Diagnosis Detection Support based on Time Series Similarity of Patients Physiological Parameters

Facilitate clinicians in their complex decision- making processes allows to improve patient outcomes along disease-specific pathways and healthcare delivery. This could be realized by enhancing medical decisions with targeted clinical knowledge, patient information, and other health data. Nowadays, a key paradigm in healthcare, designed to be a direct aid to clinical-decision making, is the Clinical Decision Support System (CDSS). Along this line, the work presented in the paper focuses on detecting patients diagnosis, proposing a learning methodology that, on the basis of the current patient status, clinical history, diagnostic and results from their pathological reports, provides insights for clinicians in the diagnosis and therapy processes. The patients physiological signals have been modeled as time series and the similarity among them has been exploited. The main idea is that patients with similar patterns of vital signs are affected by the same or similar health problems and, therefore, may have the same or very close diagnoses. The diagnosis detection method is formulated as a classification problem, combining time series similarity and an ad-hoc multi- label k-nearest neighbor approach (ML-KNN). The proposed classifier exploits the semantic similarity of diagnoses catched through sentence embedding. Results, performed over a real- world clinical dataset, show that the proposed approach is able to successfully detect diagnoses with a precision up to about 75%.