Technology, big data, and the future of paediatric neuroscience: let us go then, you and AI.

Abstract

We have all observed how great pioneers of paediatric neurology effortlessly drew on their vast reserves of knowledge and experience to make diagnoses based primarily on observation. In today's environment, individual clinicians see fewer cases, while medical science uncovers increasing numbers of often rare conditions. Artificial intelligence (AI), machine learning, and big data promise to emulate, permanently represent, and make widely available the fruits of clinical experience. The concept of big data is not new: petabyte (1015 bytes)-sized resources have informed particle physics, climate science, and genomics for decades.1 One may ask how big data can benefit a field increasingly dominated by rare diseases. It is axiomatic that genomic data is uninterpretable without phenotypic information, and readers will be familiar with SimulConsult (https://simulconsult.com), a machine learning-based platform to assist the diagnosis of rare conditions in child neurology, sometimes with unexpected results.2 SimulConsult's models are based on large quantities of data, but represent only a subset of available information. AI can now rapidly identify, extract, understand, and integrate data from multiple sources. All components of the diagnostic assessment – demographics, history, examination, imaging, laboratory tests, genomic data – contribute to outputs tailored to decision support. Seizure prediction and brain tumour classification already benefit from this approach.3 There are crucial differences between AI and the encyclopaedic knowledge of our forebears. Wisdom can be communicated in books, papers, lectures, and tutorials, but these are slow, indirect methods of accessing knowledge. Moreover, clinical diagnosis is typically hypothesis driven, with much available information discarded as irrelevant. This ‘top down’ problem solving approach is notoriously susceptible to cognitive biases,4 while a fully ‘bottom up’ approach would test the relevance of all data points, discarding no information, allowing novel phenotypes to emerge, new associations to be recognized, and unknowns to come to light. A decision-making human has neither the time nor computational resources to accomplish such a feat. Data integration is counterintuitive: how is the signal in a magnetic resonance imaging voxel related to cerebrospinal fluid cell count? It may not be, but in a fully semanticized data structure – one that represents not only facts but also the links between them – a connection that is invisible to the human eye can become a critical insight. Data capture is another challenge: most information is represented not as a value in a field but in words. AI has led to improvements in natural language processing and automatic speech recognition that allow specific quantitative and qualitative information to be extracted rapidly, reliably, and at scale from written and spoken language.5 There will be barriers to the exploitation of AI, one of which is culture. Clinicians and data scientists are not natural collaborators; data and methods tend to be carefully guarded rather than shared between specialties and institutions. Yet open access tools and shared data from large-scale initiatives such as the European Union Human Brain Project (www.humanbrainproject.eu) suggest that culture change is underway. Data sharing necessitates strict governance; the temptation to monetize clinical data has led to serious breaches (https://www.newscientist.com/article/2088056-did-googles-nhs-patient-data-deal-need-ethical-approval/). The recently adopted General Data Protection Regulations enshrine a framework of confidentiality, consent, and data ownership within which big data can be usefully and safely shared.