Random Weights Neural Network for Low-Cost Readout of Colorimetric Reactions: Accurate Detection of Antioxidant Levels

The introduction of Point Of Care (POC) devices is revolutionizing the field of diagnostics, thanks to their ease of use, portability, and real-time results. However, despite such advantages, POCs are still less accurate than traditional laboratory-based methods. In most cases, this is due to the qualitative on-off response of POCs along with readout procedures involving methods that are easily influenced by the environmental conditions or by the acquisition step of the result. Automation of the readout using machine learning supported by frugal devices and lowcost sensing systems can significantly enhance the quality of the analysis performed by POC devices, while maintaining the aforementioned advantages. This paper proposes the use of random-based neural networks to accurately assess the salivary antioxidant level detected through a colorimetric reaction. As a test case, a low-cost IoT device equipped with a trained neural network that infers the antioxidant level in the user's saliva was designed and tested. The experiments performed on real-world data confirm that the proposed solution outperforms the previously proposed readout strategy.