What's New in Surgical AI: 01/09/23
Vol 8: A deep dive into chatGPT. Can it give you research superpowers?
Welcome welcome welcome to all of our new friends. You’ll notice a nifty button here:
If you know one other person who wants to be at the intersection of surgery and artificial intelligence, but doesn’t have time to keep up with the latest tweetstorms, you can impress them by sending them this lovely newsletter.
And if you landed here as a non-subscriber, you might want to stay informed:
Now, onto the main event…
Table of Contents
🔬What is chatGPT in the first place
🏎️ 6 ways YOU (and your lab) can use it to speed up the writing process (and one trap to avoid)
💬 ChatGPT - What Busy Clinicians and Clinical Researchers Need to Know
We recently gained a lot of new subscribers (welcome!). As a result, we thought we’d spend the next few weeks reviewing the frantic A.I. hype cycle to distill out actionable insights for you.
Lets start with chatGPT.
Five weeks ago, OpenAI released a chat-based user interface for their latest artificial intelligence model called ChatGPT. OpenAI is a company (valued at $29B…though they call themselves a “research organization”) that builds models for artificial intelligence applications. They’re very well known in the AI community, but made broader headlines last year when their models could generate “AI-Art” - in fact, some of these A.I. Art pieces have won art competitions. We’ve discussed these models and their implications in previous newsletters. Their newest endeavor - chatGPT, is a chatbot which follows commands and responds with striking intelligence, accuracy and similarity to a human:
What ChatGPT is: a powerful and responsive AI model that anyone can use
ChatGPT is a natural language processing (NLP) interface that uses machine learning (reinforcement learning with human feedback) to generate human-like text based upon a large dataset (the entire available internet through 2021). chatGPT is one example from a family of AI models called “large language models”, which are neural networks containing billions of parameters trained on a massive text dataset. At a high level of abstraction, each parameter represents a piece of information that the model can use to make predictions or decisions. chatGPT has 175 billion parameters, requiring massive computational power to generate and run on a daily basis. These models input a series of text and output the next series of letters or words based on the context of the words that come before it. Models can interpret punctuation, sentiment, and even remember (to variable extent) earlier parts of the conversation.
Users can pose a wide range of topics and questions to chatGPT and receive lifelike, human-quality (ok, internet human-quality) responses.
What makes chatGPT so great?
It is so easy to use.
Here’s an example of asking it to brainstorm some vacation destinations
Or (Dhiraj) getting an email to send to his wedding planner:
It responds to feedback well. Like really well.
That email was pretty formulaic, stiff, and overly-formal. So, I asked chatGPT to make it more casual:
These are just a taste of what chatGPT can do. Much more below.
ChatGPT is not: a replacement for humans
chatGPT is not a perfect replica of human language and thinking. It is not capable of independently creating new ideas or concepts, and it may sometimes produce nonsensical or "hallucinated" responses. This is because chatGPT, like all machine learning models, is limited by the data it was trained on and the algorithms used to build it.
It is also not capable of understanding or processing images, videos, or other non-textual data, and it may struggle to understand or respond to certain types of questions or prompts.
Those that say these types of models have no place in our workflows, or is just a fad, are probably wrong. Similarly, those who say artificial intelligence that gains broad “general knowledge” (known as Artificial General Intelligence, or AGI) will displace radiologists, artists, etc., are also likely disillusioned. Technology has the unique ability to disappoint both the optimists and the pessimists.
Chat is the tip of the iceberg: APIs and custom models lie below the waterline
Although the chat user interface got the most “buzz” because of its human-like output, the real power behind GPT and other large language models is available to everyone for a small charge (some are free, depending on the model and implementation) through programmatic access.
Using the OpenAI application programming interface (API), a software application can pass text to the model and receive output. There are myriad use cases for: medical education, streamlining administrative tasks (call schedules, anyone?), content generation, and even clinical workflow improvement.
As another example, we created a webapp (www.dotphrAIse.com) to streamline clinical documentation workflow and bring the entire medical staff (physician, nurse, billing, coding) together by applying generative capabilities of OpenAI to unstructured medical text. Take the brief notes from an interview, or even just the assessment and plan, and drop them into a dotphrAIse. Then - voila - a clinic note, letter to referring doc, after visit summary customized to the patient, CPT codes, MCC checkboxes for inpatients, custom templates and more.
6 Ways You Can Gain Research Superpowers with chatGPT (and one trap to avoid)
ChatGPT is great for creating “boilerplate” content. It reduces time because it takes you from 0 to 80% complete within seconds. One thing chatGPT does NOT do well is literature searching in the chat interface. It will, 100% of the time, come up with a plausible sounding paper in a real journal with some real authors, that simply does not exist. If you want to turbocharge literature reviews, keep reading …
Here’s a list of examples of how you can use chatGPT today to solve some of your research headaches.
Example 1- Outlining Research Project Steps
First time doing a literature review? Or performing bench-work? Let chatGPT be your guide.
Prompt: “[Info about the project]…I would like to create an outline of how to go about this project, from initial data quality control and cleaning, through analysis, through data visualization and publication. Please outline this project, step by step, be very thorough and provide explicit techniques, steps, softwares or approaches whenever possible.”
Example 2: Make grant writing go faster - have chatGPT come up with a first draft of your milestone table.
Prompt: “Simulate an NIH reviewer assessing a grant. The grant is titled "Building a functioning pipeline for ingesting, storing and analyzing large volumes of surgical video safely, efficiently and at scale." Create a table showing key milestones for the grant (24 month period). Be specific and provide examples in each milestone. Be creative and ambitious. Format as a table.”
Example 3- Take the pain out of consolidating data from multiple sources.
Are you working on analyzing data from a on a multi-center project? Dealing with inconsistencies in the data submission? Stop pulling your hair out and let chatGPT write code to fix that.
Prompt: “I just got 7 different excel documents with a standardized data table. Show me sample code of how to consolidate all of them. There may be some documents with a slightly different structure (example: a column was added). I still want to consolidate all the documents, but would want to add a new column at that location where the "rogue" column is, so that all the other columns match up still.”
Cool, but what if I don’t have any coding experience? how do I even get started?
Example 4 - Ask chatGPT how to execute the code it wrote
Example 5: Cleaning up data
Ask chatGPT to remove prefixes, remove duplicates, screen for outliers, and just about anything else you can imagine within the world of data analysis.
Prompt: “Some of the columns in my data begin with 'Clinical_Symptoms' or 'Endocrine_Status' or other column prefixes before the actual information which should be the column heading. Write me a piece of code which removes a prefix from the data which I specify”
Example 6: Rewrite (or writing) titles
Prompt: Offer 5 titles for a journal submission for this abstract, and 5 titles for a blog post
Prompt: “Offer 5 titles for this journal submission and 5 titles for a blog post” [Copy/Paste abstract]
What about Literature Reviews using GPT (don’t do it)?
Good old Dr. Google wins again…
Imagine writing a literature review and typing “rate of subarachnoid hemorrhage pediatric patients united states” and receiving a formatted sentence in your text with the output. The underlying capabilities exist, but the models have not yet been customized (see the opportunity?). This is NOT a current capability of either chatGPT (do not trust citations from chat GPT). There are better implementations of LLM’s, such as elicit.org, a LLM that is only trained to retrieve, summarize and find connections between real papers on Pubmed. But it struggled with my simple query, as did Pubmed search. But I found two good recent articles (thankfully, written in 2021 and 2022) from google and google scholar.
No, I am not telling you to use google to perform a systematic literature review for many reasons (reproducibility, SEO page ranking, unclear underlying data universe, possibility of misinformation, and others). New, critical tools should be explicitly brought into the literature review workflow once they can meet the same standards to which we hold ourselves.
For the chatGPT “veterans”: What did we miss in the examples above? Drop a note in the comments! Subscribe and share so you don’t miss out on future tips to squeeze the most juice out of these tools.
Feeling inspired? Drop us a line and let us know what you liked.
Like all surgeons, we are always looking to get better. Send us your M&M style roastings or favorable Press-Gainey ratings by email at ctrl.alt.operate@gmail.com