hundreds of millions of humans are living in regions that may become only marginally or intermittently habitable due to climate change.
- “The experience teaches more than the foundations of medicine. Treating the donor, who is viewed as a doctor’s first patient, with respect and care gives students a grounding in ethics and professionalism, said Joy Balta, the chair of the American Association for Anatomy’s human body donation committee.” The New York Times’ April Rubin takes us inside medical school “gratitude ceremonies” that honor the people who donate their bodies for medical education.
- “In this planned final analysis of overall survival from the phase 3 ADAURA trial, adjuvant osimertinib resulted in significantly longer overall survival than placebo among patients with completely resected, EGFR-mutated, stage II to IIIA NSCLC as well as in the overall population (patients with stage IB to IIIA disease). Kaplan–Meier curves showed early separation that was sustained beyond 5 years.” A study published in the New England Journal of Medicine by Tsuboi and colleagues (with results presented last week at the American Society of Clinical Oncology annual meeting) presents head-turning results in the treatment of non-small cell lung cancer. (Additional explanation of findings from STAT News’ Angus Chen and Matthew Herper.)
- “The AI tool — developed in collaboration with the company ArteraAI — digitally read the patients’ biopsies and clinical data to discern which of the men could be spared long-term ADT with no increased risk of the cancer spreading…The researchers found that the predictive AI biomarker identified 34% of men who could benefit from short-term androgen deprivation therapy (ADT), avoiding the side effects of prolonged ADT without compromising efficacy. It also identified 43% of intermediate-risk men who would benefit from long-term ADT to reduce their risk of metastases over time.” And in other ASCO news: Duke’s Sarah Avery reports on investigators from Duke Cancer Institute, who presented results of a study that used artificial intelligence to identify predictive biomarkers for tailoring prostate cancer treatments.
COMMUNICATION, Health Equity & Policy
- “There are increasing signs that academic science has lost its allure for many talented researchers. More life scientists than ever are leaving academia, with Ph.D. graduates skipping postdocs to jump into lucrative positions in private industry. The number of biomedical postdocs, which had risen for decades, has flatlined and now has begun to decline. Individual faculty and entire research institutes are having a harder time hiring postdocs, and those who do join academia are demanding better working conditions…” A feature at STAT News by Jonathan Wosen probes the extraordinary stresses currently affecting the US system of academic postdoctoral research in life sciences.
- “The standards around generative AI and copyright have not yet been settled legally, which is causing companies to hold off using generative AI in their business operations…The Firefly model is trained on stock images for which Adobe already holds the rights, as well as on openly licensed content (for example, Creative Commons images) and public-domain content.” At Fast Company, Chris Stokel-Walker reports that Adobe seems to feel sufficiently confident in the air-tightness of their image-generating Firefly AI’s copyright status that they are offering to fully indemnify any users of the enterprise version of the tool against copyright suits…
- …and on the other side of that legal coin, Reuters reports that stock photo giant Getty is requesting courts in Great Britain to clamp down on image generator Stability AI, which it claims is violating its copyright.
- “Doctors are people, like anyone else; although they are trained to use verbal communication, visuals help increase their understanding and prevent them from having to keep as much information in their ‘working memory’ at any one time.” A Medium article by Katie McCurdy and Chethan Sarabu explores the often-neglected importance of visual explanations in healthcare.
- “We found that although insurance coverage increased for all groups during our analysis period, the change was greatest for partnered LGBT people, and for this group, the gains were driven by an increase in private health insurance coverage. This suggests the importance of dependent coverage. It is important to note that with our descriptive study design, it was not possible to disentangle the causal effect of the Obergefell ruling from other factors.” A research article published in Health Affairs by Bolibol and colleagues examines recent trends in health insurance coverage for LGBT adults.