Computer-aided chest X-ray interpretation can be cost saving yet effective for TB diagnosis

Kevin Schwartzman, MD, MPH

12/15/2021

Before the COVID-19 pandemic, tuberculosis was the leading cause of death from a single infectious agent. Unfortunately, all signs point to the same situation once the pandemic subsides, with TB-associated morbidity and mortality having substantially worsened during the pandemic’s first year. Missed diagnoses are a key contributor to TB-related deaths; the World Health Organization indicated that in 2020, only 5.8 million people were diagnosed and reported with TB out of an estimated 10 million worldwide who developed the disease. (1)

The public health context, diagnostic algorithms and testing capacity for TB differ substantially in lower- and middle-income countries with high TB incidence compared with high-income countries with low TB incidence. Limited availability of mycobacterial cultures and phenotypic and genotypic drug susceptibility testing, as well as chest imaging, can compound gaps in the diagnostic and treatment cascade.

In settings where TB is most common, individuals with persistent cough and/or constitutional symptoms are often referred directly for microbiologic testing. Indeed, upfront smear microscopy for those with persistent cough is the cornerstone of the traditional WHO directly observed treatment, short-course strategy. Reflecting the limited sensitivity of this approach, there is also provision for empiric treatment of patients with persistent symptoms but negative smears, particularly after a trial of an antibacterial agent. In many areas with high TB incidence, diagnostic performance has been substantially enhanced by automated nucleic acid amplification testing, notably the GeneXpert® platform. However, this adds considerable expense. And even if the pretest probability of TB disease is high, most patients will have negative results.

The role of chest X-rays in TB diagnosis in high-incidence, lower-income countries

In high-income countries like the United States and Canada, chest X-rays (CXRs) are taken for granted as part of the initial diagnostic approach for any person presenting with significant, persistent respiratory symptoms. However, because of the cost and limited availability of equipment and interpretation, CXRs have not previously been the focus of diagnostic algorithms for TB in high-incidence, lower-income settings.

Yet the cost of chest radiography may be offset by savings from unnecessary microbiologic tests averted. CXRs have high sensitivity and high negative predictive value for the diagnosis of TB, notably in symptomatic adults without HIV infection. This means that most people with normal CXRs can safely be discharged from further testing, particularly if there are mechanisms for later reevaluation of those with persistent symptoms. And the scarcity of human resources for on-site, high-quality CXR interpretation has been mitigated by the advent of increasingly accurate, automated artificial intelligence–based, computer-aided detection. (2,3)

Along these lines, WHO has recently emphasized the potential role of computer-aided detection for CXR-based triage of patients presenting with possible TB symptoms, as well as for screening of people at risk of TB disease. (4) It has articulated target characteristics for the necessary software and infrastructure. (5)

Evaluating computer-aided detection software for triaging patients

Members of our team conducted a field evaluation of two commercially available software packages for automated CXR interpretation, among 2,198 adults evaluated for potential TB symptoms in Karachi, Pakistan. (6) Virtually all were HIV-negative, while 12% had culture-confirmed TB disease (of whom 78% had smear-positive disease). In this group, the estimated sensitivity of both software packages was 93%, while estimated specificity was 69%-75%, meeting WHO’s suggested minimum requirement of 90% sensitivity and 70% specificity for community-based triage.

However, scaling up digital chest radiography with computer-aided detection imposes substantial infrastructure and software costs. In resource-limited settings, it is especially relevant to consider cost, benefit and cost-effectiveness. We therefore used the results of the field evaluation to populate a decision analysis model, designed to predict clinical outcomes and costs that would result from several potential diagnostic pathways with computer-aided CXR interpretation, for patients presenting with possible TB symptoms. For comparison, we also considered “status quo” strategies with upfront microbiologic testing using either sputum smears or Xpert, followed by empiric treatment for those with negative tests but persistent symptoms.

Our latest work, reported in Open Forum Infectious Diseases, incorporates key findings from the parent trial regarding diagnostic yield (7). It extends the trial by considering alternative use cases for CXR along the diagnostic pathway, as well as health system costs and clinical outcomes related to diagnosis, treatment and missed diagnoses.

Based on the high negative predictive value, and thus the capacity to reduce unnecessary testing and empiric treatment, our analysis suggests that computer-aided detection using CXR is likely to be cost-saving, reducing overall health system costs by 19%-37%, while maintaining or even improving clinical outcomes.

Computer-aided detection is also useful for focusing follow-up efforts among persons with initially negative microbiologic tests: Those with CXRs suggestive of possible TB disease can potentially be recalled for repeat evaluation, particularly if they remain symptomatic. Indeed, this was the most cost-effective strategy. The possibility of immediate on-site CXR interpretation and hence more focused referral for microbiologic testing could also potentially reduce losses to follow-up in the diagnostic cascade.

An important caveat is that our conclusions may not be generalizable to settings of high HIV prevalence. That's because the original study that informed our analysis was undertaken in an HIV-negative population. A recent individual patient data meta-analysis of deep learning-based computer-aided CXR detection showed that sensitivity and specificity are both significantly lower amongst people living with HIV. (8)

Important impacts beyond cost savings

Reducing unnecessary testing and treatment is not purely a matter of saving dollars. TB treatment imposes costs on patients, their families and the health system, but it also involves some risk of serious toxicity. And minimizing unnecessary examinations and treatments frees up personnel and equipment to better focus on those who truly need it. For example, some funds and personnel now used for unnecessary first-line microbiologic testing could then be redirected to more advanced laboratory diagnostics. Similarly, health care worker time could be redirected to better retain persons in care during diagnosis and treatment, which would substantially improve their outcomes and reduce ongoing transmission.

References

Science Speaks