We are moving our newsletter to Substack for a better experience!
In Week #218 of the Doctor Penguin newsletter, the following papers caught our attention:
1. Middle Ear Infection. Acute otitis media (AOM), also known as middle ear infection, is a common illness in children. However, the accuracy of AOM diagnosis has consistently been 75% or lower among primary care and pediatric practitioners. Can AI improve the diagnostic accuracy of AOM?
Shaikh et al. developed and validated a deep residual-recurrent neural network for detecting AOM in children from otoscopic videos of the tympanic membrane. The data was collected from a primary care setting and included non-ideal or partially obstructed images, captured using a smartphone connected to an endoscope or an otoscope during outpatient clinic visits at two sites in Pennsylvania between 2018 and 2023. The model demonstrated a sensitivity of 93.8% and specificity of 93.5%, exhibiting higher accuracy than pediatricians, primary care physicians, and advanced practice clinicians. These findings suggest that given its high accuracy, the model could reasonably be used in primary care or acute care settings to aid with automated diagnosis of AOM and decisions regarding treatment; Given its high specificity, the tool could also be employed at the time of triage by a trained nonphysician to prevent repeated examinations that may occur in some teaching environments.
Read paper | JAMA Pediatrics
2. Endovascular Robot. Treating vascular diseases in the brain requires access to the affected region inside the body. Navigating tortuous vessels using long, thin devices, such as wires and tubes, is challenging for surgeons to perform. This process often requires multiple attempts, repeatedly stressing the vessel walls, which increases the risk of vasospasm and vessel dissections.
Dreyfus et al. developed a highly dexterous, magnetically steered continuum robotic device designed to navigate the dense and tortuous arterial structure of the brain. The device features a helical protrusion on its outer surface, which enables forward motion when rotated due to the engagement between the outer surface and the vessel wall. Embedded with magnetic material, the device's tip is wirelessly steered by an external magnetic field that applies a magnetic torque on the embedded magnetic material. The effectiveness of this helical device has been demonstrated through successful navigation experiments conducted in models of the human vasculature and in blood vessels of a live pig.
Read Paper | Science Robotics
3. Multimodal AI. How could multimodal AI help reduce cardiovascular diseases?
In this review, Muse and Topol emphasize the transformative potential of the current phase of AI, which leverages transformer models and generative AI to integrate multimodal datasets for self-supervised learning and prediction. This advancement has unlocked the potential to combine intricate data sources, including medical imaging, labs, clinic notes, genomics, microbiomes, and digital sensors. The authors delve into the areas where multimodal AI is likely to have the greatest impact on reducing cardiometabolic disease. These areas include risk assessment of coronary artery disease, blood pressure monitoring and management, sleep optimization via improved diagnoses of sleep disorders, quantification of stress and depression, physical activity tracking and personalized exercise coaching, enhanced diagnosis and treatment strategies for diabetes, and precise nutrition based on predicted personalized glucose response and the integration of gut microbiome information.
Read Paper | Cell Metabolism
4. AI for Science. AI might make us believe we understand more about the world than we actually do.
In this Perspective, Messeri and Crockett explore the potential epistemic risks associated with treating AI as scientific collaborators rather than mere tools. They argue that this approach may make scientists susceptible to various illusions of understanding, including the illusion of explanatory depth (overestimating one's level of understanding), the illusion of exploratory breadth (falsely believing that AI tools enable the exploration of all testable hypotheses, whereas they are actually exploring a narrower space of hypotheses testable using AI tools), and the illusion of objectivity (assuming that AI tools are unbiased and represent all possible standpoints, whereas AI tools actually embed the standpoints of their training data and developers). The authors identify four distinct visions of AI: Oracle, Surrogate, Quant, and Arbiter, and encourage researchers to be clear about why they want to use AI in their research. By doing so, scientists can better recognize the epistemic risks associated with each vision and take steps to mitigate them. Additionally, the authors suggest that working in cognitively and demographically diverse teams can help individuals further reduce these epistemic risks.
Read Paper | Nature
-- Emma Chen, Pranav Rajpurkar & Eric Topol