In this article, Volodymyr Andrushchak, Data Science Team Lead at Lemberg Solutions and Ph.D. in Data Science, explores the potential that large language models (LLMs) bring to the healthcare industry.
Over the past few years, the healthcare industry has turned its sights to the vast potential of large language models.
Commonly referred to as AI models, these systems have capabilities to improve healthcare services, reduce workload, and forecast patient outcomes and risk on a local and — even — national level.
Let’s explore why LLMs are the next step in healthcare, what makes them unique, and how they can benefit — and already do benefit — the industry.
The evolution of AI models in healthcare
Healthcare professionals have applied data science for decades, using it to predict patient outcomes and improve the use of electronic health records.
Traditionally, these AI systems relied only on structured data — such as diagnostic codes and vital signs. However, the whole picture was overlooked.
While useful, early models struggled to process another valuable portion of healthcare data, which is called unstructured and includes clinical notes and discharge summaries.
As machine learning evolved, researchers found a way to extract insights from unstructured clinical data using natural language processing. But there were still some limitations.
The early NLP systems had a hard time understanding the context, which led to them misinterpreting things or struggling with abbreviations and misspellings.
The true leap forward came over the past few years.
LLMs are here. What exactly are they?
The development of AI systems that can understand and produce natural language has been advancing since the late 2010s.
This progress gained significant attention with the release of ChatGPT — one of the most high-profile launches — followed by a wave of AI models from other major technology companies.
As these models advanced, their applications expanded across various fields.
What makes LLMs so powerful in healthcare software development is their ability to process valuable, unstructured clinical data. Large language models can capture human language nuances within context.
Based on transformer architecture, they learn from vast amounts of text to recognise patterns and analyse relationships between words.
Volodymyr Andrushchak
Unlike other models, LLMs are pre-trained on a huge amount of data — way more than can be comprehended. This scale is both their strength and their challenge.
To ensure these models perform effectively in healthcare, data science teams often fine-tune them on domain-specific datasets.
For a better understanding of the scale, such a dataset might contain billions of words from clinical notes representing thousands of patients.
However, LLMs have their weaknesses. One of these is hallucination, which happens when a model’s response is made up or doesn’t make any sense.
Another concern is the risk of biases, which could negatively impact patients by leading to missed diagnoses. Finally, one of the most debatable issues is the privacy and security of patient data used in training AI models.
Considering the capabilities and limitations of LLMs, the healthcare industry is exploring how they can serve as effective tools — and the goals for these tools are highly ambitious.
The major ways AI models empower healthcare
As challenges in the healthcare industry increase — including data complexity and the need to prevent healthcare events rather than react to them — technology empowers organizations to discover effective solutions to these issues.
There are various ways in which AI models can be beneficial; this section will focus on two of the most significant aspects.
Supporting healthcare professionals with daily tasks
AI models hold the promise of reaching a level of medical understanding that effectively supports healthcare providers in their decision-making processes.
And they are already demonstrating this potential.
LLMs perform various complex tasks: answering medical questions, summarising clinical texts, analysing patient data and medical images, and even suggesting potential diagnoses and treatment options.
Some of them can be used to answer common questions from patients — and reduce the workload for healthcare professionals.
One notable benchmark of AI models’ progress in medical understanding is their ability to pass medical exams.
For example, Google’s Med-PaLM surpassed the pass mark (>60 per cent) in the U.S. Medical Licensing Examination (USMLE) style questions.
Later, Med-PaLM 2 performed at an “expert” test-taker level performance on the MedQA dataset of USMLE-style questions, reaching over 85 per cent accuracy.
In addition to general medical applications, AI models can also be fine-tuned for specialised domains such as radiation oncology, mental health, and drug discovery.
When trained on domain-specific data and aligned with appropriate clinical frameworks, these models can function as second opinions.
Predicting medical events with the capabilities of AI models
AI models are showing strong potential in predicting medical events and health outcomes — shaping preventive healthcare.
For instance, some models have been developed to predict the likelihood of a patient being readmitted to hospitals.
While the concept of predicting readmission risk is not new, advanced AI models can analyse a vast amount of structured and unstructured data, including medical records of thousands of patients.
Other models are being explored in the field of mental health. By using different types of assessments, these models aim to predict the risks of mental health issues within specific groups of people.
In some cases, these models are tested to identify underlying causes, allowing healthcare providers to recognise a person’s high-risk status before severe symptoms appear.
Additionally, some AI models are designed to assist healthcare professionals in identifying mental health issues through online social content.
One of the potential applications of AI models is their ability to forecast larger healthcare trends and risks on a broader level.
One example is the Foresight model, trained on de-identified data from 57 million people under the control of the National Health Service in England.
With access to the vast amount of national-scale data, the AI model might be expected to predict potential health outcomes across various demographics.
Initiatives like this aim to enable healthcare providers to predict health issues before they arise, forecast broader future health trends, and highlight inequalities in healthcare.
The role of open-source AI models
As interest in AI models’ capabilities grows, healthcare organisations are exploring the benefits of these innovations.
However, not every healthcare organisation has the resources and time to develop large AI systems from scratch. Yet, this doesn’t mean they can’t take advantage of the potential that LLMs offer.
For many organisations, open-source models serve as an excellent starting point for incorporating AI into their healthcare services.
One significant advantage of open-source models is that they can be hosted locally. This means sensitive data stays within the organisation, addressing the concerns around privacy and security of medical records.
By keeping data inside the organisation, this approach enables compliance with regulations such as GDPR or HIPAA.
Another key aspect is the ability to customise open-source models.
To do that, healthcare providers can use data science services from companies like Lemberg Solutions to adjust the model to their specific needs.
A data science team can fine-tune or pre-train the model using the organisation’s dataset — whether it’s data from a specific hospital department or a particular disease area.
The dataset may include actual treatment protocols and examples of how healthcare professionals make decisions, such as the phrasing of recommendations or the appearance of advice.
Healthcare providers can also choose to work with AI models known as small language models.
These models are smaller in size and scope, which makes them more cost-effective; they also might require less computational power and memory.
This option may be ideal for organisations considering AI implementation to test its effectiveness.
Transforming healthcare with LLMs
The application of LLMs in healthcare is still in its early stages.
Many AI models are currently in the research phase, with some available only to a limited number of organisations for feedback to test them before they are launched on the market.
Yet, healthcare providers are actively exploring how to integrate AI into their existing systems — and take the most of their data.
As adoption grows, so does the understanding that these models are only effective when fine-tuned and trained properly. This requires careful attention to risks, ethical considerations, and compliance with the recent regulations.
Looking ahead, AI models have all the chances to be an excellent tool for improving healthcare services.
They could be helpful in early disease detection, predicting patient readmission risks, and, moreover, supporting public health by forecasting trends.
link