Diagnosis prediction based on similarity of patients physiological parameters

Medical staff can be considerably supported in patient healthcare delivery thanks to the adoption of machine learning and deep learning methods by enhancing clinicians decisions and analysis with targeted clinical knowledge, patient information, and other health data. This paper proposes a learning methodology that, on the basis of the current patient health status, clinical history, diagnostic and laboratory results, provides insights for clinicians in the diagnosis and therapy decision processes. The approach relies on the concept that patients with similar vital signs patterns are, in all probability, affected by the same or very similar health problems. Thus, they can have the same or very similar diagnoses. Patients physiological signals are modeled as time series and the similarity among them is exploited. The method is formulated as a classification problem in which an ad-hoc multi-label k-nearest neighbor approach is combined with similarity concepts based on word embedding. Experimental results on real-world clinical data have shown that the proposed approach allows detecting diagnoses with a precision up to about 75%.