Artificial intelligence and genomics in paediatric neurology: Promise, pitfalls, and the importance of clinical judgement

Abstract

The exponential rise of computing power, coupled with increased availability of data, has propelled machine learning into rapid development – the artificial intelligence (AI) revolution. AI, particularly machine learning, promises to transform healthcare, and paediatric neurology is no exception. Our specialty relies on data-heavy investigations such as high-resolution neuroimaging and multi-channel electroencephalography. In this setting, AI and machine learning can offer three distinct contributions.1 ‘As good as the best’. If AI can perform at least as well as a specialist paediatric radiologist or neurophysiologist, it could save a significant amount of clinician time. This may translate into cost savings for care systems and address inequalities in access, because highly specialized skills and tests are often confined to major centres. ‘Better than the best’. If AI can outperform a human specialist, patients may gain earlier detection of disease and/or better treatment. New paradigms. Machine learning can detect patterns that humans have not considered or cannot see, potentially defining completely new diagnoses. Whilst these prospects sound like progress, we must exercise caution. More sensitive diagnostics can deliver problems as well as solutions, a theme explored by Suzanne O'Sullivan in The Age of Diagnosis.2 Lower or blurred diagnostic thresholds can profoundly affect individuals – putting them at risk of developing ‘self-fulfilling’ psychosomatic or functional symptoms that may not otherwise have troubled them. Young people with such symptoms, particularly if at the severe end, often present to paediatric neurology. In an AI-enabled future, health professionals will be required more for counselling skills and psychological support than for diagnostic skills. However, it is unlikely that AI will solve medical workforce shortages. Similarly, genomics is rapidly lifting our diagnostic ceiling in paediatric neurology – the centre of gravity is shifting from tests that label disease to therapies that modify it. Targeted small molecules, antisense oligonucleotides, and gene therapies are beginning to change trajectories in conditions we once considered untreatable. The practical message is simple. Instead of ending with a static report, each additional molecular diagnosis increasingly carries the chance of a licensed therapy, a targeted trial, or precision surveillance via natural-history studies (https://clinicaltrials.gov/study/NCT06504511). Equity demands that we enable access through national networks and shared standards. That promise does not diminish the skill of phenotyping; it raises the bar. Advanced language models can mine clinic letters, electroencephalogram reports, and imaging summaries at scale, but they cannot rescue poor inputs: garbage in – garbage out. We need thorough standardized examination, agreed descriptors, structured ontologies such as the Human Phenotype Ontology, and consistent data capture so automated pipelines reflect the child in front of us rather than the biases of our records. Newborn screening and early genomic diagnosis are paving the way for timely intervention, yet they reveal ethical and clinical dilemmas.3 Missense variants of uncertain significance are common in asymptomatic infants, and most will be harmless. Interpreting and communicating these nuances matter.4 A benign but poorly communicated finding can fuel months of parental concern, extra appointments, and avoidable anxiety – proof that sharing genetic findings without careful framing can cause harm. These themes sit at the heart of the 2026 British Paediatric Neurology Association (BPNA) conference in Glasgow. Sessions on advances and controversies in technology and AI will run alongside the enduring genotype–phenotype debate. Can an algorithm make a genetic diagnosis without a clinician? Perhaps it can flag the right variant faster, but only clinicians synthesize family history, phenotype, context, and values; only clinicians carry the duty to counsel, consent, and care.5 Our task is to build workflows where genomics and AI inform rather than replace clinical judgement, delivering earlier, kinder, and more effective care for the children and young people we serve. Not required.

Ähnliche Arbeiten