Event Overview: The Future of Multi-modal Healthcare AI Workshop

The AdSoLve multimodal medical diagnostics & AI workshop on November 13th 2025 featured presentations from the team, invited talks from international experts in cancer research, paediatric urology and AI for mental health, a panel and break-out discussions.

AdSolve Team Presentations

The workshop opened with Prof. Maria Liakata (QMUL), who introduced AdSoLve’s mission to address the sociotechnical limitations of AI systems in healthcare and other high-stakes domains. She highlighted current progress on multimodal evaluation frameworks, temporal reasoning modules, and responsible dataset development, emphasising the need for trustworthy AI that aligns with clinical goals and integrates safely into NHS workflows. Presentation slides HERE.

Prof. Claude Chelala (Barts Cancer Institute) presented a decade of work building advanced multimodal cancer biobanks, including the Pancreatic Cancer Research Fund Tissue Bank and the Breast Cancer Now Biobank. These resources combine longitudinal clinical records, radiology, histopathology, and multi-omics data to support precision diagnostics and patient-journey reconstruction. Chelala illustrated how aligned and high-quality multimodal datasets unlock new possibilities for early signal detection, disease progression modelling, and more equitable clinical research. Presentation slides HERE.

Prof. Greg Slabaugh (QMUL) discussed key methodological and practical challenges in multimodal AI for healthcare, from heterogeneous sampling frequencies to missing modalities and generalisation failures in real-world deployments. Through examples in neuro-oncology, he showed how fusing imaging, structured data, and text can improve prediction quality and interpretability—but also requires robust evaluation, transparency, and closer collaboration between clinicians and data scientists. His message echoed a theme of the day: multimodality is powerful, but must be bounded by clinical validity. Presentation slides HERE.

Iqra Ali (QMUL) followed with a systematic overview of multimodal diagnostic tasks, comparing unimodal multi-visit, and longitudinal multimodal settings. She outlined multimodal fusion strategies, dataset challenges, and the ethical and governance demands of multimodal systems—including privacy constraints, safe model alignment, and bias amplification through uneven modality coverage. She identified emerging research needs such as uncertainty-aware modelling, temporal alignment, and scalable evaluation frameworks that reflect clinical relevance rather than purely technical benchmarks. Presentation slides HERE.

The AdSoLve evaluation platform—Sebastian Löbbers, Cian Kennedy, and Mahmud Akhter (QMUL)—then demonstrated the project’s model-agnostic evaluation platform. Using radiology report generation as a case study, they showcased “evaluation bundles” that combine semantic metrics, clinical error detection, and human–AI comparison tools. The platform provides instance-level interpretability, flexible metric configuration, and could be deployed in secure evaluation environments designed for the NHS, setting performance expectations and informing procurement. Audience discussion highlighted demand for evaluation approaches that reflect real-world workflows, human oversight requirements, and organisational constraints. Presentation slides HERE.

Invited Talks

The clinical keynote was delivered by Prof. Hemant Kocher (Professor of Liver and Pancreas Surgery at Barts Cancer Institute, London) , who examined the acute challenges in pancreatic cancer diagnosis. He described how early-stage pancreatic cancer often presents with occult metastases invisible to current imaging modalities, making surgical planning exceptionally difficult. Kocher outlined the critical role of high-quality biobanks for early detection research, noting that multimodal datasets enable retrospective identification of pre-diagnostic signals in primary care and more accurate risk stratification. The talk underscored the urgency of developing AI tools that are truly clinically actionable, not merely accurate on retrospective datasets. Presentation slides HERE.

Dr. Gregory Dean (Chief of Paediatric Urology, St Christophers Hospital for Children, Clinical Professor of Urology, Temple University and Founder of DriQ Health, US) discussed multimodal AI in paediatric urology, a field characterised by scarce specialist expertise and substantial diagnostic uncertainty. He described multiple imaging-based challenges—hydronephrosis grading, obstruction detection, urodynamic interpretation—where variability among clinicians limits diagnostic reliability. Dean emphasised the opportunity for LLM based systems to support both general practitioners (by providing specialist-level guidance) and specialists (by distilling complex comorbidities across modalities). At the same time, he warned of over-reliance on “black-box” systems and advocated for interpretable reasoning pathways and rigorous safety expectations comparable to high-reliability industries such as aviation. Presentation slides HERE.

The final presentation, Prof. Philip Resnik (Institute for Advanced Computer Studies and Department of Linguistics, University of Maryland, US), provided a deep dive into multimodal modelling for mental-health assessment and early intervention. Drawing on newly collected and clinically validated datasets—including text, speech, facial video, and structured psychometric measures—he demonstrated how symptom-level modelling for schizophrenia and depression can outperform coarse diagnostic categories, offering clinicians more actionable insights. Resnik also introduced population-relative risk-ranking for suicidality, using pairwise LLM judgements guided by clinical criteria. His results showed promising ability to prioritise individuals at imminent risk. He highlighted three systemic challenges seen across the workshop: the scarcity of shareable clinically validated data; the need to foreground fine-grained, symptom-level insights; and the importance of shifting from diagnosis to triage given chronic workforce shortages. Presentation slides HERE.

Panel: Opportunities and Challenges in Multimodal Medical AI

The panel brought together Kathrin Cresswell (Professor of Digital Innovations in Health and Care at the University of Edinburgh), Steven Newhouse (Deputy Chief Information Officer at Barts Health NHS Trust), Chetan Kaher (Chief Innovation Officer, JIVA.AI), Prabhu Arumugam (Clinical Innovation Lead, AWS), and Dan Schofield (AI Technical Specialist, NHS England).

Panelists discussed trust, integration, and real-world readiness as primary barriers. They stressed that multimodal tools will only be adopted if they demonstrably improve workflows rather than add cognitive or administrative burden. Infrastructure gaps—such as inconsistent digital pathology systems, siloed genomic and imaging pipelines, and limited access to standardised data across Trusts—were noted as major bottlenecks.

Automation bias and the necessity of maintaining clinician autonomy were recurring concerns. Participants emphasised the need for interpretable outputs, reliable uncertainty estimates, and design approaches that avoid de-skilling clinicians while supporting decision-making under pressure. Several drew parallels to aviation: rare errors in high-stakes environments demand rigorous safety engineering and continuous monitoring.

The panel also touched on privacy, risk, equity, and governance. Experts warned of a potential “two-tier health system” if multimodal diagnostic tools are only validated on narrow clinical populations or deployed unevenly across regions. They called for

nuanced public dialogue that balances risks and benefits, moving beyond unrealistic expectations of perfect privacy or perfect accuracy.

Training and workforce development were highlighted: clinicians need practical, workflow-integrated education rather than theoretical AI instruction. Panelists also discussed the importance of transparency about limitations, model update policies, and post-deployment monitoring to ensure models remain safe over time.

Breakout Sessions

Participants engaged in facilitated discussions exploring the potential, challenges, and requirements for multimodal AI in diagnostics. Across groups, several themes emerged:

• Highest-value use cases included multidisciplinary team decision support, emergency medicine triage, oncology, cardiology, ophthalmology, and chronic disease prevention—domains where multiple modalities already shape clinical judgement.
• Data challenges were repeatedly raised: missing modalities, uneven sampling, inconsistent equipment, limited consented datasets, and organisational barriers to retrieving primary-care longitudinal data. Many noted tensions between multimodal ambitions and data-minimisation norms under existing privacy frameworks.
• Economic and organisational constraints were identified as core barriers to adoption. Beyond technical accuracy, Trusts require a compelling business case that justifies integration costs, workflow disruption, and ongoing validation resources.
• Regulatory complexity was widely acknowledged. Participants highlighted the MHRA’s evolving guidance, the burden of re-certifying software after each model update, and the mismatch between traditional medical-device regulation and rapidly evolving AI systems.
• Stakeholder involvement was seen as essential: health economists, patient safety teams, procurement officers, regulators, clinical champions, and IT infrastructure specialists must all be engaged early to ensure practical, safe, and sustainable deployment.
• Success metrics were considered broader than accuracy alone—covering time saved, reduction of missed findings, improved patient experience, lower clinician burden, and avoidance of unintended downstream strain elsewhere in the system.

Closing Reflections

Across the workshop, a clear message emerged: multimodal AI holds exceptional promise for improving diagnostic accuracy, early detection, and personalised care—but only if developed and evaluated within the realities of clinical practice. Success will depend not only on technical innovation but also on governance, data stewardship, regulatory evolution, and deep collaboration with clinicians, patients, and NHS organisations.

AdSoLve’s work—spanning robust evaluation frameworks, multimodal data alignment, temporal reasoning, and responsible deployment—aims to provide these foundations and accelerate the safe and effective introduction of multimodal medical AI across the NHS.