Generative AI vs Classical Statistics Healthcare Analytics

With generative AI based analytic models, there are many open questions in the industry about efficacy performance, and use cases. This space is just starting to be explored. It goes much deeper than what Large Language Models can do for healthcare, instead, we’ll be talking about Large Medical Models.

• What are the underlying reasons why Generative AI models may be better?
• Are Generative AI models like transformers better than classical statistics based healthcare analytic models (like linear models, SVMs, Random Forests, XGBoost)?
• What are the previously untenable challenges that can be addressed by Generative AI based healthcare analytics?

Why may Gen AI be better?

Classical statistics-based healthcare analytic models, such as linear models, SVMs, Random Forests, Bag of Words, and XGBoost, have been widely used in the industry. These models typically rely on a set of observed variables to make predictions or draw conclusions. They require a large number of observations relative to the number of variables to achieve accurate results.

The variable counts include interaction effects between multiple individual variables. So if you have 5 million providers, and 1 million medications, procedures, diagnoses codes, etc., you’ll need 5,000,000,000,000 (5 trillion) observations just to understand relationships between a provider and any intervention. Compound that to 5 or 100 interactions and you have an intractable problem for classical statistics based analytics.

However, generative AI models, like transformers, offer several advantages over classical statistics-based models in healthcare analytics. These models are capable of generating new data points based on existing patterns and learning from a vast amount of data that is not subject to the same multivariate interaction limitations. This ability to generate new data is particularly valuable when the available data is limited or the complexity of interactions are deep such as in healthcare. Generative AI models can fill in the gaps and provide insights that may not be apparent from the available observations using sequence to sequence models which are in many ways auto-regressive.

Furthermore, they are quite explainable, and you can modify various parameters or inspect parts of the neural network to better understand the output. All of these capabilities lend themselves to an ability for generative AI best models to support use cases that have not been planned for. For example, the underlying technology for chat, GPT is a GPT model. The GPT model can be used for a chat bot, for a question answer agent, to write articles like this, and many more things that we haven't thought about yet. Similarly with Generative AI based healthcare analytics, we do not need to build separate models for risk adjustment, another for cost production, one for patient matching, and yet another for provider network design. Instead, all of those use cases can be run off of a single model like ours.

Is Gen AI Better?

Of course this depends on the use case but we are seeing across a broad set of problems Generative AI based models are better than commercially available models built on classical statistics.

For example GenHealth.ai’s model is performing better than the current state of the art at predicting the next year’s cost of a patient based on the prior year’s sequence of healthcare data. Typically classical models are good at predicting within +- 100% of a patient’s next year cost based on the prior year’s data. So if a prediction is $100,000, the actual cost will be anywhere between 0 and $200,000. With GenHealth.ai’s Generative AI approach we are seeing early results with +- 80% error so $20,000 - $180,000 actual outcome for a patient’s future cost when predicting $100k.

One use case is in insurance reserves. Typical health plans hold about $500 in reserve per patient. This could be shrunk to $400 using a model such as the one from GenHealth. That $100 per patient cost savings just from interest at today's 5% interest rates for a health plan with 1 million patients would equate to $5,000,000.

We are working on additional evaluation metrics that will be published to demonstrate how well Generative AI models like ours perform against classical statistic base models for other use cases, than cost.

Generative AI models, such as Large Medical Models, have the potential to revolutionize healthcare analytics. These models offer a significant improvement in performance and detail compared to classical statistics-based models.

For instance, in the field of natural language processing (NLP), the introduction of ChatGPT-like models has brought about a leap in performance and detail compared to classical statistics-based NLP models that were used historically. Similarly, with Large Medical Models, we are witnessing a similar leap in performance and detail compared to classical statistics models. The ability of generative AI models to understand complex patterns and generate new insights can lead to more accurate predictions, personalized treatments, and improved patient outcomes.

What can you do now that you could never do before?

With generative AI models in healthcare analytics, we can now address previously untenable challenges. These models have the potential to:

• Discover hidden patterns and relationships in complex healthcare data.
• Generate synthetic data to augment limited datasets, enabling better analysis and predictions.
• Provide personalized recommendations and treatment plans based on individual patient characteristics.
• Improve diagnostic accuracy by leveraging a broader knowledge base and learning from a wide range of medical cases.

Generative AI models empower healthcare professionals with advanced tools and capabilities that were not possible with classical statistics-based models alone. They pave the way for more accurate, efficient, and personalized healthcare analytics. With our team’s work, we are seeing that this isn't just promise, but it is being realized.