Hard problems in neurodegeneration

Multimodal AI for neurodegeneration

The challenge

The questions that matter most are the hardest to answer during life:

• Which diseases are present, and in what mix?
• Is amyloid or tau present, and where has it spread?
• Is someone who feels fine already declining?

Today the answers need PET scans, spinal taps or autopsy, or they come years too late. Clues exist in MRI, cognitive tests, medical history, genetics and blood markers, but no patient has all of them, and every clinic and cohort collects a different subset on different equipment. Most AI assumes complete, uniform data, so it breaks down where it's needed most. The open problem is learning from data that is heterogeneous, partially observed and shifting from site to site, while holding the answers to the reference standard.

How we are tackling it

• We began by fusing brain MRI with cognitive and clinical data, then built models that accept any combination of available inputs by design, trained across tens of thousands of people from many cohorts.
• We estimate multiple co-occurring causes of dementia and check them against autopsy-confirmed neuropathology.
• We predict amyloid and tau status, including where tau has spread, from routinely collected data, and validate it against PET and pathology staging.
• For early detection outside the clinic, we study voice recordings, smartphone cognitive tests and other digital signals.
• To make models work at new sites, we develop domain generalization, modality-agnostic, representation-learning and graph-based methods.

Code: Differential dementia diagnosis · Alzheimer's biomarker assessment · Multimodal AD dementia assessment · Interpretable AD classification · Disease-driven domain generalization · Dementia detection from voice · Speech features on GPU

AI for digital microscopy

The challenge

Many diseases are ultimately confirmed under a microscope. Pathologists read tissue slides, electron micrographs show ultrastructure at nanometre scale, and fluorescence and other stains map specific molecules. For the neurodegenerative diseases we study, brain tissue examined at autopsy remains the reference standard. Digitizing these images has made them available to AI, but they are unlike everyday pictures. A single image can hold billions of pixels yet carry one label for the whole specimen, the informative structures may be tiny and scattered, and stains, microscopes and scanners vary from lab to lab. Expert reading is slow and scarce, and how tissue appearance relates to molecular state and clinical outcome is still largely unmapped.

How we are tackling it

• We build models that learn from gigapixel whole-slide images using only specimen-level labels, including graph-based, transformer and multiple-instance approaches, among them a Fourier-filtering method that picks out informative regions.
• We link what tissue looks like to what it is doing molecularly, by fusing microscopy with gene expression to predict outcomes and stratify premalignant lesions.
• We work across stains, organs and resolutions, from fibrosis in trichrome-stained biopsies and glomerular segmentation to electron-microscopy ultrastructure of the kidney filtration barrier.
• We bring these tools to the point of care, for example with a web tool that checks whether a kidney biopsy is adequate while the patient is still in the procedure room.

Code: Graph-transformer for WSIs · Pathology + gene expression fusion · FourierMIL

From predictions to decisions

The challenge

Clinicians don't need another risk score at the end of a workup. They need help along the way: which test adds the most for its cost, who needs a specialist or a scan, and who will benefit from a new therapy or belongs in a trial. Every unnecessary test costs money and patient burden, and trials fail when they enroll the wrong people. To be actionable, AI has to reason about uncertainty and cost, and fit into a real clinical workflow.

How we are tackling it

• We use active learning to decide which measurement to obtain next, guided by what the model can explain.
• We study how AI can make neurologists more efficient without taking decisions out of their hands.
• We stratify people with cognitive impairment by their future risk, and design AI-augmented approaches to Alzheimer's clinical trials.

Code: Explainability-driven feature acquisition

Domain-specific language and vision-language models

The challenge

General-purpose language models pass medical exams but stumble on the specifics of neurology and dementia care. They rarely cite evidence, and they can't read a brain MRI alongside a clinical note. Clinicians and researchers need models that know the field, reason over images and text together, and back every answer with sources that can be checked.

How we are tackling it

• We build domain-specific language models for research and education.
• We develop agentic systems that search textbooks, the literature and clinical trial records and ground each answer in cited evidence. The underlying collections of texts are openly available.
• We are building vision-language models for neuroimaging that interpret scans in clinical context.

Code: PodGPT · Agentic medical QA