Tuberculosis and AI
Artificial Intelligence may contribute to combatting tuberculosis; the leading cause of death of HIV victims.
Introduction
Tuberculosis is an airborne infectious disease, caused by the bacterium mycobacterium tuberculosis. In 2017 alone there were around 1.3 million related deaths, and TB is the leading killer of HIV patients. Applications of AI in both detecting the bacterium, and ensuring medication protocols are adhered to have started to be developed.
Tuberculosis in the body
TB is present in 1/3 of the global population, but predominantly in a form known as ‘latent TB’, which is when a person carries the bacteria but does not develop symptoms. This form is not contagious, but if the immune system is compromised, such as from HIV, latent TB can develop int active TB, in which case the patient will become contagious, and experience symptoms of the infection.
TB enters the body by invading the mucus in the respiratory system and entering the alveoli. After having been engulfed by macrophages (a type of white blood cell) in the immune system, TB produces a protein which inhibits the fusion of the phagosome (a vesicle within the macrophage) containing TB with a lysosome (lysosomes contain hydrolytic enzymes, which digest pathogens), allowing the mycobacterium to survive, unlike most pathogens, and proliferate.
Systemic miliary TB occurs when the infection spreads throughout the body - it affects many different organs, including the meninges of the brain, leading to meningitis; the adrenal glands, leading to Addison’s disease; and the liver, leading to hepatitis.
Detection of tuberculosis
Skin test: a fluid called tuberculin is injected into lower arm - if swelling or bump detected within 48-72 hours.
Blood test: analyse sample of blood to detect if there has been an immune response to Mycobacterium tuberculosis.
Chest x-ray: can be done to check for abnormalities.
Samples from the sputum / bronchoalveolar lavage collected from patients with symptoms (fevers, night sweats, coughing up blood, weight loss) - samples can be analysed to detect the bacterium.
Detection of mycobacterium tuberculosis with AI
Bacilli are very small, and detection can be a strenuous and time-consuming job, and often experienced pathologists fail to detect the presence of TB – hence, the possible role for AI.
It has proved difficult providing AI with precise parameters of the morphological features of mycobacterium tuberculosis, and AI has not currently learned the all the different forms in which the bacillus can be found, so AI diagnosis requires the confirmation from a human pathologist.
The Department of Pathology at Peking University First Hospital ran a study from January 2016 to June 2017, in which the diagnosis rate by TB-AI was 76.6% in accordance with confirmation by pathologists on the first run. In the second run, after improvements had been made to the model, TB-AI scored 83.65% on specificity, and 97.94% on sensitivity to TB bacilli, in comparison to diagnoses by pathologists with both microscope and digital slides.
Monitoring medication adherence to TB treatment with AI
Lack of adherence to antibiotic courses can lead to increased drug resistance, as well as prolonged infection and the related consequences. Certain areas of the world, suffer from a scarcity of medically trained professionals, leading to the introduction of VDOT (video-based directly observed therapy), which was officially recognised for monitoring TB treatment by WHO in 2017. Both live and pre-recorded videos can be reviewed by healthcare workers, to monitor patients’ medication intake, overcoming challenges facing patients living in remote areas. However, VDOT is a time-consuming, monotonous task for people, and tiredness can lead to poor assessment quality.
A study conducted on patients in Kampala, Uganda, tested the ability of AI to monitor medication adherence, after having utilised deep-learning models to teach AI visual signs such as facial gestures, which would correspond with taking / not taking medication. The study recorded an accuracy in diagnosis of 72.,5% to 77.3%, which is similar to that of a doctor.
Bibliography:
https://www.youtube.com/watch?v=v_j-LD2YEqg
https://www.youtube.com/watch?v=0fVoz4V75_E
https://www.medicalnewstoday.com/articles/8856
https://www.youtube.com/watch?v=6P6zBHpWiGA
What do you think about using AI to detect and monitor TB treatment? Do you believe this could result in better patient outcomes? Share your thoughts below!