Automated diagnoses from clinical narratives: A medical system based on computerized medical records, natural language processing, and neural network technology

Abstract

A collection of artificial, associative neural networks (PROMNET) interfaced to a computerized medical record is described. Clinical narratives were subject to automated natural language processing, and relations were established between 14,323 diagnoses and 31,381 patient findings. Patient diagnoses and findings were counted, grouped into clinical entities, and used to train PROMNET. Training was completed in a few minutes. PROMNET's dictionary contained about 20,000 words, and the neural network recognized more than 2800 disorders. Its performance was evaluated by an automated, double-blind Turing test. PROMNET made clinical decisions in a few seconds with sensitivity of 96.6%, and specificity of 95.7%. The most pertinent clinical entity was usually ranked highest. PROMNET is a powerful inference engine that learns from clinical narratives and interacts with medical personnel or patients in natural language.