The future of AI in the NHS - part three

Earlier this year, I completed the ‘Artificial Intelligence in Health Care’ course at the MIT Sloan School of Management. I found it to be fascinating, thought provoking and very worthwhile. Specifically, it took a subject I was generally interested in and helped me focus on the precise issues of relevance to not only my practice but the future of NHS health care overall.

In this series of blog posts, I am repurposing my written assignments as articles so colleagues in the NHS and globally can share in that thought process. You will see that, at this stage, there are more questions than answers, but hopefully together we can work to move the theoretical towards the practical.

What are the advantages and challenges of using natural language processing and data analytics in healthcare?

As outlined in my previous post, the UK NHS generates a vast amount of data daily. The sources include primary care, secondary care (hospitals) and ancillary services including diagnostics, social care, and other support services. Although some of that data is coded using SNOMED system, the majority is stored in textual format. Furthermore, a sizable number of hospitals are still reliant on paper-based records. Similarly, most care and nursing homes are also reliant on the same format.

I carried out brief research on using Natural Language Processing (NLP) in healthcare here in the UK and came across Data Lens, which uses NLP and other AI tools to create a universal search engine for health and social care data catalogues and metadata (1). It joins up data catalogues from NHS Digital, the Health Innovation Gateway, MDXCube, NHS Data Catalogue, Public Health England Fingertips, and the Office for National Statistics (ONS).

According to Health Data Research (HDR), electronic health records (EHR) contain a rich history of the patient journey with huge potential improvement in direct patient care and NLP can utilize free text information recorded in electronic health records to help us learn how to tailor treatments more accurately for individual patients. According to HDR, and as I stated earlier, the challenge is that most of the information contained within the EHR is unstructured text.

I have been involved with Microsoft and their subsidiary Nuance in discussing and demoing their Dragon Ambient eXperience (DAX) which is a direct implementation of NLP use in healthcare. It offers both physicians and patients an uninterrupted interaction and naturally flowing consultation. In terms of using NLP to extract meaningful data from consultations text, I wondered how NLP would handle abbreviations that are used by clinicians as a time-saving measure when documenting their findings. I realized that adapting models like DAX would likely result in cleaner data with less contaminants such as abbreviations. Additionally, as AI is using NLP to process the interactions, it can also extract patterns and data that can be fed to other relevant AI algorithms and analytics models. The following questions are therefore relevant:

1. Should there be a national or even global discussion on the issue of adopting universal standards for developing basic building blocks to create algorithms to be used in NLP and deep machine learning (DML)? I believe that there should be a global basic standardization of data sets used to build NLP and DML tools. In doing so, we can achieve worldwide standardization of data collection, annotation and subsequently higher quality data for NLP training.
   

2. If the response to the question above is yes, who should be tasked in creating such standards? It would be feasible for the World Health Organization (WHO) to take a lead in coordinating efforts between various healthcare systems and to provide guidelines on what standards should be followed to create the data set required for NLP training. However, such guidelines must consider discrepancies between various healthcare systems and the resources available to them to adopt such guidelines. We have seen the WHO taking a lead and guiding the world during the pandemic of COVID-19; it is likely that it can lead similarly on the issue at hand.
   

3. Can NLP help in joining up data sets generated by various NHS components? In my previous posts, I alluded to some of the influences exerted on the decision-making process within the NHS. One such influence is lack of coordinated political will to allow better communication and collaboration between various components of the healthcare system. In my opinion, this represents an obstacle to a joined-up healthcare system despite the best efforts and large amount of human and financial capital. It is possible to envisage that the introduction of NLP to harness the data contained in the electronic health records and to allow better decision-making could ultimately lead to a more coordinated and smoother-running healthcare system.
   

4. Human healthcare professionals are generally trained using standard and uniform programs. However, these professionals remain unique individuals which is reflected in their data input styles. Can NLP and AI cope with myriad variations? It is possible that NLP will, initially, struggle with the large number of variations that exist in EHR. This is due to the various ways, styles and choice of phrases by human healthcare professional in documenting consultations in EHR. This is in addition to the use of non-standardized abbreviations that may not be understood by NLP models. LLM and deep machine learning may be able to cope with this challenge. It is also possible that with increasing compute power and more sophisticated algorithms this challenge can be overcome.
   

The use of large quantities of correctly annotated data would be ideal to train NLP in the United Kingdom NHS. Such a model will inevitably provide a higher quality and cleaner data to train NLP. However, given the significant fragmentation of the available data within the NHS and the need for very complicated and prolonged annotation of such data, there will be a need for significant resources and a prolonged period to achieve such annotation and make the data available to train NLP. To put this in perspective, the NHS has been an existence for the last 76 years and digitization of primary care medical records in the United Kingdom primary care only began in the 80s. Before that, data was generated and stored in mostly handwritten paper format.

The use of domain and NLP experts to build a customized model that can make predictions based on limited data, in contrast to the model above, may yield faster results and require less resources to build. However, I would be concerned about the accuracy of the resulting predictions, given the limited access to the data used in the training. Although the quality of the data is likely to be higher caliber, I do not believe that this model would be suitable for the vast data available with the NHS and I am unable to comment on the availability of domain and NLP experts in the UK, although it is possible that such experts are available. This aspect would require further research.

I wonder if it would be feasible to construct a model that combines the two models highlighted above, where structured and annotated data can be used initially to generate tools that can then annotate a larger data set. This would institute further retraining and development to improve the predictability and output of NLP. I am uncertain about at what point further retraining becomes unnecessary and whether NLP will exceed the human capacity to provide structured and annotated data.

I intend to explore this aspect further in future posts and invite you to contribute your thoughts!

Connect with Contempo and join the discussion!