The benefits of AI in healthcare are becoming clearer with real-world testing of new AI solutions. From improving attendance to tracking wound care with AI-powered technology. It just shows when used right and in the appropriate places, AI technology can be a great benefit.
At the NHS Providers Annual Conference the UK Health Secretary made clear that technology will be an important part of the their plan. So it’s likely that AI use in healthcare will continue to grow in the UK. A change NHS staff are supportive of. 76% of NHS staff support using AI for patient care and 81% support the use of it in administration.
Let's start with the basics
First we’ll go over some of the jargon in the AI field, that can be a barrier for organisations who want to explore the potential for AI in their systems. Then we’ll have a look at some case studies to put the terminology into practice.
Automation
The use of technology to automate repetitive tasks is a major application of AI in healthcare. The machine cannot do the task if a human doesn’t provide the instructions. Patient monitoring and administrative tasks such as sending appointment reminders can be automated. .
Artificial intelligence (AI)
A field of study into the capability of computer systems to mimic human speech. AI learns from data to simulate human understanding and make decisions.
Data
Can take the form of numbers and statistics, symbols, text, or multimedia. It is information that can be analysed to draw insights from. Patient records, clinical studies, and health monitoring information are examples of data.
Model
A real-world representation in program form. Trained on a set of data to recognise patterns and make decisions without human involvement. A model could be trained on images of moles and whether they are cancerous to help detect cancer in future patients.
Machine learning (ML)
A subset of AI that allows machines to learn and improve based on experience without being completely programmed. ML allows systems to learn from data, enhancing their ability to improve patient outcomes.
Prompt
A question or a command input is given to an AI model.
Understanding models and data
Accuracy
How often does the model get it right overall. Essentially you want closer to 100%.
Algorithmic Impact Assessment (AIA)
A method to evaluate how an algorithm, such as one used for patient risk scoring, affects patients, healthcare staff, and the broader healthcare system. It ensures that the AI system does not have unintended harmful consequences.
Anonymisation
The process of removing personally identifiable information from data so that individuals cannot be identified. In healthcare, this is important for protecting patient privacy when using data for research or training AI models.
Application Programming Interface (API)
A set of rules that allows different software applications to communicate. In healthcare, APIs allow electronic health record (EHR) systems to interact with AI models to provide clinical decision support or automate tasks.
Bias
AI systems can reflect human biases in their training data.
Binary
A system of classification with two categories for example healthy vs. sick.
Classification
A type of machine learning where AI models categorise data into classes. In healthcare, classification models can be used to categorise medical images (e.g., identifying tumor types) or patient conditions (e.g. diabetic vs. non-diabetic).
Computer vision
A field of AI that enables computers to interpret and understand visual information, such as medical imaging. In healthcare, computer vision is used to analyze X-rays, MRIs, or CT scans to identify diseases like cancer.
Explainability
How easy to understand are the AI system’s decisions to humans. It must be able to explain why it has made the predictions it has. If it has prioritised patients by those most in need then it needs to be able to explain this to humans.
Explainable AI (XIA)
This is a method of describing an AI model. It’s accuracy, potential biases, and how it got to the result that it did.
Model drift
This is when the model starts to ‘drift’ from its original parameters. If the model’s training doesn’t line up with the new data then it can’t use it to make accurate predictions. It’s a large part of AI governance.
Narrow Artificial Intelligence
AI systems that are designed to perform a specific task. In healthcare, narrow AI is widely used for tasks like analysing radiology images or predicting patient readmission risk, but it cannot perform tasks outside its specific function.
Precision
The higher the precision then the less likelihood that there are false positives. If we imagine the hits on a dart board are tightly clustered then this is high precision. Precision is especially important when the consequence of false positives is highly sensitive.
Scalability
A model is scalable when it can used with much larger data sets and produce results from the entire thing. Scalability refers to how well the model can do this without compromising the precision and quality of its results. Its capacity to do this is vital as the volume of data sets increases and the number of users increases.
Sensitivity or recall
In the same vein as accuracy and precision, sensitivity also called recall answers if the model can find all instances of the positive class. For example, for a medical test, the sensitivity measures how often the test correctly identified the positive result. The higher the recall the better.
Specificity
The same as sensitivity but for how well it correctly identified negative results.
Training
The process of teaching a machine learning algorithm to make decisions based on data. Similar to training with clinical data to improve diagnosis and treatment for healthcare.
Data types used
Big data
Refers to gigantic and diverse collections of data that grow with time. Datasets this large cannot be stored by traditional data management systems. In healthcare, this can include processing multiple data sources like genetic data, medical history, and other info for personalised treatment.
Labelled/ unlabelled data
Labelled data contains tags and is used in supervised learning while unlabelled data doesn’t contain tags and is used in unsupervised learning.
Structured data
Organised and formatted data readable for humans and machines. A patient’s electronic health record (EHR) with named fields like blood type name etc is an example of structured data.
Synthetic data
Not real data but designed to replicate real-world data.
Test data
A way of checking a machine learning (ML) algorithm has been sufficiently trained by testing it on data to see how well it works.
Training data
The data is used to train the machine learning model to teach it how to do a task.
Unstructured data
This is data with no structure, a vast collection of unsorted data.
Validation data
This data is not in the model train data that is used to test how well the model is doing with new data. The metrics used to determine its performance are accuracy, specificity, precision, and sensitivity.
Types of machine learning
Neural network
A type of machine learning that simulates how the human brain makes decisions using neurons. ChatGPT has 176 billion neurons in its neural network, and the human brain has around 100 billion neurons.
Reinforcement machine learning
Enables the system to learn from trial and error based on feedback.
Semi-supervised machine learning
Combines unsupervised and supervised learning using both labeled and unlabelled data.
Supervised machine learning
Uses labelled datasets to train algorithms.
Unsupervised machine learning
Uses unlabelled data to discover patterns without any instructions.
AI Solutions for Healthcare: Tackling Real-World Problems
Sensitivity in Healthcare with AI
AI is changing how organisations provide information and resources to the public, particularly in sensitive areas like sexual health. For healthcare providers, AI is an accessible, efficient and reliable way to reach users confidentially while ensuring sound medical advice.
Positive East and Japeto's sexual health chatbot Pat
Japeto and Positive East created the chatbot named Pat, to address user questions, locate nearby testing sites, and triage STI concerns via an automated conversational interface.
The chatbot was built using supervised machine learning (ML) models trained on labelled data derived from user surveys and existing sexual health resources. These models, powered by AWS Lex, were designed to handle the nuances of natural language processing (NLP). By incorporating training data with diverse examples, such as slang and common misspellings, Pat ensures high accuracy and specificity in recognising and responding to user prompts.
Impact
Pat offers around-the-clock support, enabling users to access advice outside traditional working hours. So unlike traditional phone lines, Pat can handle increased traffic, simultaneous queries from multiple users, demonstrating strong scalability. With a knowledge base of over 300 scripted responses, from structured data, Pat addresses a wide range of common sexual health questions with high sensitivity and recall.
Pat’s success highlights the potential of narrow artificial intelligence to benefit the healthcare sector. AI-powered technology coupled with ethical considerations like anonymisation, can address sensitive issues effectively and confidentially.
Reducing DNAs and Last-Minute Cancellations
DNAs often referred to as “did not attends” cost the NHS around £1 billion every year. By using AI for patient engagement, hospitals can reduce these inefficiencies. Missed appointments lead to inefficiencies and wasted time. AI can help reduce them by predicting patient behaviours and sending automated reminders.
The Mid and South Essex NHS Foundation Trust's DNA prevention programme
The Mid and South Essex NHS Foundation Trust used a model that used machine learning reduced their DNAs rate by almost a third. Encouraged by the success, the Trust is using the AI solution across the whole organisation as of July 2024.. A tool of efficiency increasing productivity with 80,000 more patients that can be seen at the trust each year.
The predictive analytics is developed by the deep learning model, which is trained, fine-tuned with validation data then tested. Once the accuracy, precision, specificity and sensitivity is ensured it is ready to deploy.
ML systems refine their predictions over time, offering insights into factors influencing patient behaviour. Factors such as patients’ jobs, childcare commitments, as well as traffic and weather can be included in the training data a model uses. Reducing DNAs can contribute to more efficient use of healthcare resources, enhancing service delivery and accessibility for other patients.
Clinical Applications of AI in Wound Care
One of the most interesting applications of AI in healthcare is AI-powered wound care technology One of the most promising applications is the use of AI-powered colour recognition technology. This technology helps monitor and assess wounds more accurately and efficiently.
North Cumbria Integrated Care's wound tracking system
At North Cumbria Integrated Care NCIC, approximately 50% of community nurses’ workload involves managing acute and chronic wounds, amounting to a staggering £41.7 million annually. Nationally, wound care accounts for £8.3 billion per year of NHS costs. Community nurses relied mostly on manual tape measures for wound assessments. In more complex cases, staff would send photos to higher-level staff for remote advice.
The AI solution NCIC used, when integrated with cameras or imaging devices, analysed wound images to assess their healing progress. Algorithms recognise subtle changes in wound appearances—such as colour, size, or texture—accounting for lighting, size, depth and user. Using ML to provide insights into how the wound is responding to treatment. The AI technology captures a precise 3D model of the wound, standardising the assessment process in alignment with clinical guidelines. Ensuring that all staff—regardless of their level—can assess wounds confidently and safely.
Streamlining Healthcare Administrative Tasks with AI
AI in healthcare administration offers a practical solution, automating routine processes to save time and resources. Hospitals and clinics often deal with massive amounts of paperwork and administrative tasks, which can be overwhelming for staff and contribute to burnout. Automating routine tasks is a good step to reducing the pressure on the NHS.
A Southeast Primary Care Surgery's Automated Call Handling System
A primary care surgery in the Southeast, serving 14,500 patients, struggled with long call wait times leading to almost a quarter of callers abandoning their call. The reception team of eight staff was overwhelmed, and far from meeting their targets.
The surgery used an automated call-handling system leverages Natural Language Processing (NLP) to understand and respond to patient inquiries in real time. The AI answers and captures patient details, and processes this data into a structured data format. By integrating directly with the surgery’s practice management system (PMS), the AI can access and update patient records, schedule appointments, and retrieve relevant information without manual intervention.
The surgery saved the equivalent of 15 full workdays each week. This was a significant reduction in the workload for the reception team and more than 90% of patients reported an improvement in the service.
The Future of AI in the NHS
The integration of AI into healthcare holds many measurable benefits. From reducing missed appointments and improving patient engagement to improving wound care and easing administrative burdens, AI has the potential to address some of the NHS’s most pressing challenges.
If the UK leverages its position as a global tech leader, the collaboration between innovative AI solutions and the NHS’s healthcare framework can set a benchmark for life sciences worldwide. By investing in AI for clinical settings and AI-driven healthcare innovations, the NHS aims to tackle its waiting list challenges while enhancing patient outcomes.
