Machine-Learning-Enhanced Fluorescent Nanosensor Based on Carbon Quantum Dots for Heavy Metal Detection

Fluorescent nanomaterials such as carbon quantum dots (CQDs) have been widely utilized as optical nanosensors for the detection and imaging of chemical or biological species with their superior sensitivity and spatiotemporal monitoring capabilities. Even though the nanosensors originally provide high sensitivity, the potential limit of detection (LOD) could not be accurately utilized for real-world applications due to the absence of signal extraction near the noise level, which is significantly influenced by various environmental factors and conventional statistical methods. To address this issue, we have developed an optical signal processing technique based on machine learning to enhance the practical sensing performance of CQD nanosensors for heavy metal ion detection. Fluorescence spectra of the nanosensor are acquired under both normal conditions (absence of analytes, n = 786) and analyte conditions (presence of analytes with varying concentrations, n = 288). Seven different types of machine learning algorithms are systematically applied to a training data set of fluorescence features (n = 200). Our best model could distinguish the spectra of 10 nM and 100 pM Cu2+ from just noise signals without analytes, which are 104 and 106 times enhanced LOD compared to the original detection limit (95 μM) providing accuracies of 90.69 and 70.41%, respectively. We further demonstrate that this technique could be universally applied for different types of analytes such as Hg2+, Pb2+, Cr6+, and Fe3+ ions with distinct sensing performances. This technique can be up to 106 times lower than the LOD of the fluorescent nanosensors. Our approach is not limited to CQD nanosensors or metal ions but can also be directly applied to various types of fluorescent nanosensors and biochemical analytes.