BrainTech Alliance, Paris Brain Institute, September 15, 2026

Speakers: Sylvie Bothorel (Chief Data Officer, Paris Brain Institute), Ira Hebold Haraldsen (AI-Mind Project Coordinator, BrainSymph Chair), Thomas Nedelec (Research Scientist, Paris Brain Institute) Facilitator: Géraldine Farjot (Director Innovation, Paris Brain Institute)

The central question

Neuroscience has never had access to so much data: imaging, EEG, genetics, blood biomarkers, cognitive scores, sleep, cardiovascular and autonomic signals, and electronic health records. Yet this abundance of data still does not allow researchers to predict individual patient trajectories with sufficient accuracy.

Mild cognitive impairment (MCI) illustrates the challenge. Roughly 40% of patients with MCI progress to dementia, while 60% remain stable. The challenge is therefore to identify, early and cost-effectively, which patients are actually likely to develop dementia.

Biomarkers such as p-tau 217 can improve risk detection years before symptoms appear, but they are not enough to predict an individual patient’s trajectory. A positive 80-year-old may never develop dementia, while a positive 60-year-old may experience rapid cognitive decline.

For the panel, the main bottleneck is therefore not the power of AI models, but our ability to generate sufficiently rich, reliable, standardized and longitudinal data. As one speaker put it, AI will magnify both the strengths and the weaknesses of our data.

From single-purpose collection to reusable data

Most health data were collected for care, billing or operations, not research. They carry biases that vary by country, institution, software and clinical workflow. Using them well requires understanding how they were generated. That is why speakers argued that curation must stay close to the clinicians producing the data, even as others suggested researchers should be freed from infrastructure work to focus on research. The paradigm shift called for is to move away from collecting the minimum data for one declared purpose. The alternative is to collect rich data that can be reused across many questions. This enables longitudinal, multi-task, whole-patient views and becomes more cost-efficient with each reuse. It also sits in tension with regulatory logic that still favors narrow purposes and limited retention.

Retrospective data remain valuable for generating hypotheses. But the irreplaceable target is prospective, standardized, longitudinal follow-up of the same individuals. Speakers also described disease as a process rather than a snapshot. Dementia in particular was framed as a condition of the whole body, with immunological and metabolic signals essential to accurate risk assessment.

Standardization is the most tractable problem, and the most neglected

Standardization drew the strongest consensus. Format disputes, such as DICOM versus BIDS, were described as a waste of scarce intellectual bandwidth. Panelists also criticized the use of proprietary standards that lock customers into specific MRI or EEG hardware.

The call was therefore for internationally agreed, interoperable standards, backed by genuine industry commitment. But standardization should go beyond data formats: it should also cover how data is collected, documented, and used to develop and validate AI tools. European efforts to harmonize these requirements, including by bringing AI regulation closer to established clinical research standards, were seen as an important step forward, although implementation remains too slow.

Infrastructure exists, but exchange does not

Europe has built federated infrastructure, and EBRAINS was discussed directly. Speakers stressed the distinction between having infrastructure and actually exchanging data: the platforms are in place, but data rarely flows between them in practice. Two concrete blockers were cited: Security certification. AI factories, for example in Germany, cannot yet receive high-security patient data because they lack the required certification. Retention rules. French hospital data warehouses keep data for 10 to 20 years, while longitudinal research needs far longer horizons.

How much data is enough?

The panel genuinely disagreed on scale. One view held that research should start with deeply characterized, well-defined groups rather than chasing millions of records. The counterpoint was that scale is indispensable for rare diseases and subgroup discovery. UK Biobank, with about 500,000 participants, is an excellent resource, yet it contains only around 300 ALS patients, too few for stratification. The proposed resolution was disease-specific, deeply phenotyped initiatives modeled on PPMI (Parkinson’s Precision Medicine Initiative), extended to ALS, MS and other disorders. These cohorts would start small and be enriched progressively. Crucially, they would be designed to be comparable across diseases, so researchers can separate shared mechanisms from disease-specific ones.

What routine care data can already reveal

Paris Brain Institute researchers combine two sources:

  1. France's national health administrative data (SNDS), available since 2009
  2. The AP-HP clinical data warehouse, covering Paris public hospitals

Linking the two connects phenotype, care trajectory and pre-diagnosis history. Current work focuses on automatically extracting disease form, diagnosis timing and clinical scores from routine clinical notes, with MS and ALS as priorities. One finding illustrates the potential: type 2 diabetes is associated with increased Parkinson's risk and lower ALS risk. The next step is to enrich these trajectories with genetic and biomarker data from resources such as UK Biobank and PPMI.

Speakers also highlighted that relevant signals for assessing dementia risk often emerge outside neurology. Cardiovascular and brain health are closely connected, and early signs of cognitive decline may be detected during cardiology follow-up without triggering a neurological assessment. Similarly, post-operative delirium in older patients, including after hip replacement, can indicate increased cognitive vulnerability but is rarely followed up systematically. The broader challenge is therefore to connect signals that already exist across different specialties and care pathways, rather than treating neurological data in isolation.

The Paris Brain Institute is addressing this fragmentation by building a research data warehouse around a “virtual cycle of sharing”. Raw data enter the system, while enriched outputs such as segmented images, synthetic data and intermediate results are made available again to users. In parallel, teams in Sweden, Denmark and Germany are conducting comparable analyses of national health data. Because each country’s datasets reflect different healthcare systems and inherent biases, comparing results across countries is essential to distinguish robust signals from context-specific effects.

Economics: someone has to pay

Research data warehouses are largely free to users and borne by public systems, which is unsustainable. Maintenance is costly, and GPU infrastructure for model training adds substantially to the bill. Panelists saw public-private collaboration as the only viable path. They also argued that large IT companies can build federated infrastructure far faster than public actors alone.

AI-Mind offers a working model: Roche and Bosch co-finance analytics staff and use the project's databank to validate multimodal biomarkers and algorithms. Private investment, however, requires protected environments where companies can test hypotheses without competitors seeing their work. Fair cost-sharing between data collectors and data users is still unresolved, as is giving researchers time to publish before their data are opened. Speakers also argued for lifecycle-based economic modeling of prevention, in place of today's per-episode health economics, to make the policy case for investment.

Lessons from AI-Mind for large research projects

Ira Hebold Haraldsen shared lessons from AI-Mind, a Horizon 2020 project spanning eight European countries. The project's goal from day one was a product, not just scientific output. It brought in a notified body from the shipping industry early to enforce industrial-grade quality standards. Her recommendations for new EU-funded projects receiving €10–50M: Split databanks into development and validation sets from the start. Build algorithms designed for regulatory approval before the project ends. Control data quality continuously rather than retrospectively. These steps address the failure mode she described: well-funded cohorts built in isolation that never reach patients.

Participation and collection beyond the hospital

Speakers cited a stark gap in research participation between Nordic countries and France. They proposed three remedies: Software that reduces clinician burden Proactive invitations to patients Population-based screening, such as inviting an entire age cohort

They also argued data collection should move outside hospitals, through apps, connected devices and mobile units, including in low-resource settings. The European health data catalog, expected by 2029, will extend data-sharing obligations to device and wellness manufacturers. Its governance and financial terms are still being defined.

Europe's position

Panelists saw Europe's longitudinal depth and data quality, from national records to the long-running Nordic registries, as a genuine competitive advantage, at risk of being outpaced. They called for more self-confidence and for deliberate complementarity between European and US actors rather than duplicated effort. They also urged the field to borrow simulation and modeling expertise from other industries instead of reinventing it. Access for startups remains an open problem: small companies struggle to reach validated datasets that large pharma can obtain.

Open questions

  • How can federated data exchange be made operational when the infrastructure already exists?
  • What security certification would allow AI factories to process high-security patient data?
  • How should the costs of data collection and warehouse maintenance be shared between public funders, hospitals and industry?
  • What governance lets researchers publish first while making data available to others promptly?
  • How large must prospective cohorts be for rare disease stratification, and who funds them?
  • How can research participation in countries like France approach Nordic levels?
  • What business model lets a research data warehouse serve public researchers, startups and large industries at once?