Generating evidence and developing deployable AI technologies for global health and humanitarian response
We build AI systems that work where they are needed most by prioritizing openness, real-world validation, and local relevance.
This means collaborating with users at every stage: from defining use cases to integrating tools into clinical workflows, while ensuring models are trained on high-quality, contextually grounded data.
We emphasize open-source development, ethical governance, and capacity building, designing systems that can operate under real constraints such as limited data, connectivity, and infrastructure, and that can be sustained and owned locally.
3
Contextualized Models
2
Privacy-preserving distributed learning
1
Real world
clinical studies
4
Clinician-centered AI applications
MoBaye
MoBayes constructs an explicit clinical knowledge base, then iteratively gathers evidence through LLM-parsed patient dialogue and updates beliefs with Bayesian inference — outperforming much larger standalone LLM doctors at a fraction of the cost. The LLM is confined to parsing and verbalization, while a deterministic Bayesian module handles posterior tracking, question selection, and calibrated abstention.
MultiModN
MultiModN explores modular neural architectures for multimodal, multi-task learning, with a focus on interpretable fusion across heterogeneous data sources.
The project introduces a flexible approach to combining modalities sequentially, supporting robust prediction even when some information is missing not at random.MoDN
Modular clinical prediction models that operate with incomplete patient information and fragmented healthcare records.
IMCI-Plus
An interdisciplinary, Pan-African-European Union research initiative which aims to improve the management of childhood pneumonia by pioneering the use of point-of-care lung ultrasound in everyday clinical practice, and policy.
MOOVE
MOOVE (Massive Open Online Validation and Evaluation) is a global, expert-led initiative that enables clinicians, humanitarian actors, and local experts to evaluate AI models against the realities of their own settings. Rather than a single study or platform, MOOVE is a progressive clinical validation pathway: partners begin with hypMOOVE, where experts evaluate AI using hypothetical clinical scenarios; advance to silentMOOVE and retroMOOVE, where AI is evaluated on real patient data without influencing clinical decisions; and, when sufficient evidence has been established, progress to trueMOOVE, where AI is evaluated in randomized controlled trials. Across all stages, MOOVE generates expert preference data, context-adapted models, locally owned datasets, and research-grade publications.
The MOOVE programme currently comprises MOOVE Africa (Tanzania, Rwanda, Kenya, Malawi, and Ethiopia), MEDUSE (Switzerland), and MOOVE India, advancing AI evaluation and clinical validation across diverse healthcare settings.
MoBayes
MoBayes constructs an explicit clinical knowledge base, then iteratively gathers evidence through LLM-parsed patient dialogue and updates beliefs with Bayesian inference — outperforming much larger standalone LLM doctors at a fraction of the cost. The LLM is confined to parsing and verbalization, while a deterministic Bayesian module handles posterior tracking, question selection, and calibrated abstention.
MultiModN
MultiModN explores modular neural architectures for multimodal, multi-task learning, with a focus on interpretable fusion across heterogeneous data sources. The project introduces a flexible approach to combining modalities sequentially, supporting robust prediction even when some information is missing not at random.
DISCO
DIStributed COllaborative Learning, Train AI Models Together. Keep Data Private. Build and train AI models without sharing any data. Machine Learning directly in your browser.
AI POCUS
An international collaborative network that brings together clinicians, researchers, developers, implementers, policymakers, donors, industry, and health organizations to accelerate the safe, equitable, and evidence-based adoption of AI-enabled point-of-care ultrasound (AI-POCUS) in low- and middle-income countries (LMICs).
MOOVE
MOOVE is a global, expert-led platform that enables clinicians and humanitarian professionals to rigorously evaluate AI systems against real-world local healthcare contexts, generating trusted evidence on safety, quality, and contextual relevance. By combining community governance, data sovereignty, and continuous validation, MOOVE helps adapt AI models to the populations and settings they are intended to serve, particularly in underserved and low-resource environments.
MoBaye
MoBayes constructs an explicit clinical knowledge base, then iteratively gathers evidence through LLM-parsed patient dialogue and updates beliefs with Bayesian inference — outperforming much larger standalone LLM doctors at a fraction of the cost. The LLM is confined to parsing and verbalization, while a deterministic Bayesian module handles posterior tracking, question selection, and calibrated abstention.
MultiModN
MultiModN explores modular neural architectures for multimodal, multi-task learning, with a focus on interpretable fusion across heterogeneous data sources.
The project introduces a flexible approach to combining modalities sequentially, supporting robust prediction even when some information is missing not at random.MoDN
Modular clinical prediction models that operate with incomplete patient information and fragmented healthcare records.
IMCI-Plus
An interdisciplinary, Pan-African-European Union research initiative which aims to improve the management of childhood pneumonia by pioneering the use of point-of-care lung ultrasound in everyday clinical practice, and policy.
MOOVE
MOOVE (Massive Open Online Validation and Evaluation) is a global, expert-led initiative that enables clinicians, humanitarian actors, and local experts to evaluate AI models against the realities of their own settings. Rather than a single study or platform, MOOVE is a progressive clinical validation pathway: partners begin with hypMOOVE, where experts evaluate AI using hypothetical clinical scenarios; advance to silentMOOVE and retroMOOVE, where AI is evaluated on real patient data without influencing clinical decisions; and, when sufficient evidence has been established, progress to trueMOOVE, where AI is evaluated in randomized controlled trials. Across all stages, MOOVE generates expert preference data, context-adapted models, locally owned datasets, and research-grade publications.
The MOOVE programme currently comprises MOOVE Africa (Tanzania, Rwanda, Kenya, Malawi, and Ethiopia), MEDUSE (Switzerland), and MOOVE India, advancing AI evaluation and clinical validation across diverse healthcare settings.
MoBayes
MoBayes constructs an explicit clinical knowledge base, then iteratively gathers evidence through LLM-parsed patient dialogue and updates beliefs with Bayesian inference — outperforming much larger standalone LLM doctors at a fraction of the cost. The LLM is confined to parsing and verbalization, while a deterministic Bayesian module handles posterior tracking, question selection, and calibrated abstention.
MultiModN
MultiModN explores modular neural architectures for multimodal, multi-task learning, with a focus on interpretable fusion across heterogeneous data sources. The project introduces a flexible approach to combining modalities sequentially, supporting robust prediction even when some information is missing not at random.
AI POCUS
An international collaborative network that brings together clinicians, researchers, developers, implementers, policymakers, donors, industry, and health organizations to accelerate the safe, equitable, and evidence-based adoption of AI-enabled point-of-care ultrasound (AI-POCUS) in low- and middle-income countries (LMICs).
MOOVE
MOOVE is a global, expert-led platform that enables clinicians and humanitarian professionals to rigorously evaluate AI systems against real-world local healthcare contexts, generating trusted evidence on safety, quality, and contextual relevance. By combining community governance, data sovereignty, and continuous validation, MOOVE helps adapt AI models to the populations and settings they are intended to serve, particularly in underserved and low-resource environments.
.png)
Where we work
Where we work
EPFL, School of Computer Science, Switzerland
Harvard, Ariadne Labs, Harvard T.H Chan School of Public Health, USA
Ashoka University, Koita Center for Digital Health, India
C4IR, Africa AI Scaling Hub, Rwanda
PROJECT HIGHLIGHTS
PrAlmaan MOOVE
The Massive Open Online Validation and Evaluation (MOOVE) is a multi-country platform for evaluating generative AI-enabled clinical decision support tools, involved in large-scale randomized controlled trials in Africa. As of 2026, MOOVE has also launched in India in the PrAImaan project. PrAimaan will establish a centralized yet federated authority under ICMR to govern health AI evaluation; define national standards; map India’s landscape of models, datasets, and compute environments; and build a Technical Facilitation Unit to adapt and operate MOOVE for Indian needs.
​
Collaborators: ICMR, Koita Centre for Digital Health at Ashoka
PROJECT HIGHLIGHTS
ChitChat
ChitChat develops frameworks and methods to ensure AI aligns with the International Committee of the Red Cross ethical values, including evaluation tools for LMLMs and a human-feedback web platform for continuous alignment.
The project is funded through the Engineering for Humanitarian Action initiative, a partnership between the ICRC, École polytechnique fédérale de Lausanne and ETH Zurich.