On the state of artificial intelligence development in laboratory medicine

Artificial intelligence (AI) models in healthcare have the potential to improve the accuracy and speed of personalized medicine for patients, and in some cases help determine the best treatment or preventive care. Clinicians are already implementing these models in areas such as early detection of sepsis and analysis of radiological images for the diagnosis of prostate cancer and other diseases. This is a growing area of interest that laboratory medicine professionals should pay attention to, as data obtained from laboratory tests is a key component of AI tools for clinical decision-making.

Clinical artificial intelligence and a subset of artificial intelligence known as machine learning (ML) can be used for tasks across a wide range of fields, from precision medicine to public health. The main advantage is speed, as these tools rely on computerized rather than manual tasks. There is great interest in how to make artificial intelligence algorithms more advanced so that they can predict more complex outcomes, such as response to cancer treatment or the risk of side effects from surgery.

Clinicians focus on understanding individual patients. But laboratory technicians have excellent skills in aggregating and processing data. A natural extension of providing laboratory results is to provide risk analysis or probable diagnosis. That is what all these predictive analysis tools are designed for.

CURRENT AND POTENTIAL USE OF ARTIFICIAL INTELLIGENCE IN THE LABORATORY.

In laboratory settings, there has been some adoption of artificial intelligence and machine learning methods, mainly in molecular pathology (e.g., classification of central nervous system tumors using DNA methylation profiling) and digital pathology (e.g., image analysis), but this has been slow.

In a review article (Clin Biochem 2022; doi:10.1016/j.clinbiochem.2022.02.011), highlighted some examples of artificial intelligence/machine learning being studied in the laboratory, such as predicting laboratory test values, improving laboratory utilization, automating laboratory processes, facilitating accurate interpretation of laboratory tests, and improving laboratory medicine information systems—some with impressive accuracy.

For example, the study discussed in the article used a neural network model to predict iron deficiency anemia and serum iron levels based on routine complete blood count (CBC) findings. Another study discussed the development of a machine learning model capable of recommending which laboratory tests a physician should order. In general, artificial intelligence/machine learning technology has the potential to use large amounts of medical data to create more personalized interpretations of test results. Thus, the paradigm may shift from defining normal hemoglobin levels in general to defining them for a specific individual.

Chemistry and immunology laboratories are particularly well suited to machine learning because they generate large, highly structured datasets (Clin Chem 2021; doi: 10.1093/clinchem/hvab165). The labor-intensive processes used to interpret and quality control electrophoresis and mass spectrometry records could benefit from automation as technologies improve. The article notes that clinical chemistry laboratories also generate digital images, such as urine sediment analysis, which may be highly amenable to semi-automated analysis given advances in computer vision.

There are two general classes of artificial intelligence models. The first solves internal problems in the laboratory, such as providing clinicians with more accurate results, and the second seeks to identify patient cohorts and care processes to eliminate gaps in quality in healthcare systems.

OBSTACLES TO THE ROUTINE IMPLEMENTATION OF AI.

Laboratories will still face significant challenges before this technology can be used more widely. These include the need to collect high-quality data from diverse populations and manage the costs associated with computing infrastructure and personnel to develop and update algorithms and software tools.

Understanding the different tasks that AI can help with can help solve some problems. For people who are not as familiar with AI, trusting an AI model to perform a diagnostic task that is performed by highly specialized personnel or biologists can create a major barrier to trust, as opposed to using AI for something that seems less risky, such as evaluating work processes in areas that can be optimized, or identifying patterns in test usage that can be improved. Another approach is to implement an AI program in parallel with the manual process, evaluating its performance in the process to facilitate the use of the program. 

There are also ethical considerations regarding the use of artificial intelligence in medicine. One is access to data, and another is how to properly obtain patient consent to include their data in larger pools. And how can clinicians ensure that they do not introduce additional biases when using the data? 

Artificial intelligence models may also be ineffective if the dataset used does not reflect the population served by a particular laboratory or healthcare system. 

BUILDING A BRIDGE TO THE FUTURE.

To assist with data access and related issues, the US National Institutes of Health has launched the Bridge2AI program to create new flagship biomedical and behavioral data sets. The program also aims to identify best practices for collecting and preparing artificial intelligence/machine learning data for biomedical and behavioral research. 

This is a very specialized program that aims to generate large amounts of data with the appropriate consent structure so that the data can be used to create artificial intelligence models that drive innovation in artificial intelligence that can lead to improvements in human health.

In the future, artificial intelligence will be directly integrated into more devices and tools, and laboratory managers may not choose separate AI programs. Instead, AI functions may be included in larger software packages that they are considering. 

Material prepared by

Karen Bloom, freelance medical/science journalist from Owings Mills, Maryland. Email: karen_blum@verizon.net