By Bart Steverink

Estimated Reading Time: 5-7 min

You’ve probably noticed—there’s a lot of hype around AI. Attend any tech event, and you’ll experience it firsthand. The healthcare sector has also joined in the AI buzz. At CES, the world’s largest tech event held annually in Las Vegas, nearly 80% of the booths in the health section advertise that their service or product involves AI in some way.

Looking at current trends, healthcare is at the forefront of data growth, with an ever-increasing amount of data being generated by new applications, such as wearables and increasingly sophisticated hospital equipment. Year after year, the volume of data expands, yet only a small fraction is effectively utilized. However, buried within these data are valuable insights that can help guide daily healthcare operations and improve future healthcare services. This is where AI, and more specifically Machine Learning (ML), can play a significant role.

What is Machine Learning?

Machine Learning is a subset of AI and uses (semi-)structured data and machine learning algorithms to create models. These models enable us to make predictions about new situations based on patterns observed in known situations (inference).

For example, an ML model can predict the likelihood of a patient staying in the hospital longer than planned or the probability of an infection after surgery. There are numerous ML algorithms, each with its own strengths and weaknesses. The best algorithm is chosen based on the specific application and the performance of the trained model on test data.

Applications in Healthcare

ML applications in healthcare are mainly in the areas of imaging (for example, recognizing abnormalities on an MRI scan or ultrasound), predicting complications, and advising on medication and treatment. These applications are considered Clinical Decision Support (CDS), assisting healthcare professionals in their decision-making process. It is important to emphasize that the healthcare professional makes the final decision, not the model. The model provides additional information to give the healthcare professional a more complete picture of the situation.

In addition to clinical applications, there are also AI/ML applications in the process and logistics side of healthcare, such as predicting readmissions or length of stay and reducing administrative burdens, for example, using LLMs for converting speech to text.

Developing ML Applications

The development process of an ML application naturally begins with identifying a relevant (clinical) problem or area for improvement. Through a literature review, it is determined whether there is already a theoretical basis for the predictive factors (features) that the model should use. The availability and quality of data are also assessed.

Is the data, for example, already well-structured with a fixed table and column structure? Are all fields always present, or are data points often missing? Besides structure and availability, it is essential to understand the origins of the data. Is blood pressure measured manually or with a device, and is this always the case or only sometimes?

Data processing and cleaning represent the bulk of the work in this phase of the development process.

Testing Models

Based on insights from the literature and the available data, the ML algorithm and the model’s features are selected. The total available dataset is divided into a training set, a validation set, and a test set. The model’s performance on the test set is used to choose the best-performing model. At this stage, we now have a trained model that can be used to make predictions. However, this is not the final step. The model was developed and tested in a “laboratory” setting, but this does not guarantee that it will achieve its intended purpose in real-world practice. Moreover, what we have at this point is just a model, a file stored on a computer’s hard drive. It is not yet a usable application for clinicians.

Implementation in the Clinic

For the model to be used in practice, it first needs to be packaged into an application. This could be a computer program with a user interface or an integration with the interface of the Electronic Health Record (EHR) system. Code must be written to handle and log errors in a system so that notifications can be generated if something goes wrong.

The performance of the model and the characteristics of the data must be continuously monitored to detect issues such as model drift, concept drift, and data drift. Additionally, it is crucial to assess the model’s impact in the clinic—does it effectively contribute to solving the problem, or does it introduce unintended side effects?

The Future of AI and ML in Healthcare

AI and ML hold enormous potential in healthcare. The increasing volume of data, trends toward standardized data exchange, and growing pressure on healthcare professionals and the system as a whole make it worthwhile to invest time and effort in these technologies. The AI landscape is also evolving rapidly. Today, much attention is focused on generative AI and Large Language Models (LLMs), but in just a few months, the discussion may shift and we may be talking about agent-based AI models at the coffee machine.

One thing is certain: AI and Machine Learning will soon become indispensable in the healthcare system. By learning, experimenting, and understanding, we can fully harness AI’s potential to responsibly and efficiently improve healthcare.