Generative AI in Healthcare: The Definitive Guide

Ricky Sahu
by Ricky Sahu


The emergence of generative AI is delivering a new era of possibilities for healthcare. With the ability to create original content based on learned data, generative AI can redefine the way industries function. The healthcare sector is now next in line to experience these transformative changes with the integration of generative AI. This will lead to breakthroughs in personalized care, administrative efficiency, drug development, and far more in the industry. This guide offers a comprehensive understanding of generative AI technologies, their application in different healthcare markets, and potential use cases.

Various Generative AI Modalities in Healthcare

One way to think of generative AI in healthcare is based on the type of sequential data used to train the transformers. Five primary categories exist today. Language, Medical, Audio, Image, and Chemical models. Audio, image, and chemical models are highly specialized today. Language and Medical models can have broader applicability. But they are all based on making the input data sequential and feeding it into a specialized model called a transformer.

The Transformer

Generative AI has been made possible by the development of the transformer model, which has its roots in sequence to sequence models such as RNNs (recurrent neural networks) and LSTM (long short term memory) neural networks. The transformer model, introduced in the 2017 paper "Attention Is All You Need" by Vaswani et al., is a type of neural network architecture designed to predict the next token in sequential data, making it a powerful tool for language processing, audio analysis, and image interpretation tasks among other sequential series like patient timelines.

Transformers are a massive improvement over prior sequence models due to their ability to process that data more efficiently. The attention mechanism introduced in transformers allows the model to focus on relevant parts of the input sequence simultaneously, resulting in more accurate outputs with more efficient compute resources. By training on large datasets, transformers can generate original content based on learned data. One of the most famous transformer models is ChatGPT in which the GPT stands for Generative Pre-trained Transformer. The underlying transformer technology (like that in ChatGPT) is beginning to be incorporated into healthcare by enabling the creation of domain specific large language and domain native medical models, as well as audio, image, and chemistry models.

Large Language Models (LLM)

Large language models, such as the ones used in ChatGPT, are prime examples of transformers trained on comprehensive text datasets. These models can be fine-tuned on healthcare-specific data to generate insights and recommendations for healthcare providers. By simulating human language understanding and generation capabilities, these models can assist in patient communication, medical research, and administrative tasks. Usually medical research papers and doctor notes are incorporated in these healthcare specific LLMs. As the use of these models becomes more prevalent in healthcare and research meets commercialization, the quality and accuracy of patient care is likely to improve.

LLM Solutions: This is the most active space in generative AI for healthcare with many solutions including Med-PaLM, Notable Health, Hippocratic AI, and more. Some solutions in this space have trained their own foundation model from scratch, others are fine-tuning, but most are building on top of standard LLM models like OpenAI’s GPT4, Google’s LaMDA / PaLM2, and Anthropic’s Claude.

Large Medical Models (LMM)

These transformer models are specifically trained on medical events and codes such as ICD-10 and CPT codes, transitions of care between different Provider NPIs, and cost data. The models' proficiency with these codes assists in simplifying medical billing, creating accurate disease forecasting models, and improving patient care by generating valuable insights into symptoms, diagnoses, treatments, among other areas. LMMs are trained on a large dataset of medical records, claims, and other healthcare information including demographics, providers, and prices. This information is often missing from the actual doctor notes or from research papers. Instead they are contained in standardized healthcare feeds like FHIR, HL7v2, and 837s. An LMM is highly specialized, and its training is focused on medical data, which makes it more accurate and reliable than other models for certain applications and broadens the set of use cases to those which are beyond the human communicated text of healthcare.

LMM Solutions: created the LMM space and we are still the only solution approaching transformers from a native medical standards and terminology lens. We find that LMMs support a broad set of use cases which LLMs cannot like provider network analytics, cost analytics, risk stratification, and population and patient level forecasting. The health tech space is like an iceberg, most people see the tip (providers interacting through natural language with their EHR), but our experience has shown us what lies beneath (the financials, claims, prior auth, networks, formularies, workflows, etc.) and that’s what LMMs are best suited for.

Audio Models

Audio models, which are built by utilizing transformers that have been trained on audio data, have the ability to convert spoken language into text. This is done by analyzing audio sequences and processing them via a transformer architecture into written format. Because of this, they are extremely useful in situations where transcribing doctor-patient conversations, decoding spoken instructions, and transforming speech-based notes into written format are necessary. By doing so, audio models are able to significantly reduce administrative burdens on healthcare providers, allowing them to focus more on providing quality care to their patients. Moreover, audio models have the potential to improve the accuracy and efficiency of healthcare services by minimizing errors that could arise from manual transcription, especially when dealing with complex medical terminologies.

Audio Solutions: Nuance, Abridge, Suki, Amazon Transcribe Medical, are all solutions in the audio to text for healthcare space. Again some of these are built on their own foundation models, others are putting apps on top of commercially available audio models.

Fun fact: the first GenHealth model was actually a fork of an audio transformer model just substituted with medical tokens instead demonstrating the versatility of this technology.

Image Models

Image models, including transformers and stable diffusion tech trained on radiology and pathology images are pivotal in detecting and interpreting anomalies in medical images. Such models can play a crucial role in early detection and monitoring of diseases like cancer.

In addition to transformers, diffusion models are also being incorporated into healthcare. Diffusion models, such as those developed by Stability AI and MidJourney, are a type of generative AI that are used for image generation. These models use a process called the diffusion to generate high-quality images that are often indistinguishable from real images by diffusing noise of random pixels and then using an trained neural network to caption that noise. Then more noise is diffused into the image to come closer to the desired caption. While the application of diffusion models to healthcare is still relatively new, they have the potential to significantly improve the accuracy and efficiency of medical imaging.

Chemistry, DNA, and Protein Models

In the domains of bioinformatics and biomedicine, transformers are employed to decipher sequences of amino acids in proteins, DNA, or RNA. By analyzing these sequences, these models can predict the functionality of unknown proteins or mutations and are progressively being utilized for drug discovery and personalized medicine. Given the generative nature of these models, scientists are now beginning to create entirely new proteins and antibodies and bring them into clinical trials.

The future of DNA/RNA models is rapidly evolving. These models have the ability to create first of a kind synthetic sequences which can be used themselves to produce new proteins via biological processes. However the length of nucleic acid sequences and the high cost of memory in compute will mean it will be some time before DNA models are able to create an entirely new organism. That is the future however.

Multi-modal Models

Generative AI is also being used to build multi-modal models that can integrate language, audio, and image data to generate more comprehensive insights into healthcare data. By training on a range of data types, these models can identify previously unknown relationships between medical data and various health outcomes.

Multi-modal models are still relatively new in the broader industry, but in healthcare, they offer the potential for comprehensive insights into healthcare data by integrating language, medical events, audio, and image data to identify previously unknown relationships between the data and various health outcomes. This is probably the most exciting frontier in AI.

The Impact of Generative AI on Healthcare Markets

Generative AI has ushered in remarkable changes in various healthcare markets.

Health Insurance Plans

For health insurance companies, generative AI is a powerful tool for cost prediction, billing streamlining, plan personalization, and fraud detection. By training on historical claims data, AI models can generate precise predictions about future claims, thereby helping insurers manage their risk-profit balance effectively.

Pharma Industry

Generative AI offers immense potential for the pharmaceutical industry, particularly in drug discovery and clinical trials. AI can assist in creating novel drug molecules and predicting their interactions within the human body. Additionally, in clinical trials, AI proves invaluable in data analysis, patient selection, and monitoring trial progress, thereby accelerating the drug development process.

Healthcare Providers

Healthcare providers, including hospitals and clinics, can utilize generative AI to enhance patient care. AI can predict disease progression, tailor treatment plans, and decrease healthcare professionals' administrative load. Telemedicine services can also benefit from AI, which can assist in patient triage and follow-up care.

Life Insurance

In the life insurance sector, generative AI can aid in predicting lifespan and disease risk, enabling insurers to create more personalized and accurate life insurance policies. AI can also simplify the underwriting process, improving efficiency, and reducing costs.

Application developers

App developers can use generative AI to build chatbots or virtual assistants that can help patients with their medical queries. With the help of natural language processing and machine learning algorithms, these AIs can provide personalized insights to patients, such as helping them schedule appointments, remind them to take their medications, and answer their medical questions. Often generative AI companies make APIs available to app developers and rather than coding every logical decision by hand, applications can incorporate generative AI models to support the long tail use cases.

Generative AI Use Cases in Healthcare

Generative AI offers a myriad of use cases for different parties within the healthcare sector.

Health insurance companies can employ generative AI to develop personalized insurance plans based on individual health data, lifestyle, and genetic factors. AI can also identify fraudulent claims by recognizing patterns indicative of fraud.

In the pharmaceutical industry, generative AI can significantly expedite drug discovery by creating and in silico testing millions of potential molecules, thereby bypassing the time and cost-consuming traditional drug discovery methods. AI can also facilitate patient recruitment for clinical trials by identifying suitable candidates based on their health data.

Healthcare providers can leverage generative AI for disease prediction and treatment. AI can analyze patient health data to predict disease development likelihood and recommend preventive measures. It can also assist in tailoring treatment plans based on a patient's response to different medications.

Life insurance companies can utilize generative AI to predict a person's lifespan based on health data, genetic information, and lifestyle, enabling the creation of more personalized and accurate life insurance policies.

A Glimpse into the Future

As we stand at the accent of a generative AI revolution in healthcare, the possibilities extends far beyond the horizon. What was once confined to the realm of science fiction is rapidly becoming a reality, with the potential to transform (pun intended) every facet of healthcare. From enabling more accurate diagnoses and personalized treatments to driving efficiency in drug discovery and clinical trials, generative AI holds immense promise for the future of healthcare.

In the coming years, we can expect to see generative AI increasingly embedded within healthcare systems. Advanced models could provide more nuanced and comprehensive patient assessments, drawing from a wealth of medical literature and individual health records. They might even predict potential health risks and recommend preventative measures, driving a shift towards more proactive healthcare. Furthermore, generative AI has the potential to democratize access to quality healthcare, especially in remote or underserved areas, by performing tasks that typically require specialist knowledge.

The art of the possible with generative AI in healthcare holds both potential to help people live longer healthier lives and the opportunity to create massive businesses in the space. As the industry develops, it is vital that we navigate this new terrain with care, ensuring that ethical guidelines and regulations evolve in tandem with technological advances. The path to the future may be uncharted, but the promise it holds is clear.