Monitoring medication adherence for TB treatment in Africa using AI

By Dr. Priyom Bose, Ph.D.Reviewed by Danielle Ellis, B.Sc.Apr 11 2022

It has been estimated that 1.7 million people die from Tuberculosis (TB), and more than 10.4 million new cases are reported every year worldwide. The global 'End TB' strategy aims to eliminate the disease by 2030. However, realizing this goal would be challenging if there were to be a gap in treatment adherence to prescribed medication.

This news article was a review of a preliminary scientific report that had not undergone peer-review at the time of publication. Since its initial publication, the scientific report has now been peer reviewed and accepted for publication in a Scientific Journal. Links to the preliminary and peer-reviewed reports are available in the Sources section at the bottom of this article. View Sources

Background

In the context of TB and HIV coinfection, non-adherence to the medication has been associated with the incidence of drug resistance, prolonged infection, unsuccessful treatments, and death. Africa experiences a severe shortage of healthcare workers, making delivering proper healthcare difficult.

The recent application of digital adherence technologies (DATs) has helped improve healthcare services substantially. In 2017, the World Health Organization recognized the use of video-based directly observed therapy (VDOT) as an appropriate alternative to DOT for monitoring TB treatment. VDOT has played an important role in monitoring the adherence to TB treatments, as it enables health providers to monitor patients' medication intake activity directly through synchronous or asynchronous recording. One of the key advantages of VDOT is that it overcomes the challenges of geographical locations by presenting an opportunity to healthcare providers to reach out to individuals in remote areas.

Asynchronous VDOT requires human effort to review videos and determine the medicine intake practices of individuals. However, the task of manual review is often monotonous and can get repetitive. There is a high risk of inaccurate assessment owing to human fatigue when the workload is extremely high. This is the reason why the application of artificial intelligence (AI) could be a logical step to obtaining a better result.

Researchers have stated that the application of AI in the healthcare domain has the potential to transform several clinical practice areas, such as medical imaging. This technology has significantly enhanced the efficacy of care delivery by appropriately arranging workflows in the healthcare system. One of the key advantages of utilizing AI has been faster delivery of care and optimal management of limited resources.

Previous studies have shown that modern computer vision techniques in combination with deep learning convolutional neural networks (DCNNs) could be applied in developing medical videos, medical imaging, and clinical deployment. Scientists expressed that deep learning techniques could be utilized to effectively and efficiently monitor TB. However, implementation of deep learning methods has been limited due to a lack of access to large, well-curated, and labeled datasets. Additionally, the lack of technical skillset required to develop deep learning models in most healthcare professionals makes the application of deep learning in the healthcare setting difficult.

A new study

A new pilot study, available on Preprints with The Lancet*, has focused on determining the technical feasibility of applying AI to analyze a raw dataset of videos from TB patients taking medications. This study was conducted by a multidisciplinary team led by a public health physician specializing in TB medication adherence and three computer scientists specializing in deep learning models. In this study, researchers aimed to develop an AI system that can evaluate medication adherence and non-adherence activities of TB patients based on their visual attributes obtained from videos, such as facial gestures and jaw-drop.

In this study, researchers used a secondary dataset containing 861 self-recorded medication intake videos of 50 TB patients. These videos were intended for VDOT. The study cohort consisted of both male and female patients between 18 and 65 years with a confirmed diagnosis of TB. All the patients attended public clinics in Kampala, Uganda, and their socio-demographic characteristics were recorded.

Key findings

Researchers tested several deep learning models and found that the 3D ResNet performed effectively at an AUC of 0.84 and a speed of 0.54 seconds per video review. They observed a diagnostic accuracy ranging from 72.5% to 77.3%, which is comparable to or higher than the expert clinical accuracy of doctors

In this study, all the DCNN models exhibited comparable discriminative performance to state-of-the-art performing deep learning algorithms. This finding supports the utility of deep learning models in the binary classification of medication video frames to predict adherence. Scientists stated that this is an important step for building more effective models with relevant applications.

Conclusion

One of the study's limitations is the inability to incorporate all the recommended methodological features for clinical validation of AI performance in real-world practice. However, the authors stated that the high performance of the deep learning models, especially the 3D ResNet model, emphasizes the power of AI tools in monitoring medication in a drug efficacy trial. Scientists stated that the classification accuracy of DCNN models in medication adherence should be improved along many dimensions in the future, including the open-sourcing of large labeled datasets to train the algorithms.

This news article was a review of a preliminary scientific report that had not undergone peer-review at the time of publication. Since its initial publication, the scientific report has now been peer reviewed and accepted for publication in a Scientific Journal. Links to the preliminary and peer-reviewed reports are available in the Sources section at the bottom of this article. View Sources