Closing the Diagnostic Gap in Active Tuberculosis with AI
Current Gaps in Detection, The Impacts of Prediction
TLDR:
- Tuberculosis kills over 1.6 million people per year; however, most of those fatalities are preventable
- Current tuberculosis detection methods cannot differentiate between latent and active infections nor predict active tuberculosis infections
- AI has the potential to predict active tuberculosis infections by processing risk-related patient data and certain biomarkers
- Early detection through AI can reduce fatalities, lower healthcare costs, decrease transmission rates, and identify patterns to form effective disease control strategies
Early detection can be the deciding factor between life and death when it comes to tuberculosis (TB), an infectious disease known to kill over 1.6 million people annually as of 2021, making it the second leading infectious killer after COVID-19.
However, as with many health conditions, most deaths caused by TB are preventable. Through combinations of antibiotic treatment plans, TB is known to have a high global treatment success rate of 85% according to the World Health Organization (WHO) report of 2020.
But why are millions still dying?
Overview of The Diagnostic Gap in TB Detection
Vaccine research, drug development, and diagnostic methods for TB have been neglected for decades. In fact, the BCG (Bacillus Calmette–Guérin) vaccine — developed over 100 years ago — is the only vaccine available for TB with an effectiveness of around 50%. Bedaquiline, a generic antibiotic drug discovered by Johnson & Johnson for the treatment of multidrug-resistant TB received FDA approval in 2012; however, the pharma giant has enforced patents to this day, preventing the drug’s wider distribution to lower-income countries — countries in need of the drug the most.
In regards to detection however, the healthcare infrastructure's inability to support TB detection has contributed to millions of deaths for decades and even now. Spending on TB diagnostic, treatment and prevention services fell from $5.8 billion to $5.3 billion, which is less than half of the global target for fully funding the tuberculosis response of $13 billion annually by 2022. Before COVID-19, approximately 1.4 million people were diagnosed with TB in sub-Saharan Africa alone in 2019; however, epidemiologists estimated that 1 million more had TB but were neither diagnosed nor treated. With the death rate for untreated TB being approximately 50%, the impacts of insufficient disease detection are severe and noticeable even within a short time period.
Furthermore, on a global scale, the impacts of disruptions to essential detection services are directly correlated with the increase in deaths caused by TB.
An example is the recent increased TB mortality count over the past few years, which is consistent with the World Health Organization (WHO) projections in the Global Tuberculosis Report 2021. It reflects the impact of disruptions to essential TB detection services due to the COVID-19 pandemic.
However, there is another component/ gap that makes TB infections unique and different when compared to other infectious diseases, and that is its dependency on early detection. Because of TB’s two main infection types — latent and active — it is one thing to detect TB, but it is another to detect it early on.
For those with latent TB infections (one-third of the global population), there are only trace amounts of Mycobacterium tuberculosis (M. tb), which are forced into a quiescent state after the process of phagocytosis — the body’s immediate immune response. This stage can last for years or up to an entire lifetime where the disease is asymptomatic and non-infectious.
However, around 5–10% of latent TB infections develop into active TB infections where M. tb manipulates the macrophages they were encapsulated in to undergo necrosis — a type of cell death that allows the bacteria to infect neighboring cells. Unlike latent TB, the pulmonary and extrapulmonary manifestations of active TB may cause serious symptoms (e.g., chest pain, breathing difficulty, blood coughs, etc.) that can become fatal in many cases. As of 2021, 1.6 million deaths from TB were recorded.
Active TB is typically diagnosed based on clinical symptoms and the results yielded from certain TB detection methods. On the other hand, latent TB is currently only diagnosed when there is a positive reaction to TB detection tests without clinical symptoms.
Unfortunately, the current lack of focus on early detection has had far-reaching implications on public health and is largely responsible for millions of deaths. With current diagnostic methods for TB detection, it is difficult to differentiate between latent and active TB infections, and nearly impossible to predict which patients with latent TB infections will develop an active TB infection and which patients will not. The struggle to do so prevents our ability for early disease treatment and collectively forms the diagnostic gap in TB.
Current Diagnostic Methods for TB Detection
Sputum Smear Microscopy — Examines for the presence of acid-fast bacilli (a trait of M. tb) in sputum sample stains, typically used in resource-limited settings. However, the sensitivity index ranges from 50% to 70%; requires larger sample loads, is less accurate compared to newer molecular methods; and cannot differentiate between latent and active TB.
Sputum Culture — Inoculating and cultivating sputum cultures in a laboratory for identification of active pulmonary TB and drug susceptibility testing. It has an average sensitivity of over 80% and is highly accurate; however, the culturing process takes several weeks to yield results and does not distinguish between latent and active TB.
GeneXpert (Xpert MTB/RIF) — Utilizes polymerase chain reaction (PCR) techniques to detect specific genetic sequences of M. tb and the rifampin resistance gene (rpoB). It has a high sensitivity of approximately 90%, is highly accurate, and provides results within a few hours. However, this method cannot differentiate between latent and active TB and requires advanced healthcare infrastructure that may be costly.
Chess X-ray & CT Scan — Utilizes medical imaging technology to analyze biomarkers such as cavities, infiltrates, nodules, and more that are indicative of TB infections. The sensitivity of these methods is subjective to the disease stage and does not differentiate between latent and active TB.
Line Probe Assays (LPA) — Detects specific genetic markers associated with M. tb and its resistance to anti-TB drugs through DNA hybridization. They have an average of 90% sensitivity for detecting pulmonary active TB and drug resistance biomarkers; however, they are unable to distinguish between latent and active TB.
Tuberculin Skin Test — Injects small amounts of purified protein derivative (PPD) called tuberculin under the skin, measuring the delayed-type hypersensitivity response to TB antigens and ultimately detecting previous exposure to TB. However, this test can be falsely positive in patients vaccinated with BCG and may be falsely negative in those who are immunocompromised. In addition, this test does not confirm pulmonary TB nor is able to differentiate between latent and active TB.
Interferon-Gamma Release Assays (IGRAs) — Measures the release of interferon-gamma by white blood cells when exposed to TB antigens, ultimately indicating TB infections. They are more accurate and less affected by BSG vaccinations as compared to tuberculin skin tests; however, IGRAs do not confirm active disease nor distinguish between latent and active TB.
A noticeable commonality for all current TB detection methods is that they are unable to differentiate between latent and active TB infections. But more importantly, there are currently no detection methods to determine or predict which latent TB infections will progress to active TB infections.
Diagnostic Potential of AI in Active TB Prediction
To address the current absence of any TB prediction methods, the potential implications of using AI to predict TB pose incredible impacts. Predictive AI models’ capacity to process large amounts of data quickly and identify relationships within the data given makes it a potential diagnostic tool for the prediction of active TB infections. By feeding a combination of risk-related patient data and specific biomarkers to an AI model, physicians can determine which latent TB patients will likely develop an active TB infection and which latent TB patients will not.
Scientifically-validated Risk Assessment and Diagnoses for LTBI Treatment
Upon determining the likelihood of developing active TB infections, patients would currently be able to receive preventative treatment for latent TB infections (LTBI) through antibiotic medications such as Isoniazid, Rifapentine, or Rifampin exclusively or in combination to kill sleeping M. tb.
However, as with many other antibiotic treatments, low initiation and completion rates affect the effectiveness of LTBI treatment. In the Ottawa Hospital TB Clinic, a seven-year study was conducted from 2010 to 2016 to ascertain non-initiation and non-completion of preventative TB treatment.
The graph above represents the percentage of patients remaining in the cascade compared to the number of individuals screened for LTBI treatment. The incredibly low completion rate of LTBI treatment was associated with the lack of “initial consultation to increase understanding of LTBI” and the patient’s uncertainty of diagnoses as they lack scientific validation.
Utilizing AI to determine the risk levels of active TB development involves processing tangible factors — risk-related data and specific biomarkers — to form results rather than the current method of “yes symptoms or no symptoms.” This will not only reduce the amount of incomplete LTBI treatment but also assist physicians with the decision-making process for need-based diagnoses.
Currently, physicians make judgments just from positive LTBI indications and basic immunosuppression factors. This method of assessing whether an LTBI patient should be treated with antibiotic courses (usually ranging from 3–9 months) or not has room for massive improvement. Therefore, determining the probability of active TB development with the help of AI can remove diagnostic uncertainties and drastically increase the rate of LTBI treatment completion for patients.
Long-term Benefits and Reduced TB Transmission Rates
Like many other infectious diseases, TB imposes a substantial burden on both the healthcare system and the local economy. As early detection and treatment through predictive AI models result in reduced active TB cases, healthcare expenditures can be significantly decreased while hospital admissions and other healthcare burdens are lowered. In addition, the prevention of active TB cases reduces the number of people unable to attend work due to the disease’s symptoms.
Furthermore, as TB infections can only be transmitted during active stages, the prevention treatments collectively reduce transmission rates of the infectious disease.
Public Health Surveillance and Control for TB
In the pursuit of achieving international TB control goals, predictive AI models can play a pivotal role in identifying high-burden regions and populations, enabling targeted allocation of resources and precise targeting strategies for TB control.
By analyzing patient data derived from the AI models, organizations like the World Health Organization (WHO) may gain a more comprehensive understanding of how TB spreads within communities, the multifaceted factors that influence its transmission, and the effectiveness of certain policies/ initiatives. This improved understanding of TB pathogenesis, epidemiology, and risk factors has the potential to lead to the development of more efficient prevention and control strategies, ultimately working towards the United Nation’s Sustainable Development Goals.
