Machine Learning-Based Rapid Diagnostic-Test Reader for Albuminuria Using Smartphone

Abstract

Albuminuria is an excellent marker for early diagnosis of kidney and cardiovascular disease. Urine dipsticks that are widely used for rapid screening of albumin, lack sensitivity and specificity in lower concentrations (<300 mg/dL) which is clinically very significant for early diagnosis and often provide qualitative or semi-quantitative results. Precise quantification of lower concentrations is based on urinary analyzers that are not portable and cannot be used in point-of-care (PoC) settings. Here, the feasibility of an accessory free analytical device has been demonstrated using a smartphone. Amalgamation of a smartphone with a dipstick enables rapid and inexpensive diagnosis. It estimates not only the standard five concentrations used in dipstick method, but ten different concentrations. This enables accurate detection and quantification of albumin at lower concentration, clinically significant in the early diagnosis of kidney disease. In order to mitigate ambient light conditions and shadow of the smartphone, images of strips were taken in a smartphone camera with “Flash ON” mode. Machine learning algorithms were used to classify ten different albumin concentrations, corresponding to normal, micro albuminuria and macro albuminuria conditions. The study was performed under varying illumination conditions using multiple smartphones. Random Forest algorithm yields an accuracy of 92% in constant illumination and variable smartphone conditions. In variable smartphone and illumination condition, it yields 82% accuracy on the test data. The detection limit of the proposed method is 7.8125 mg / dL.