When was the last time you asked someone what an email was? Or what it meant to Google something? Nearly 30 years ago, when the internet exploded onto the scene, these were common questions. Now, these terms are ubiquitous regardless of age or generation.
We’re seeing the same thing happen today with AI, generative AI, and large language models or LLMs. Aside from the meteoric advancement of LLMs making us change our vocabulary, people often assume every AI conversation is about LLMs. Everything that came before LLMs is known as “traditional” AI. Traditional AI technologies, like machine learning (ML) and neural networks, are proven, pragmatic, and still the AI workhorses of healthcare.
There’s hesitancy around using AI in healthcare, especially when people equate the general term with generative AI instead of trusted, traditional technologies. That doesn’t have to be the case as this blog post will outline.
Traditional algorithms vs. AI models
Prior to the dawn of AI techniques, most computer programs were created the same way using traditional, rule-based approaches. A smart person would think up an algorithm, implement it in code, compile the code into a program, and input data into the program. The data would flow through the algorithm and generate output. If you wanted better results, you needed to invent a better algorithm.
Machine learning has since reversed that paradigm. For some categories of problems, you can use the exact same algorithm for any purpose by fine tuning it to a particular data set. This is sometimes called a model. To get better results, you need a larger, richer data set.
An example helps illustrate the difference between a traditional algorithm and an AI model. In Rhapsody EMPI, we have implemented a machine learning algorithm to automate data quality tasks. It solves the problem of figuring out whether two records represent the same or different individuals when a certain match, one with no ambiguity, cannot be made.
For example, if two records have exactly the same first and last name, sex, date of birth, and address, we can use pre-programmed logic to determine that they represent the same person (a traditional algorithm does this). But if some fields are the same and some different – especially if the differences are near matches – then a traditional algorithm cannot make the determination. A human, or an ML model like we have in Rhapsody EMPI, trained to make the same decisions as a human, must judge if the records match.
In this example Rhapsody EMPI solves the problem using traditional, probabilistic, and ML-driven matching. Knowing how healthcare organizations must constantly evaluate costs and find ways to accomplish more with the same or fewer resources, Rhapsody EMPI uses multiple matching technologies including an ML feature called Autopilot – a model trained to make the same decisions as a human data steward, engineer, or data scientist. It’s like cloning the organization’s data stewards, engineers, or scientists to automatically resolve data quality issues.
Strengths of AI
AI needs enough data to train a model effectively. Typically, this means hundreds or thousands of records. When data is sparse, ML can’t work. Fortunately, for the Rhapsody EMPI use case, our customers have abundant data and the humans needed to provide correct answers.
The right data elements also need to be present in the requisite medical records. For example, if a valuable field for matching person records happens to be their ZIP code, then that’s extremely important to capture in every record. That doesn’t mean more fields are always better. Sometimes extraneous fields can create a model that overvalues factors that are, in fact, irrelevant. Or, in the case of patients filling out paperwork, more fields mean more time and poorer satisfaction. It’s a balance of having enough of the right data elements in each record.
Before highlighting some of the challenges of traditional AI, let’s review the strengths of this technology. Machine learning, which is considered traditional AI, does not require figuring out an algorithm. You can simply solve a problem based on having a large set of sample data, and healthcare data continues to grow at an incredible rate. ML also gives good, validated results after training by humans. Lastly, and in healthcare most importantly, there are no hallucinations with this traditional AI as there are with generative AI.
Weaknesses of AI
One drawback of using traditional AI is the explainability of ML algorithms. Algorithms trained by ML techniques are usually very difficult for humans to understand. Because the model cannot report out what factors it considers, we don’t know how it generates its output. In other words, ML algorithms have poor explainability.
There are ways to overcome this challenge. Rhapsody ensures all decisions made by our ML algorithm are traceable, and human experts can reverse any decision made if needed. Ultimately, after sufficient training, customers have confidence in the results, rarely needing to reverse decisions yet having peace of mind if something arises.
Training the model by feeding it results, validating its responses, and correcting any inaccuracies, is a technique called supervised learning. After it’s presented with enough examples and the correct, human-provided answers as training, the ML model will learn to accurately judge records and make the right decisions.
A second drawback of AI is that if the model training data sets contain bias, the ML algorithm will capture and automate the bias. ML algorithms need to be carefully validated to ensure they work as intended, free of bias with clear data lineage. They also need to be updated because the underlying forces may change over time. A data set that’s generated today may be qualitatively different from the same data generated two years from now, which means that a model trained on the newer data would behave differently.
Rhapsody EMPI models are trained by the customer’s data quality experts with a consensus needed for each decision – multiple people providing responses and discussing for consensus. They complete the training at the same time with the same guidelines to ensure decision alignment and mitigate bias during the training process.
LLM vs. traditional AI
LLMs are the next evolutionary step for ML. An LLM is an ML algorithm applied to the problem of generating human-like text in response to a prompt. Many of the same machine learning techniques are used in LLMs as with traditional AI, but the inputs, outputs, and training data sets are even larger, enormous. Some LLMs are trained on the entire internet, for example.
LLMs produce text (or “write”); they also produce text in response to a document (essentially, “read”). They excel when there is not one right answer. “Are these two records the same or different” is a prompt that will likely produce poor results. “Generate a summary of this patient chart” is a prompt that achieves good results from an LLM because there are many correct ways to summarize a chart.
That is a crucial difference between traditional AI and LLMs. Traditional AI, like what we use in Rhapsody EMPI with Autopilot, is proven and trusted when there is one right answer. When talking about matching a person’s record that answer will be definitely the same, definitely different, or unclear.
LLMs, though, are like students taking an essay test where guessing may provide the right answer, which is better than leaving it blank. They will produce a reasonable-sounding, confident answer even if it’s wrong. Hallucinations with this type of technology have been written about and discussed at length. Some AI experts claim such made-up-answers are unavoidable and inevitable, and, even with efforts to mitigate or reduce them, we may never get to zero hallucinations. While this hallucination problem may improve over time, for now that limits the ways LLMs are useful in healthcare.
Conclusion
The hype around LLMs may make AI seem like something new, but in fact it is a mature discipline of computer science that is having a renaissance. Some facets of AI have been present throughout its history:
- For certain types of problems, AI produces impressive results. It’s important to use the right techniques to match the problem.
- Larger, richer data sets drive better results. Improving algorithms is no longer the right place to focus. Improving data quality, quantity, completeness, and richness is.
- Bias, explainability, hallucinations, and model drift are topics of concern and areas to address when dealing with AI, especially in healthcare.
No LLMs were used or harmed in the production of this blog post.
