Aim

To develop and test an AI-supported point-of-care ultrasound (AI-POCUS) to improve the reach and quality of tuberculosis screening in vulnerable adults and children through an international collaboration between researchers, developers, clinicians, communities affected by TB and policy makers.

Objectives

1. To develop a network of multi disciplinary researchers, clinicians, communities affected by  TB, industry partners and policy makers to co-design and evaluate AI-POCUS for TB screening in children and adults
2. To evaluate and optimise different POCUS acquisition protocols in different epidemiological contexts
3. To broaden understanding around the ethics of data sharing in AI and pilot a collaborative machine learning approach
4. To evaluate the clinical performance (sensitivity, specificity and predictive value), acceptability and costs of implementing AI-POCUS
5. To maximise dissemination of research findings and uptake of an AI-POCUS
6. To build capacity for inter-disciplinary research in four low and middle-income countries (LMIC)

This research aims to develop a simple, cheap imaging solution to screen for tuberculosis (TB), the world’s leading infectious disease killer. The World Health Organisation has set a target to "End TB” by 2035, and developed a strategy to help countries achieve this. One of the key activities in this strategy is to find and treat more people with TB. Current screening guidelines include chest x-ray, but this is inaccessible to many of the poorest people worldwide who are most at need of screening for TB. Ultrasound may provide a solution. Ultrasound is a simple and safe procedure that can be used for adults and children. Portable ultrasound probes (“Point-of-care” ultrasound, POCUS) that can connect to mobile phones are available and could be placed at the lowest levels of the health care system in countries most affected by TB. Initial research is promising, but highly trained health workers able to interpret ultrasound images are not widely available. Therefore, artificial intelligence (AI) could provide a solution.

This Global Health Research Group (GHRG) will bring together researchers from four LMIC with different TB epidemiology contexts; Mozambique and Zambia (high burden of TB/HIV), Bolivia and Peru (high burden of MDR-TB), alongside a medical imaging partner to develop an AI-supported POCUS tool.

The project will start with co-design of the AI-POCUS with communities affected by TB, health care providers, and regulators in these settings to understand what they would like from such a product. We will conduct clinical research to determine whether POCUS is an effective and practicable tool for TB screening including which scanning protocol provides the best images, yet can be easily taught and does not overburden health workers or patients, and to start collecting clinical, microbiological and imaging data.

To develop the AI reading, large numbers of images and data are needed, however there can be challenges with data privacy and ownership if data is transferred to a central source. The GHRG will therefore conduct research around the broader ethics of data privacy and transfer, and develop capacity to pilot a “federated machine learning approach”, which involves all sites retaining full ownership of their data and concurrently training a general machine learning model on-site, before combining these models into a centralised one.

The finalised AI-supported POCUS product will then be tested across the four epidemiologically different sites in adults and children being investigated for TB. We will assess the clinical performance and the costs of the system.

The GHRG will finally convene a workshop with communities affected by TB and policy makers to present the findings and discuss the roadmap to the further development of the product. We will also produce a training package for health workers to facilitate adoption beyond the facilities and countries in this GHRG. If successful, this product has the potential to be a game-changer for screening of TB in high-burden countries by dramatically increasing access to better quality and co-ordinated TB diagnosis amongst the world’s population that are both the most vulnerable to TB and face the greatest barriers to being diagnosed.

Work packages

WP1 - Workshop consortium

The first activity of the consortium will be a workshop of all consortium members, who will then be joined by other researchers, communities affected by TB, plus international and national policy makers to discuss the current knowledge base around POCUS for TB screening and ensure research and policy gaps are identified. Throughout the project coordination between the different work packages and across the sites will be achieved by regular virtual group meetings and cross-site interaction. WP1 will ensure high standards of research governance are maintained including ethical and regulatory compliance and monitor progress against milestones.

WP2 - Formative research on POCUS

The first study will be qualitative research with TB patients Health Care Workers (HCW), and regulators to understand how POCUS tools could best be deployed, including the optimal user and patient experience. Clinical research with both adults and children will identify optimal POCUS acquisition protocols, including combining lung, FASH and mediastinal views, to obtain and compare images with clinical data and microbiological results. Several acquisition protocols have been proposed but a balance between image quality, training complexity, and HCW and patient time needs to be evaluated. The research will identify whether protocols should differ in children and adults, and based on HIV or other clinical factors. 

WP3 - Artificial Intelligence

Machine learning, and in particular deep learning, have become a major trend in medical image analysis. In order to have good clinical applicability large medical datasets derived from various sites are needed. However, the current centralised learning techniques raise concerns about patient privacy, confidentiality, security, and data ownership. Even in anonymized datasets, an individual could be re-identified correctly with a high likelihood. Prof. Bond, a social anthropologist, will supervise a PhD student to conduct research around the broader ethics of data ownership and privacy in medical research. Ir. Swinkels, a deep learning engineer, will lead a pilot of federated learning, a potential solution to develop the AI model. Federated learning may solve data privacy and ownership issues by training a consensus model collaboratively between different decentralized datasets. The sites will be capacitated to manage their own data to train the model locally. Thereafter, updated model parameters will be sent to a central server that aggregates the separately trained local models to generate a new global model. This new model could be personalized for each client to include data heterogeneity.

WP4 - Evaluation of clinical performance, acceptability and economic effectiveness of AI-POCUS

Data from WP2 and WP3 will be used to develop an AI-POCUS product that will be evaluated in the different epidemiological situations provided by the sites (TBHIV adults Zambia, children Mozambique, non-HIV TB adults Peru, children Bolivia). WP4 will develop unified tools and standardised outcomes for the protocols across the sites to enable evaluation within and across countries. Clinical performance will be assessed by measuring the sensitivity, specificity and predictive value of the AI-POCUS compared to standardised endpoints, in the hands of intended users. Qualitative research will assess usability and acceptability, with an economic evaluation to assess costs. 

WP5 - Dissemination and uptake

To ensure uptake of the proposed AI-POCUS, tools and training materials will be needed for HCW to ensure that the system can be deployed. We will work with communities and HCW to co-develop a training package based on the final acquisition protocol. WP5 will engage with regulators and policy makers to communicate findings and understand pathways to inclusion of the tool in screening guidelines. 

WP6 - Research capacity development

Underpinning the research activities of the Global Health Research Group (GHRG) will be development of research capacity including one PhD in each country engaged to conduct different research studies. The PhDs will include one social scientist to look at broader issues of ethics in AI, a clinician who will focus on the evidence needed for the AI-POCUS model, a computer scientist and an implementation researcher. Institutional capacity development will include development of infrastructure and data skills for federated learning as well as broader research and finance governance tasks.