Ten percent of admissions to Australian acute-care hospitals are associated with harm to patients or adverse events. The reporting of critical incidents by health professionals is now well established and the rate of reporting continues to increase worldwide. Current methods, which rely on retrospective manual review of incident reports, do not permit timely detection of safety problems and can no longer keep up with this growing volume of data. In New South Wales alone, more than 137,000 patient-safety incidents were reported in 2011.  

We are evaluating text classification methods to capture incident reports automatically by type and risk rating. The goal is to track ten types of patient-safety problems nationally working in collaboration with St Vincent’s Hospital, Sydney and the NSW Clinical Excellence Commission. Working with the Australian Commission on Safety and Quality in Health Care we have shown that text classifiers based on well-evaluated machine-learning techniques such as Naïve Bayes and Support Vector Machines can be effective in automatically identifying incidents in two priority areas – clinical handover and patient identification. More recently we have shown the feasibility of using machine learning to identify IT incidents.