Scientifically Validated AI-based Solution for Self-Triage

Oscar Garcia-Esquirol, Elvira Moreno Barriga1 , Irene Pueyo Ferrer2 , Miquel Sánchez Sánchez2 , Montserrat Martı´n Baranera3 , Josep Masip Utset1 Emergency Department, Hospital Clı´nic de Barcelona, Bareclona, Spain 1 Intensive Medicine Unit, Consorci Sanitari Integral; 2 Emergency Department, Hospital Clínic de Barcelona; and 3 Epidemiology and Research Department, Consorci Sanitari Integral Background: Prospective observational study conducted in the Emergency Department of Hospital Clínic de Barcelona (HCB).

This is a tertiary level (high complexity) 800 bed hospital, located in the city center, and reference hospital for a population catchment of 550,000 people approximately. At present, approximately 95,000 general emergencies are attended annually, excepting the specialties of obstetrics and gynecology, pediatrics and ophthalmology, which are performed in other centers. Since February 2009 the triage of patients in the emergency department is performed by nursing staff, using the allocation of level of urgency according to the Andorran Triage Model (MAT). The research was performed in a 4-month period, from October 1, 2016 until January 31, 2017. The selected patients were those assigned MAT triage levels III, IV and V (which did not require immediate attention) in order of arrival and in different shifts to ensure obtaining a representative sample.

Methods: Observational, prospective study performed in the Emergency area of a tertiary university hospital. Patients with medical and surgical (including orthopedic pathologies) which did not require immediate medical assistance were given a tablet to answer Mediktor’s questioning. According to their answers, each patient was attributed a list of 10 pre-diagnoses sorted by probability, which were concealed both from the patient as well as the attending physician, in order to avoid modifying the regular process. Subsequently, the degree of agreement between the medical diagnosis and the diagnoses offered by Mediktor was analyzed.

Results: 1,015 patients were surveyed, and 622 patients from this group were considered valid cases for analysis. Patients who did not meet the inclusion criteria, without a diagnosis on discharge, those who had their final diagnosis stated as a symptom and those with diagnoses not included in Mediktor were excluded from the study. Matches between the medical diagnosis (gold standard) and the 10 diagnoses of Mediktor was 91.3%, in the first three diagnoses was 75.4% and on the first diagnosis it was 42.9%. According to the most common diagnostic groups, sensitivity >92% and specificity >91% in most of them with a kappa index between 0.24 and 0.98 were observed. Mediktor is a reliable tool to help in the diagnosis of the most prevalent diseases found in an Emergency Service and it is easy to use by the general public.

Discussion: We carried out this study on Mediktor since, besides being a symptom evaluator using AI algorithms, it allowed us to perform the medical questionnaire without problems of understanding by users using natural language. After comparing the obtained diagnoses (physician & evaluator) it was observed that Mediktor had a high diagnostic match considering the list of 10 possible pathologies (ordered by probability) obtaining 91.3% with the diagnosis given by the Dr. In turn we were able to verify the high sensitivity (>92%) and specificity (>91%) of the evaluator. The patients were confident with the system and had no difficulties in answering the different questions that the symptom checker asked. Thus, more than 98% of respondents felt that Mediktor was an easy-to-use program. We consider that the results of our study provide relevant data that lead us to consider that Mediktor is an IT-aided diagnostic tool that can accelerate the diagnostic process with improved efficiency

Learn more about our “Scientifically Validated AI-based Solution for Self-Triage” study published in a leading American scientific journal Mary Ann Liebert, Inc.