In 2016, the British cognitive psychologist and computer scientist Geoffrey Hinton called for a halt to training further radiologists. With new AI-based radiology applications becoming more effective all the time, Hinton believed that AI may take on a large share of radiology work [1]. 

AI promises to have a similar impact on the lab diagnostics space, but as in radiology, some types of lab work are more prone to automation than others.

Diagnostic tasks that require triangulation between many varied datasets will be relatively hard to automate. If an AI system were to learn how to diagnose breast cancer from scratch, for example, it would have to consider more than 30 variables each time it attempts a diagnosis. This includes structured information from imaging scans and genetic tests, as well as unstructured patient narratives and family histories. To get to the same level of expertise as an expert physician, it has been estimated that the AI system would have to analyse 7 billion unique breast cancer samples.

Other diagnostic tasks may be more amenable to automation. Compared to breast cancer, fewer variables need to be considered when diagnosing coronary heart disease, mapping flu trends or recognising arrhythmia in ECGs. Over the last two years, two teams showed that AI is as good as an expert physician at diagnosing skin cancer, while another team at Stanford demonstrated the same for pneumonia [2].

AI also has potential to change the lab diagnostics space in two broader ways: by improving access to diagnostic services in low-resource settings and creating new opportunities for continuous patient monitoring.

Bringing diagnostic services to those who lack them

In many developing countries and in remote parts of the developed world, millions of people lack access to basic diagnostics services, including consultation time with physicians who can generate, interpret and explain a lab result. By taking on some of this work and reducing the cost of diagnosis, AI can play a big role in addressing this problem.

While some AI-supported diagnostic systems may issue affirmative diagnoses, others will focus specifically on ruling out severe diseases. The negative predictive value achieved by AI systems for many severe diseases is already in the 90s, reaching an accuracy often better than that of community physicians working in rural areas. For those whose tests do come back positive, these patients could be referred and managed by regional healthcare systems, where trained experts in clinical and laboratory medicine can provide them with the care they need.

Improving care with continuous monitoring

In the longer term, labs can also use AI to incorporate new types of patient data and continuously monitor patients in a more personalised way. Every year, more people are using wearable devices that track their steps, heart rate, sleep patterns, and calories in real time. We may eventually even see widespread adoption of smart toothbrushes that extract data from biomarkers in saliva while we brush our teeth.

Within the next five years, such advances will transform what we know about patients. A patient will not just be someone whose vital signs are observed at discrete points in time. That will still happen, but labs will also have access to continuous signals too. Any significant personal variations allow laboratories to support timely adjustments to existing care plans. We are beginning to understand that it is not population reference ranges that determine whether a person’s results are normal, but rather how they change over time.

Over the next five years, more practical applications of AI will emerge. Laboratories that prepare now for an AI-supported healthcare system will stand to benefit the most. To prepare for this new reality, it will be key for labs to start digitising information, indexing it and storing it in an easily accessible format.

[1] A.I. Versus M.D., What happens when diagnosis is automated? The New Yorker, 2017 

[2] Stanford algorithm can diagnose pneumonia better than radiologists, Stanford News, 2017

This article is based on a presentation “Artificial Intelligence in Healthcare and Diagnostics” at LEADx Diagnostics Leadership Summit in Mumbai, India.