ON ChatGPT's AI and its gamechanging impact for Healthcare

I was fortunate to be awake on twitter when Sam Altman recently released this tweet:

At first this sounded like one more of those "AI Hypes" that lead to nowhere when applied to the real world, but, right afterwards, I started to see this:

This is almost impossible to see: ALL these hugely respected #tech gurus pointing out in the same direction: a tipping point for AI and a massive change for society.

What is ChatGPT (for business\ non-tech people)?

It is basically an AI model that has INGESTED billions of documents\webpages on millions of topics, and that has learned a REPRESENTATION of the world's knowledge. Put it simple: Give it a large corpus of text from any domain, and its Machine Learning will read, understand and be able to talk about it like an expert in the space, within minutes.

GPT: Short for “Generative Pre-Trained Transformer,” is a breakthrough AI technique for language understanding (LLM) that uses 175 billion parameters and costs $5-10Million to train.

The magical part is that you can chat with this "robot" asking it any type of question and obtain a reasonable answer within seconds. While answering, these incredible models also Generate new hipothesis and seem to emulate what a (super knowledgeable) human would answer.

Although it does not demonstrate pure intelligence yet (strong AI\ AGI), that is not the point.

The point is that it is the best WORKING proof of what "WEAK AI" can already bring:

💡 Doing complex tasks that humans once did, 1000000X faster. And that is, by itself, a huge gamechanger for any business!

Why is this a leapfrog jump in Healthcare? 

Having dedicated a great part of my healthcare career leveraging Data & AI to help save lives, I understood how critical is TIME to action inside some healthcare settings.

It is unimaginable the amount of handcrafted work that #doctors and #nurses have today for compiling all relevant information about a clinical case, in each part of the clinical pathway. The EMR (Electronic Medical Record) remains basically a digital form of storing patient data, with no intelligence running in the backstage. Hospitals & GPs still have hundreds of disconnected systems and billions of text notes, leaving for clinicians the tedious and manual process of correlating, compiling, and analysing the information spread out in all those sources to produce an inference about the patient, and decide the next clinical step.

Now imagine this could be done in seconds, with the help of ChatGPT (clinically tuned).

How ChatGPT already understands and resonates over clinical domain

Imagine #doctors being able to have a (#clinical) robot, by their side, to simply ask questions about their clinical cases. For instance, checking if a patient's severe symptom might be an adverse reaction to the combination of the dozens of medications that he might be taking. Just by asking this robot to ingest all clinical notes and immediately crawl the clinical literature in milliseconds and obtain possible answers\relations\hipotesis in this specific case, the treatment process would speed up massively.

To get to the point, have a look at my interaction with ChatGPT, testing if it could resonate over a clinical situation:

Some reactions that immediately come to mind:

💡 How was it able to understand that CRP and WBC levels were "elevated", without that fact being explicitly stated in the text ?!

💡 How was it able to maintain context throughout the entire conversation, when I referred fever in a further separate sentence, being able to correlate that with levels of CRP and WBC previously stated, and produce an inference of a possible infection case ?!

WHY this matters?

 80% of Healthcare Data is UNSTRUCTURED.

JUST THINK ABOUT THIS:

A large hospital center has Hundreds of thousands of clinical notes written, PER YEAR! Having seen those notes I can state how valuable these can be for a model that is able to understand all relationships among concepts (embeddings). The clinical notes of a patient's episode stored in the EMR often have ALL the clinical, family, and social history of that patient, described with great detail.

• Now multiply this Data by all hospitals (public & private), GP centers, health insurance companies, pharma companies,.. Etc
• This represents decades of clinical knowledge stored in petabytes of Data, completely underused although its breathtaking potential. 

We will soon see a clinical version of ChatGPT that will be able to crawl over these billions of texts and automatically understand, correlate, and generate new clinical information about each patient in a flash of a second.

The Digital Patient

Back in the days (2016), I was fortunate to lead a huge research project that leveraged AI (NLP) to retrieve actionable insights from hundreds of thousands of clinical notes. This initiative was then called the Digital Patient

The Digital Patient was put live, in production, with the capability of summarizing any patient's history in a flash of a second (medication, allergies, surgeries, lab results..). This degree of healthcare innovation was rare to see even at a global level, mostly because of the huge Technological & Data challenge it represented.

This challenge got me introduced to how powerful #NLP models can be. But, at the same time, how hard it was to train these models to understand the clinical domain: back then, thousands of hours from a specialized clinical team were needed in order to "teach" the system about each clinical concept and the thousands of clinical idiosyncrasies and acronyms.

But models like ChatGPT seem to bring a much more scalable, unsupervised and powerfull approach to this problem. As Harvard\MIT researchers have proved, it will be only a matter of time until these Large Language Models help to decipher clinical notes. Being able to automatically infer and generate a digital representation of a patient just by reading these notes, will soon become the gold standard for a Digital Patient vision.

Data Platforms - The importance of building the right foundations upfront

It is now becoming clear that the right investment is no longer in building data science and AI research teams. The best AI models will naturally come from powerhouses like OpenAI, who can invest millions of dollars in computing power to train these huge models, and make those available to the community, as ChatGPT's example ("Models as a Service").

The real differentiation for companies and other healthcare orgs, will soon come from their proprietary Data Platforms (DP): how well these were conceived and are able to capture and correlate the most important asset of any business, as Chamath Palihapitiya signals here.

Building the right Data Platform means years of upfront investment, but will soon become the primary goal for large organizations. These foundations are key to be futurly used by novel AI models, and allow companies to Compete in the Age of AI. The best framework I've seen so far to ignite this transformation is BCG's Data & Digital Platform one.

In Healthcare, AI models like LLMs (ChatGPT) will soon evolve to absorb all Health Information, and become multi-modal - interpreting text images, sound and more. These means that, with the right DP in place, models will able to crawl along these formats and rapidly infer patterns in a very precise way (e.g: think about a tumour multimodal analysis crossing a suspected long nodule in a CT Scan, with the clinical history of family cancer in text notes).

Autonomous Healthcare Agents will boost clinician's work like never before

For those who want to foresee the future, Pedro Domingo's (2015!) article on the WSJ couldn't be more timmely applied, and is well worth a read 🚀

Soon, these powerfull models will gain autonomy and transform into intelligent assistants that will become autonomous agents of our lives, being able to, for instance, negotiate our car insurance with our insurance company on behalf of ourselves. Or, as Sam Altman also signals: these will able to "go out and discover new knowledge for ourselves"

Similarly, for the medical community, these assistants will be able to crawl all of the internet's trusted sources like PubMed to find, and filter the correct information needed for a particular research study or clinical treatment! This is a massive stepforward to the super manual and handcrafted work that currently doctors and researchers face to explore the literature, spending literally years of life crawling trough thousands of papers, something that a LLM robot could easily do at lighting speed.

With ChatGPT released just a month ago, there are already several initiatives heading this way, with the initial launch of Stanford's PubMedGPT just 15 days later (15/12), and, right afterwards, GoogleAI and DeepMind releasing an even better model that already achieves 67.6% accuracy on medical questions and answers from the US Medical Licensing Exam (USMLE) (a 17% improvement in just a few days!!):

Dangers & Trustworthiness

Current LLMs (Large Language Models) such as GPT-3 (ChatGPT) still suffer from flaws that make them unpredictable in tasks that require common sense, logic, planning, reasoning, being notoriously known for hallucinating responses, generating text that is coherent but factually false.

But we most be aware that GPT4 is already underway (100 trillion parameters! Vs 175B of GPT-3) and promises to bring a massive improvement to most of these problems. Also, Google recently released a ChatGPT-like ChatBot for Healthcare that already accounts for potential harm, and bias of misleading health information.

Summary

The pace of change on this new AI breakthrough is unstoppable and unprecedented. Healthcare, being among the most primitive areas when it comes to effectively applying AI will possibly be the one that will most benefit from it. With 80% of its clinicians activity being currently stored in unstructured formats, Large Language Models will be able to easily digest, interpret and learn the meaning of petabytes of clinical notes without any problem.

In an environment where Time to Action is key and so much manual work is still being done, my guess is that a clinical version of ChatGPT will soon become a doctor's super AI assistant, helping him to obtain actionable insights for treating his patients that would, otherwise, take days to obtain.