Generative AI in Healthcare: Solving Medical Staff Performance Issues

To put this in perspective, 1 petabyte equals 1 million gigabytes. This means a single hospital produces as much data each year as approximately 50 million smartphones, each with 1 terabyte of storage. In fact, 95% of this data goes unused because most of this data is unstructured and uncategorised.

Moreover, reviewing medical records "manually" is time-consuming and can lead to missed information—mainly when healthcare providers have only 15-20 minutes for patient consultations.

Typical challenges of manual data review in healthcare:

• Missed critical information: in cases with lots of unstructured data, it is very easy to overlook some essential information from patients' condition or history.
• Delayed decision-making: when doctors need to process a comprehensive set of records for analysis and communication, this can lead to delays in diagnosis or treatment.
• Less time with the patient: reading and comprehending reports takes away time from direct patient care, potentially compromising the quality of healthcare delivery.

Our team has developed a proof of concept (PoC) powered by generative AI (gen AI) to address these challenges. This solution not only creates patient data summaries that can be easily customised to meet the needs of various healthcare professionals, including doctors, nurses, and administrative staff, but also features a document classification capability that allows the support of different types of clinical documents. Let's examine our solution in more detail.

ELEKS’ AI-driven medical document summarisation

Our team aimed to develop a gen AI solution that could produce summaries of medical documents in a specific, predefined format suitable for healthcare professionals. For this PoC, our team researched, compared, and selected the most relevant open-to-use LLM model in compliance with healthcare industry regulations.

We have set the following guidelines for the PoC:

• PoC will use a generic predefined prompt to have a straightforward summary.
• We will focus on summarising specific healthcare documents, including nursing assessments, therapy notes, doctors' notes, plan of care (CMS form 485), lab tests, OASIS document, and discharge summary.
• All processes will adhere to HIPAA and GDPR regulations.

A major challenge in our process was acquiring document samples, given the sensitive nature of healthcare data. While we were able to find different document templates online, obtaining actual samples proved difficult. For this purpose, we used ChatGPT to generate some of the samples of these documents.

Additionally, operating within the healthcare domain necessitates strict adherence to data security protocols and compliance with regulations such as HIPAA and GDPR. This focus on data security and privacy is crucial, especially considering that 435 incidents of data breaches were reported in 2024, as of when this article was published.

Process overview: from document intake to AI-powered summarisation

Our medical document summarisation solution operates autonomously, activating as soon as new records are added to the database. The process begins when a medical document is received via email, triggering the initial processing phase.

We've observed that medical records are commonly stored as image files or PDFs. To address this, we've implemented an Optical Character Recognition (OCR) module. This was the necessary step, as it converted these documents from images and unrecognised text to recognised and searchable text.
Once the OCR process is finished and the document's content is recognised, it is stored in the database as discrete fields in text format, turning unstructured data into structured content.

Based on the kinds of clinical documents we encounter, a classification module has been included as part of the solution after an OCR was performed. For the classification of documents, we resort to generative AI and thus use the same model employed for summarisation. To this end, we have developed a separate prompt that describes each document type to match documents with its the respective type.

Once records have been recognised with OCR and classified, a summary of the document is automatically generated. At this stage, our solution supports summarisation on an individual document basis, meaning that each new document produces its standalone summary.

In cases where the type of a document is not recognised, the summarisation process will still continue to avoid leaving a document unsummarised.

ELEKS' solution: model selection and tech stack

While developing our PoC, we investigated a variety of LLMs to identify the best solution for generating summaries for medical documents. To determine the best model among all, we based our choice on three parameters: prompting capabilities, price and the final output of the summary.

We investigated both general-purposed models, such as GPT-3.5 Turbo and GPT-4, as well as specifically trained LLMs designed for medical summarisation, which we hosted ourselves, including DHEIVER/Medical Summarisation, MedLlama2, Meta-Llama/Meta-Llama-3.1-8B, and FalconsAI/Medical Summarisation. Let's take a closer look at them.

While these models produced reasonable results, they lacked flexibility. We couldn't customise the output through prompting; we uploaded documents and waited for preconfigured summaries, and prompting capabilities was a key criteria for us. Generally, the results from these models weren't bad; we found that general models combined with effective prompting delivered significantly better outcomes.

This insight led us to choose Llama 3.1, a "self-hosted" LLM that gave us more control over the summarisation process and security, as information is "self-hosted." However, running such a powerful model required considerable computing resources, and maintaining it on a Virtual Machine with adequate power was costly. To address these challenges, we transitioned from "self-hosted" to Azure AI, which allowed us to switch models in minutes, pay only for what we use, and receive immediate output.

We have also tried ChatGPT 3.5 Turbo, but it is limited in terms of the size that will be analysed (a few pages max). After realising this, we decided to switch to a newer model, like GPT-4o.

Currently, we are relying on Azure AI, where we are running OpenAI's GPT-4o model. This solution gives us both the flexibility for prompting and the power to process and summarise large volumes of medical data.

In terms of the overall architecture, our PoC is entirely based on the Microsoft Azure platform. We utilise .NET technology in the backend to manage and adjust workflows. The combination of Azure AI and .NET is effective, flexible, and extendable for summarising medical information.

Conclusions

Over only a couple of weeks, we created our AI-powered solution for medical document summarisation, which represents a significant step forward in addressing the challenges of information overload in healthcare. By leveraging advanced language models and cloud technologies, this tool offers healthcare professionals a way to quickly digest vast amounts of patient data, potentially improving decision-making and patient care.

One can easily see how other systems, formats, and models could be quickly integrated into this proof of concept. As we look to the future, the planned enhancements - including EMR integration and multi-document synthesis - promise to further streamline healthcare processes. This innovation exemplifies how AI can be harnessed to increase efficiency and improve patient outcomes by allowing healthcare providers to focus more on direct patient care. As the healthcare industry continues to evolve, solutions like this will play a crucial role in managing the ever-growing volume of medical data, ultimately contributing to a more effective and responsive healthcare system.