MOOVE is a large-scale, participatory evaluation platform for clinical AI systems, designed to produce evidence that is directly usable in real-world health systems. It addresses a core gap in medical AI: the lack of rigorous, context-aware evaluation of large language models under the clinical, linguistic, and resource constraints where they are intended to be used.

MOOVE is built with clinicians and health institutions across Sub-Saharan Africa, South Asia, Latin America, and Europe, and focuses on high-stakes decision-making in both resource-constrained and high-resource settings.

RetroMOOVE is a research project focused on building large-scale evaluation benchmarks for medical AI using retrospectively collected, real-world clinical data. The project uses routinely collected healthcare data from multiple countries, including Tanzania, Rwanda, Kenya, and others, to construct realistic evaluation settings that reflect actual clinical workflows, data quality, and decision-making constraints.

By grounding evaluation in retrospective real-world data, RetroMOOVE enables systematic comparison of large language models against established clinical baselines, moving beyond synthetic prompts or isolated benchmarks.

The project currently includes large-scale retrospective datasets across multiple clinical domains and geographies:

– Pediatric primary health care in Tanzania and Rwanda
– Maternal and child health across Kenya, India, and other Sub-Saharan African settings
– Mental health using retrospective datasets from Kenya and South Africa

Meditron-4 is the next iteration of Meditron, focused on advancing clinically aligned, guideline-faithful, and context-aware medical language modeling.

The project aims to develop open-source fine-tuning and evaluation pipelines for medical LLMs, and to produce a highly clinically aligned Meditron model built on leading open medical and general base models. In parallel, the work explores small, offline-capable models (e.g. MedGemma 4B, LFM-2) to support deployment in resource-constrained settings.

This project focuses on improving clinical reasoning capabilities in Meditron by studying explicit reasoning mechanisms in medical language models. In high-stakes clinical settings, correct answers are not sufficient on their own; models must demonstrate coherent, grounded reasoning that aligns with clinical logic, evidence, and guidelines. This project treats reasoning as a measurable, evaluable property rather than an emergent side effect.

The work explores how reasoning-oriented training methods can improve decision consistency, error detection, and robustness in medical language models.

Most medical language models are trained and evaluated primarily in English, despite the fact that clinical care is delivered in many languages, often in settings where English is not the working language. In low-resource and multilingual health systems, this mismatch limits the usefulness and safety of AI tools.

PolyglotMeditron focuses on extending Meditron to support multilingual clinical reasoning, with an emphasis on low-resource languages and real clinical usage. The project studies how medical meaning, terminology, and decision logic transfer across languages, and where existing multilingual models fail in medical contexts.

MultiMeditron focuses on making Meditron multimodal, allowing users to provide medical images in addition to text. The aim is to extend Meditron beyond text-only inputs while preserving its existing structure and capabilities.

The work is twofold. First, the Meditron codebase is adapted to support a multimodal architecture and new input modalities (currently limited to images). Second, modality-specific “expert” models are developed and improved to process these inputs and generate embeddings that are fed into Meditron.

Voice carries clinically meaningful information relevant to respiratory disease, mental health, neurological conditions, and general clinical assessment. While large speech foundation models show strong general capabilities, their reliability, robustness, and data requirements in real clinical settings remain poorly characterized. This project focuses on rigorously evaluating speech foundation models for clinical voice-based assessment, with particular attention to performance under realistic, resource-constrained conditions.

The goal is to determine when such models are useful, where they fail, and what constraints must be addressed before they can be safely integrated into health systems.

This project focuses on training a Vision Transformer (ViT) for medical imaging using unsupervised or self-supervised learning objectives (e.g. DINO-style methods) on large-scale medical image datasets.

The work emphasizes reproducing a state-of-the-art approach, building a clean and extensible training codebase, and running large-scale training experiments suitable for downstream medical imaging tasks.

Tuberculosis and pediatric pneumonia remain major causes of morbidity and mortality in high-burden, resource-constrained settings. This project focuses on building multimodal AI systems that are explicitly designed to operate under real-world clinical, infrastructural, and governance constraints, rather than idealized research conditions.

The work integrates lung ultrasound video, structured clinical data, and metadata, and emphasizes robustness to variation across operators, sites, and patient populations. All models are evaluated with deployment in mind, including explicit analysis under domain shift and contribution to ongoing prospective clinical trials.

This project studies model quantisation for medical large language models in practice, with an emphasis on careful, reproducible experimentation. Quantisation is a key enabler for deploying medical LLMs in resource-constrained settings, where memory, compute, power, and latency limitations are critical. Understanding how quantisation affects clinical model behavior is essential before such models can be used safely in real-world health systems.

This project focuses on designing and evaluating edge-deployed AI systems for maternal and child health decision support in Zanzibar. In this setting, community health workers operate with limited connectivity, constrained hardware, and high responsibility for frontline clinical decisions. Cloud-dependent AI systems are often impractical due to latency, cost, privacy, and reliability constraints.

The project studies how small language models, retrieval-augmented generation, and multimodal inputs can be combined into on-device systems that support context-aware reasoning using locally available clinical guidelines, patient records, documents, images, and speech.

In real-world healthcare settings, multimodal data are frequently incomplete. Imaging, laboratory results, clinical notes, or sensor data may be missing due to cost, infrastructure, workflow, or patient factors. Models that assume complete modality availability often fail when deployed. This project addresses that gap by studying how to build and evaluate multimodal models that remain valid when data modalities are partially or entirely missing.

The focus is on robustness and generalization across diseases and datasets, reflecting the conditions under which clinical AI systems are actually used.

This project studies knowledge distillation for medical large language models, with the goal of producing smaller, more efficient models that retain clinically relevant behavior. Distillation is a key technique for deploying medical LLMs in settings with limited compute, memory, or connectivity, but its effects on clinical performance and failure modes are not well understood. This project treats distillation as a scientific problem rather than a purely engineering shortcut.

This project is centered on in-depth literature review in artificial intelligence for healthcare. Participants select a focused topic aligned with the interests of the lab and conduct a structured, critical review of the relevant research literature.

The primary outcome is a well-written literature review on a defined topic, alongside regular contributions to the collective knowledge of the group. Topics may range from theoretical foundations of machine learning and modeling choices to applied questions around evaluation, deployment, safety, and impact of AI in health.

The LiGHT Bootcamp is an online, certificate-based MOOC web platform for hands-on training in interdisciplinary AI skills applied to healthcare. It is designed for ministries of health, policymakers, and clinicians working in low-resource settings.

The platform combines structured online content with mentored, practice-oriented learning that is continuously updated to reflect advances in AI and healthcare. Participants engage with applied tools, real-world use cases, and contemporary methods, with an emphasis on interactive, responsive, and engaging learning.

The bootcamp is developed and delivered in collaboration with ministries of health and clinical stakeholders, with the goal of providing sustained access to up-to-date knowledge, practical tools, and best practices for responsible AI use in health systems.
Students may work on this project either as a side project (1 or 2 hours per week of mentoring) or as a full time semester/optional project (there are many ways you can help improve the MOOC, check the "Research questions/tasks" below)

MMORE is LiGHT’s core research library for multimodal document intelligence. It provides the infrastructure required to process, index, retrieve, and extract structured information from complex documents that combine text, layout, tables, figures, and images. Such documents are ubiquitous in clinical and regulatory settings, yet remain difficult to handle reliably at scale. LiGHT started developing a new library, MIRAGE, that makes it way easier to process datasets either with code or with an LLM. There is no longer a need to write a processing script for each dataset, you can simply write a prompt for an LLM or a code snippet that will be applied to each sample of the dataset, MIRAGE handles the rest, including parallelization.

These libraries underpin key LiGHT research infrastructures, including MOOVE and Meditron, and are used in settings where robustness, traceability, and scalability are essential. The project aims at maintaining and developing Python libraries that are useful for data science work in the lab (and for others), allowing to focus the work on data and evaluation.