Can AI predict research impacts?

The success or failure of medical research is judged by patient outcomes far downstream of the strategic decisions that initiate it. Optimising translational impact therefore relies on long range forecasting, for which no established framework exists. The evaluation of research proposals by expert appraisal of their content is undermined by difficulties with scaling, reproducibility, generalisability, and bias. Evaluation by summary bibliometrics of public reception offers greater objectivity but doubtful fidelity. Both approaches favour the familiar, the conventional, the plausible, and the incremental; and oppose the unusual, the unorthodox, the counter-intuitive, and the disruptive: rare characteristics on which translational success increasingly depends. 

In this talk, Amy Nelson and Parashkev Nachev (UCL) advocate for a third way, founded on richly expressive models of research content, that seeks to combine the finesse of a human expert with the rigour of a machine. They argue such models can successfully capture regularities too intricate to be either intuitively apprehensible or reducible to summary metrics, thereby illuminating complex characteristics of translational success in which testable hypotheses about optimal research strategy may be grounded. 

They describe a proof-of-concept analysis of the comparative predictability of future real-world translation—as indexed by inclusion in patents, guidelines, or policy documents—from complex models of title/abstract-level published research content versus citations and metadata alone. Quantifying predictive performance out-of-sample, ahead of time, across major domains, using the entire corpus of biomedical research captured by Microsoft Academic Graph from 1990–2019, encompassing 43.3 million papers, they show that high-dimensional models of titles, abstracts, and metadata exhibit substantially higher fidelity (AUC > 0.9) than simple models, generalise across time and domain, and transfer to recognising the papers of Nobel laureates. Their talk will build on this recent paper in Patterns.

The Speakers

Amy Nelson is a Senior Research Associate in the High Dimensional Neurology Group at UCL Queen Square Institute of Neurology, Research Impact Fellow at the NIHR UCLH Biomedical Research Centre, and a junior doctor. Dr Nelson builds AI models for clinical, operational and research impact objectives across computer vision, deep representation learning, and natural language processing domains.

Parashkev Nachev is a Professor of Neurology at the UCL Institute of Neurology, and Honorary Consultant Neurologist at the National Hospital for Neurology and Neurosurgery, Queen Square. His High-Dimensional Neurology Group develops novel computational methods for drawing representational, predictive, and prescriptive intelligence from rich data.