As part of an effort to take advantage of recent advances in artificial intelligence, in particular in large language models (LLMs - like ChatGPT), I’ve been experimenting with usage of AgentGPT for the purpose of researching and developing novel mechanisms of detecting cardiac arrest or other health emergencies. The AgentGPT interface is user-friendly and doesn’t require any software skills to use. It works by chaining language models in order to execute on the task, including web searches and summarization of research publications.
If you haven’t experimented with AgentGPT, you can try it here. Note that adding a GPT-4-enabled API key will be necessary in order to get robust responses to complex tasks. I tried it without adding my API key, and while the response was interesting, it was far from complete.
AgentGPT as a 'Cardiac Arrest Detection Scientist'
One of the agents I created I called “Cardiac Arrest Detection Scientist”, and I gave it the task of “Summarize the latest photoplethysmography sensor research and come up with novel theories and strategies to accurately detect cardiac arrest on a wrist or finger device”. Note that "photoplethysmography" refers to light-based blood volume sensors that are standard in one form or another in fitness tracking devices.
I deployed my agent and it got to work. It broke the task down into subtasks, and I was impressed by its identification of those tasks, as they indicated it understood what I was asking and had an intelligent strategy. You can see the interface and some of those tasks in the image below:
Some example subtasks:
Task 1: Perform a comprehensive review and analysis of recent photoplethysmography sensor research publications, patents, and experiments, focusing specifically on their application in detecting cardiac arrest passively in wrist or finger devices.
Task 2: Identify key limitations and opportunities in existing research and technology, and generate novel theories and strategies to improve the accuracy of passive cardiac arrest detection using photoplethysmography sensors.
Complexity and specificity of tasks:
The tasks get more and more detailed and advanced as they go along. For example:
Task 27: Compare the performance of deep learning models with traditional PPG processing techniques: Conduct a comprehensive comparison between the developed deep learning models (CNNs and LSTMs) and traditional PPG processing techniques for noise reduction and motion artifact removal. Measure the improvements made by deep learning models in terms of accuracy, sensitivity, and specificity for cardiac arrest detection in wrist or finger devices. Identify the most effective combination of techniques for optimal performance in real-world applications.
What tasks does it do?
It will do what it can to perform these tasks, for instance it can do planning, estimating, theorizing, brainstorming and research, but not yet development or experimentation. What it can do is included in the output, the rest is described for the user to do. Other GPT and programming assistance tools can then be used to execute the subtasks as defined by the agent, though this would take time to implement depending on one's familiarity with the subtasks.
What I learned from the output from the Cardiac Arrest Detection Scientist, after 54 pages of output, was that there is no sure way to accurately detect cardiac arrest with PPG, there may be some methods that are as much as 90% accurate, and it identified what factors would be involved in those methods.
AgentGPT as a Death Detection Device Researcher
Next, I made an agent that researched death detection devices for me. The task I gave it was to "Survey the existing devices that people can wear that will quickly and accurately detect that they have died. Provide citations."
Again, its generated subtasks indicated it understood what I wanted from it, and it got to work on the task. It converged on the known issues (e.g., accuracy, battery life, regulation) within a few minutes, and gave me some leads on devices and techniques I wasn't yet aware of.
Some example subtasks of this second agent task:
Task 6: Explore any potential uses of AI algorithms or machine learning techniques in combination with vital sign monitoring devices to more accurately and quickly detect death, and provide any scientific or medical studies that support this approach.
Task 8: Investigate additional case studies or recent advancements in the development of wearable devices that can detect death or life-threatening situations, focusing on incorporating AI or machine learning algorithms, their accuracy, and validity.
Task 12: Investigate any recent advancements or research related to AI and machine learning algorithms in the field of wearable devices for detecting death or life-threatening situations, and assess their potential for implementation in the existing Medical Guardian, LifeFone, GreatCall Lively, and Apple Watch devices.
Outcome of second task
Unfortunately, the output and citations included many systems in the form chest straps and vests, given that I didn't specify to only include devices or systems one would easily wear day to day. I also didn't specify in the task that it had to detect death passively, so the output also included classic life alert solutions, i.e., the "I've fallen, and I can't get up" types of devices.
Summary: AgentGPT will only get better at this kind of execution, and will be helpful in ongoing research efforts in cryonics monitoring. Our plan is to automate inquiries and summaries from various LLM tools in order to accelerate and improve cryonics monitoring research. AI is almost at the point where my expertise in this subject can be encoded/accounted for in various machine intelligences, meaning large parts of research and development can be transferred to machines.