Can artificial intelligence systems be used to detect tuberculosis and silicosis among ex-miners in Southern Africa?
Heads of state worldwide gathered in New York on September 26, 2018 at the United Nations General Assembly explicitly to accelerate efforts to tackle tuberculosis (TB). An estimated 10 million people developed TB disease in 2017 and TB remains the top infectious cause of death globally.
There is growing interest in using artificial intelligence (AI) to help address daunting challenges in low-resource healthcare settings. Advancements have been made regarding the use of artificial neural networks in computer-aided detection (CAD) to diagnose TB, with some commercially available systems achieving promising results. Nonetheless, one very high-risk group has not yet received attention with this new technology: people who worked in South African mines, who collectively had TB rates higher than any other population worldwide and ~10 times the incidence the World Health Organization (WHO) considers a health emergency.
Tuberculosis in miners — a more complex scenario
Important drivers of TB in gold miners are silica dust and silicosis, the latter reaching prevalence rates of 25% in long-service miners. World efforts have been launched to directly target TB and silicosis in the ex-miner population, bringing mass screening in this low-resource setting under the international spotlight. However, none of the existing CAD TB models have been validated for use in a population with high rates of both TB and silicosis. Silicosis can result in lung changes that mask the appearance of TB or mimic it on chest x-rays, thus this process is more complicated than simply detecting TB in otherwise normal chest x-rays. This is why we are working on a pilot to test this technology as part of DFID’s Frontier Technology Livestreaming Programme.
To compound this problem, South Africa’s workers compensation system for occupational lung disease, which dates back over a century, was racially discriminatory both in payment levels and in access to medical examination. While racial discrimination was removed from the statute in 1994, rectifying the plight of black gold miners only began recently. With an estimated 130,000 active gold miners in South Africa, as well as an estimated 800,000 more ex-miners throughout the Southern African region, pressure for redress has been mounting over recent years, including in recent successful class action suits.
However, in spite of these changes, major barriers remain. Waiting times between assessment and compensation are typically more than 1 year. This is primarily due to the insufficient number of medical experts trained to interpret the clinical tests — specifically the chest x-rays, which are essential for the diagnosis of TB and silicosis. The need to assess degree of impairment from lung function tests requires expert assessment and cannot be ignored, but the efficiency of the testing might be improved. Thus, we are training AI systems to detect TB and/or silicosis in this population with high rates of both TB and silicosis. In addition, we are analyzing the systems’ abilities to be able to adequately differentiate between TB and silicosis.
Sprint 1 — our first steps
Our work began with a major meeting of various stakeholders from different disciplinary backgrounds, the mining sector, government, and others, to set out background and our plan, and to seek input.
One set of experiments we conducted assessed how well an AI system can detect TB in chest x-rays. The experiments revealed that chest x-ray AI systems are very good at differentiating TB from normal chest x-rays. However, they are much less accurate in detecting TB in a population that also had silicosis or differentiating TB from silicosis. As correctly identifying both the presence and the type of disease is of critical importance, our next steps will involve working with the companies to improve their accuracy. This will include continuing to re-assess their AI systems as they continue to “train” them with new cases, assisting in the determination of threshold “cut-off” points when classifying the chest x-rays and to work with our partners to ensure that they have enough chest x-rays to train their systems.
Another key part of the TB diagnostic process is assessing lung function. AI systems need to be able to determine how impaired lung function is, but also determine the quality of the test itself, in terms of the effort made by the person being tested and other variables that might interfere with the test’s accuracy. The aim is to ascertain how well a AI system for lung function compares with “gold standard” human assessment. Work on the lung function tests has proceeded in a different way from what we originally foreseen. Two groups within our extended team with different interests in the subject have been working on the problem. One, based in Johannesburg, is focused on developing a programmed algorithm for assessing lung function impairment levels using pre-determined standards. The other, based in Cape Town, is working on an AI assessment of spirometry quality.
Prioritizing treatment
As part of the first Sprint, it has also been important to explore how we might prioritize who should be assessed first. Of the potentially hundreds of thousands of miners and ex-miners from South African and neighbouring countries who have been exposed over the years, the majority do not have silicosis or active TB disease. We are therefore looking to find a way to more efficiently provide the health and social benefits to those in need at any one time.
To do this, we have been making progress on organizing data on exposure, including merging databases and matching job titles, so that a job exposure matrix can be created. This would enable each miner and ex-miner’s risk to be assessed based on the various job titles they held over the years, the mines they worked in, when they worked there and for how long. While the merger of the database has gone smoothly, the matching of the job titles and assigning exposure levels is taking a longer than expected because exposure data has not been as accessible as we had hoped.
What next?
In the next phase will have to tap into the expertise of occupational hygienists familiar with South Africa’s gold mining industry to see how progress can be made more swiftly to determine exposure level assessments of ex-miners. We will also will have to figure out collectively how best to take age, calendar years of exposure and latency into account. We also conducted field-testing of the AI in our first Sprint in which chest x-rays were classified by the AI before being sent to panels of occupational medical and radiological experts. We are looking to determine the results of this as part of our next phase.
Our conclusion for Sprint 1 is that progress is being made, and the eagerness of the users to get the technology into widespread use is extremely high. There are exciting opportunities for widespread use and further validation and refinement. The time is right to get all the various parties together working on each of the respective experiments to fully share our insights, debrief on the progress, and define next steps, so that we have solid scientific evidence to back up the usage.
