Artificial intelligence-aided rapid and accurate identification of clinical fungal infections by single-cell Raman spectroscopy

Abstract

Integrating artificial intelligence and new diagnostic platforms into routine clinical microbiology laboratory procedures has grown increasingly intriguing, holding promises of reducing turnaround time and cost and maximizing efficiency. At least one billion people are suffering from fungal infections, leading to over 1.6 million mortality every year. Despite the increasing demand for fungal diagnosis, current approaches suffer from manual bias, long cultivation time (from days to months), and low sensitivity (only 50% produce positive fungal cultures). Delayed and inaccurate treatments consequently lead to higher hospital costs, mobility and mortality rates. Here, we developed single-cell Raman spectroscopy and artificial intelligence to achieve rapid identification of infectious fungi. The classification between fungi and bacteria infections was initially achieved with 100% sensitivity and specificity using single-cell Raman spectra (SCRS). Then, we constructed a Raman dataset from clinical fungal isolates obtained from 94 patients, consisting of 115,129 SCRS. By training a classification model with an optimized clinical feedback loop, just 5 cells per patient (acquisition time 2 s per cell) made the most accurate classification. This protocol has achieved 100% accuracies for fungal identification at the species level. This protocol was transformed to assessing clinical samples of urinary tract infection, obtaining the correct diagnosis from raw sample-to-result within 1 h.