#18 | An AI To Monitor & Save Patients

Quote of the Week 💬 

An AI Alert System Monitors Predicts Heart Risk & Saves Lives

⌛ The Seven Second Summary: A team of researchers in Taiwan designed an AI-based alert system to notify physicians when hospitalized patients were at a high risk of death. This AI alert system helped physicians reduce patient mortality by meaningful amounts when compared to physicians who followed standard care procedures.

🔬 How It Was Done:

• The research team trained an AI system on a dataset of over 450,000 electrocardiogram (ECG) tests and the survival data of each corresponding patient. This data allowed the model to identify and predict when certain heart monitoring data reached high-risk thresholds that would result in a patient’s death without medical intervention.

• After a series of tests to refine the model’s performance, the researchers designed a complimentary alert system to integrate within a hospital setting and notify physicians when a patient’s heart data reached or surpassed the model’s high-risk thresholds.

• Afterwards, a randomized controlled trial was conducted over several months at Taiwan's National Defense Medical Center, involving 39 physicians and 15,965 patients randomly divided into two physician care groups. One group with 7,292 patients utilized this AI-based alert system, while the other group with 7,276 patients followed standard intervention procedures.

🧮 Key Results:

• The overall mortality rate for the group with the AI alert system was lower (3.6%) than the overall mortality rate of the control group (4.3%) after 90 days.

• Furthermore, when only considering the mortality rate of high-risk patients, the group with the AI alert system was significantly lower (16%) than the overall mortality rate of the control group (23%) after 90 days. This 30% mortality reduction of high-risk patients is consistent with the results of a preliminary meta analysis of this AI-based intervention system, which indicated a 29% mortality reduction in critically ill patients.

• Incredibly, the risk of a cardiac death fell by 91% for high-risk patients within the AI alert system group when compared to the high-risk group following standard intervention procedures.

💡 Why This May Matter: The researchers note their AI system was able to identify high-risk patients who did not showcase significant vital sign changes or issues that would ordinarily trigger medical interventions. This suggests a well-trained AI system may be able to identify high-risk patients who would otherwise go untreated by human experts, which results in fewer patient deaths.

🔎 Elements To Consider: It is also possible that since the AI system was capable of differentiating between high-risk patients and low-risk patients, physicians were able to pay more attention to the people in the most need. If so, this use case is a reminder of why higher quality healthcare saves lives.

📚 Learn More: Nature.

Stat of the Week 📊 

The World’s Highest Observatory Comes Online After 26 Years

5,560 meters

⌛ The Seven Second Summary: The University of Tokyo Atacama Observatory (TAO) is now officially opened after 26 years of planning and construction. Scientific observations are set to begin in 2025.

🔬 How It Was Done:

• The TAO project began in 1998 and involved years of planning and collaboration across various academic, political and indigenous groups.

• The telescope’s high altitude was chosen to minimize the impact of moisture in the atmosphere on infrared observations.

• TAO’s altitude brings unrivaled clarity and sensitivity for mid-infrared wavelengths, making it the only ground-based telescope cable of clearly viewing this range

🧮 Key Results: The 6.5-meter telescope sits 5,640 meters high (18,500 feet) on Mount Chajnantor in the Atacama Desert in Chile. This marks the highest elevation or altitude among all telescopes currently in operation in the world.

💡 Why This May Matter: The famous James Webb Space Telescope (JWST) is, like TAO, an infrared telescope. While JWST has an even clearer view of the cosmos since it sits in space, an advantage TAO has over JWST is a wider field of the cosmos on Earth. This will allow researchers to survey more extensive swaths of the sky and study multiple, large-scale celestial targets at once.

🔎 Elements To Consider: The telescope sits at such a high altitude that on-site personnel may need oxygen tanks to support their breathing and ensure they can safely work amongst such thin air.

📚 Learn More: University of Tokyo. TAO Project.

AI x Science 🤖

AI Language Models To Judge & Evaluate Other Models

Last month, we highlighted a research paper to describe how the future development of Large Language Model (LLM) reasoning appears to be morphing into a type of tree search and evaluation problem. Meaning, once an LLM receives a specific question or task, it searches through a near limitless amount permutations of responses it can give in a language, then evaluates and iteratively refines its answer until it is rewarded by providing the correct answer. This process autonomously repeats itself several billions of times over, and eventually, the model reinforces the reasoning skills its reward system is trying to teach it.

A key component to reinforce good reasoning skills in an LLM is the evaluation mechanism used to judge the quality of its responses. Since the researchers developing these AI models want to simulate reasoning tasks, they need to be able to design autonomous systems to assess the quality of an LLM’s response at least as well as a human expert. In sensitive domains, like medical diagnostics, an autonomous evaluation system may need to be even more reliable than medical experts in order to minimize the opportunities for poor reasoning skills to creep into the model and cause undesirable patient outcomes.

As we mentioned last month, Stanford’s 2024 AI Index Report indicates AI systems are approaching non-expert level human performance in various tasks, including the ability to assess the quality of responses of other LLMs. While this is a promising development over the last 2 years, AI systems are still far from reliable enough to manage complex or multi-step tasks as well as humans, let alone human experts in specific domains. Moreover, the best LLM evaluation models are supported by proprietary systems like Chat GPT-4, which presents complications with transparency, affordability and configurable controls and criteria for many researchers.

To advance the quality of open-source evaluation models, a team of researchers from KAIST AI and other major universities released a new system called Prometheus 2. The base models for this system are from Mistral-7B and Mixtral8x7B, which is significant because it probably means this system already has room to improve by simply switching one its base models to the recently released Llama-3 model. Nevertheless, in tests on several benchmark datasets, Prometheus 2 has already outperformed other open-source evaluator LLMs in its ability to correlate responses with human and Chat GPT-4 judgments.

While neither a Chat GPT-4-level judge, nor a human-level judge are sufficient evaluation systems for developing a super-intelligent AI model, the Prometheus 2 news are the sorts of incremental progress the field is making towards eventually solve machine reasoning — the ultimate search and evaluation problem. arXiv. Hugging Face. Github.

Our Full AI Index

• Medical AI: Google released Med-PaLM last year to provide the healthcare industry with a suite of AI models for medical use cases. The company has now released a new generation of AI models, based on the foundation of its Gemini AI. When they tested their new suite of Med-Gemini models on 14 different medical benchmarks, Med-Gemini achieved state-of-the-art results on 10 of them. Moreover, Med-Gemini surpassed human experts on medical text summarization, which ties back to the Stanford AI Index Report we mentioned earlier. arXiv.

• Carbon Capture: To combat excessive carbon emissions and climate change, researchers from Georgia Tech and Meta collaborated to develop an OpenDAC database. This database contains reaction data from 8,400 different materials to predict material interactions and energy outputs to help researchers identify the most promising materials for direct air capture. Georgia Tech. ACS Central Science. OpenDAC.

• AI Assistive Technology: Country music star Randy Travis just released his first new song in over a decade, and he used AI to recreate his voice years after a stroke left him unable to speak or sing. This story reminds us of The Whispp App and other assistive technology we highlighted from the Consumer Electronics Show earlier this year. Rolling Stone.

Other Observations 📰

Positive Results From An Alzheimer’s Clinical Trial

Annovis Bio announced positive results from its Phase II/III clinical trial evaluating the drug buntanetap for the treatment of mild to moderate Alzheimer's disease. This 12-week randomized, double-blind, placebo-controlled trial included 353 patients across 54 sites in the United States.

This oral medication works by selectively binding to an element in the mRNA of undesirable proteins to prevent their translation. Through this mechanism, buntanetap was shown to decrease the production of beta-amyloid, tau and other proteins which are hallmark signs of Alzheimer's disease. Furthermore, the benefits of buntanetap appeared to be dose-dependent, as more pronounced effects were observed as patients received higher doses.

The results showed patients who were given buntanetap demonstrated significant improvements in cognition, as measured by the ADAS-Cog 11 scale. This scale ranges from 0 and 70 and consists of 11 tasks to evaluate areas such as memory, language, attention, and other cognitive abilities. The buntanetap patients experienced a 3.3-point improvement on this scale compared to just 0.3 for the placebo group. While this is a meaningful clinical difference in cognitive function, there are still several areas or uncertainty about the long-term efficacy and durability of the drug. A longer-term Phase III trial is still needed to confirm its efficacy and safety with a larger pool of patients. Annovis.

Our Full Science Index

• Clean Energy Investments: Global investments in manufacturing five key clean energy technologies – solar PV, wind, batteries, electrolysers and heat pumps – rose to $200 billion USD in 2023, a 70% increase from 2022. IEA.

• Whale Communication: Researchers from MIT and Project CETI are using machine learning to analyze over 9,000 recordings sperm whale to analyze and understand how the creatures communicate through their complex patterns of clicks. If this project is successful, humans may be able to speak their language someday. NPR. Nature. Project CETI.

• Personalized Cancer Vaccines: The long-awaited Phase III clinical trial for mRNA-4157, the personalized skin cancer vaccine, is now underway and recruiting ~1,100 patients with ‘serious and high-risk’ melanoma. The treatment’s Phase II trial only had 157 patient participants, but reduced the likelihood of death or cancer recurrence after 3 years by 49%. It would be a monumental feat in cancer research if similar results are seen in this subsequent trial. The Guardian. The Lancet. Moderna.

Media of the Week 📸 

Boston Dynamics Has A New Animatronic Robot

Boston Dynamics is well known for many things, including dancing robots. They recently shared a video of their latest dog bot and it has all the moves you would expect from man’s best friend.

The Beautiful Complexity of Leaf Veins

Fossil evidence suggests that around 340 million years ago, leaf veins had a simple branching pattern. However, about 23 million years later, leaves evolved more complex vein networks, possibly in response to insect herbivores and drought conditions. This complexity improved the resiliency of leaves, and led to some the intricate beauty we can now appreciate today. The Royal Society.

JWST Views The Horsehead Nebula in Unprecedented Detail

The Horsehead Nebula is about 1,300 light-years away from Earth and astronomers have been studying it for decades. It’s amazing to see the different visual details between telescopes, and this side-by-side demonstrates why the detail of the James Webb Space Telescope is unmatched. NASA JPL.

This Week In The Cosmos 🪐

May 11: An Earthshine Night. 

This is when the unlit part of the Moon becomes visible.