)
)
Déjà vu in healthcare AI: lessons from the world’s pioneer AI clinical decision support system
Recent advances in artificial intelligence (AI) have renewed interest in the possibility of computers assisting, or even replacing, doctors in making clinical decisions. However, computerised clinical decision support (CCDS) is not new, with scientific roots going back to the 1950s.1 One of the first applications, a system for diagnosing causes of abdominal pain known as AAPHelp, was developed at the University of Leeds under Tim de Dombal’s leadership.2 To use the system, clinicians would take a structured assessment of a patient on a paper form (figure 1). Data from the form was then entered into a computer programme—a Naïve Bayes classifier, a simple machine learning algorithm that uses Bayes’ rule to estimate conditional probabilities. The output was a differential diagnosis, in which each potential diagnosis had an estimated probability. In its initial installation, the programme ran on the 4.7 tonne KDF-9 computer and diagnoses could take up to 20 min. The aim of this editorial is to summarise the key findings from AAPHelp studies, contextualising them against the current AI zeitgeist and highlighting their continued relevance for today’s AI research.