Decision support for diagnosis and trige in primary care

Purpose and goal

The project, which is based on Doctrin´s automated patient interviewing tool, has together with KTH, Lund University and Capio AB, developed a machine learning prototype for triage of patients with the 10 most common medical complaints in primary care. In order to carry out this work, a process for machine anonymization of medical data has been developed, as well as a user interface for annotation and validation of medical reports. The final results will be reported in scientific publications in 2019-2020, and distributed to other stakeholders in popular science.

Expected results and effects

We show that patient-reported structured medical data (questionnaire responses) have limited predictive value for triage, both when interpreted by human doctors and machine learning. Human interpretation of the patient´s own description of ideas, expectations and concerns is what makes a difference. Preliminary data show that AI free text analysis of the patient´s own description + chats improve the prediction of the triage level. The comparison between physicians and the machine learning algorithm was limited by the interrater variability between the physicians.

Planned approach and implementation

The project was based on 14220 machine anonymised medical reports. From these, 300 reports (30 each of the most common 10 search causes) were randomly selected for annotation (triage category and up to 3 differential diagnoses) by a specialist in general medicine. The machine learning algorithm was trained on these annotations. 5 primary care physicians doctors each diagnosed and triaged a different set of 300 reports for the same search causes. Their assessments were then compared with the results of the machine learning algorithm.