Introduction

Let's face it, the rapid growth of ChatGPT has ignited worldwide interest in Large Language Models (LLMs) and their potential applications across various industries. People on both sides of the aisle raise their voices in discussions about the challenges and opportunities that the (not so) new technology holds. One sector that has been particularly captivated by the possibilities of LLMs is healthcare. For a good reason, not because ChatGPT successfully 'passed' the US Medical Licensing exam but rather because healthcare has an enormous corpus of structured and unstructured data. 

Structured data is key to pre-training and fine-tuning models. 

Unstructured data is an excellent proxy to measure the relative opportunity for LLMs.

Unfortunately, it's also the sector where we have seen the highest amount of fake promises and snake oil stories since the introduction of ChatGPT. A recent study found that ChatGPT gave 'better' recommendations to patients than actual doctors. Unfortunately, most AI fanatics were quick to dismiss the fact that the researchers used Reddit and not, you know, a natural medical setting for their assessment. 

We can all agree that ChatGPT won't replace doctors anytime soon. Furthermore, LLMs won't be very visible nor save a lot of lives directly. But don't let that distract you; the impact of LLMs on healthcare will be omnipresent and durable. LLMs will disrupt healthcare more in the next 10 years than any other technology. Disruption comes not a minute too late, and let me tell you why:

1. Since 2000, the price of medical care has increased by 115.1%. In contrast, prices for all consumer goods and services rose by 78.2% in the same period. 
2. Between 1999 and 2014, labour productivity in the US increased by only 6% in healthcare, compared to 18% in other service industries
   

In conclusion, prices in healthcare rose more than in any other industry, and people who work in healthcare have not increased in productivity like most other traditional industries. This creates a gap in value for money (and a need for disruption). Add to the mix that more people than ever need medical care, and you get the picture.

Traditionally when a specific sector is underperforming other sectors, it's ripe for disruption. But healthcare is complicated, and so for years, we have seen amazing innovations that have massively improved the lives of patients affected but not much innovation that has a wide impact on medical–the industrial complex. But NLP, LLMs & generative AI might just be what we need. LLMs promise a huge increase in productivity and quality if implemented correctly. They have the potential to address the challenges mentioned earlier and bridge the gap in value for money in healthcare. Think:

• Differential diagnosis generation
• Summaries of patient history 
• Detection of adverse events (drug interaction or complications)
• EHR data normalisation, extraction, structuring and insertion
• Automatic enrollment in treatments and clinical trials
• Personalised health coaching and remote follow-ups
• Continuous education and optimising care coordination
• Provide suggestions for more complete and accurate records
• Provide suggestions for compliance with regulatory requirements
• …
  

Today, it shouldn't be the goal for LLMs to replace doctors, and neither will this be the case in the future. However, we should speed up the implementation of LLMs in healthcare. In this blog, we will explore how LLMs, trained on healthcare-specific data, can significantly enhance the daily workflow of medical professionals and pave the way for the rise of the 'Augmented Medical Professional'. 

So, how do we get there?

Step 1: Develop Healthcare-Specific LLMs

GPT-4 excels at predicting the next best word with the highest probability. It generates expressive sentences and follows linguistic rules. However, general-purpose LLMs like ChatGPT, despite their impressive capabilities, can cause occasional inaccuracies or inconsistencies, also known as hallucinations. In the medical domain, such errors can have serious consequences. These hallucinations aren't bad when brainstorming a new LinkedIn post, but when developing a treatment plan for a patient, a quick visit to the radiology department for a colonoscopy to test your blood pressure is far from ideal.

For reference, one emergency room doctor anonymised their patient's data and fed it to ChatGPT to find a 50% success rate in suggesting the (what he thought to be) pathology present in the patient. This 50 % hit rate is far from acceptable in our context. 

Luckily, for many applications suggested above, we can get closer to 100 %. And remember, we don't need 100 % as long as you use it as a tool and don't see it as a replacement. 

We need to use healthcare-specific LLMs that are developed using the same model structure as GPT-4 but trained on medical data, offering a more tailored solution for medical professionals. One of these models is Google's Med-PaLM 2, which was the first LLM to perform at an "expert" test-taker level performance on the MedQA dataset of US Medical Licensing Examination (USMLE)-style questions, reaching 85%+ accuracy (there are also a lot of Open-source alternatives available)

Step 2: Fine-Tune LLMs with Organization-Specific Data

Once you've selected your foundational model (preferably healthcare specific), you can connect it to your database (e.g. EHR). Connecting your entire healthcare database to a model like ChatGPT may seem to be a comprehensive solution, but it has some limitations:

1. You must be comfortable sharing your data with a third-party vendor and adhere to data protection regulations (big nono in healthcare).
2. It is still challenging (both for humans and LLM models) to determine the source of the data (coming from the pre-trained model or from your database) and whether it is up-to-date and reliable.
3. Fine-tuning a model with your data requires good data, structure and a secure environment.

For all of these challenges, there exist time-tested solutions. Working with a trusted partner can help reduce costs and de-risk your investments. We can help.

Step 3: Integrate LLMs into Existing Systems

To unlock the full potential of LLMs in healthcare, your newly fine-tuned LLM should be accessible in applications and tools without being immersive in the traditional workflow. A well-integrated system can automatically process medical records and treatment plans, promote patient engagement and adherence and provide medical professionals with augmented context for better patient care. And all this at a lower cost. LLM can halt the trend of an 'inflated inflation' in healthcare.

Step 4 - General Medical AI (GMAI)

GMAT is the holy grail of AI in healthcare. Imagine a model capable of carrying out a diverse set of tasks and owning a diverse set of skills like interpreting medical data from different sources, including data from non-text sources, interpreting graphs or CT scans. Stay tuned for my next blog post, which is 100% dedicated to this topic.

So, what can you do?

I want to end this blog post with some borrowed insights from my friend, Thomas Beuls (Ooho). Introducing new technology in healthcare is complex. Regional systems require a tailored approach because of regulatory complexity, and people who take the risk are seldom rewarded. LLM technology is still early, but please be open-minded. Create isolated environments where you can test without creating security risks or decreasing the quality of care for your patients. 

If you want to be part of the leaders who help make healthcare affordable, then we're more than happy to guide you through developing and implementing your AI projects.

‍