Combining orthogonal analytical techniques to identify substandard or falsified formulations of pharmaceuticals | Poster Board #610

In this work, we explored deep learning approaches to assess the salient features of two very different field screening technologies for more accurate identification of substandard and falsified pharmaceuticals (SFPs). The two technologies used were paper analytical devices (PADs) and near-infrared spectroscopy (NIR). PADs are microfluidic devices that perform multiple chemical color tests on the constituents of a pharmaceutical dosage form. The PAD is engineered to store chemical reagents and to test powder from a capsule or tablet. When water flows through the microfluidic channels, the reagents mix with the sample to generate unique color patterns. These color patterns were used to train a convolutional neural net and other machine learning tools to discriminate substandard or falsified formulations of pharmaceuticals from genuine formulations. NIR is a vibrational spectroscopy tool that responds to both molecular structure and solid-state arrangement of molecules. 3D-printed sample holders for capsules and tablets were designed to allow non-destructive analysis of pharmaceutical dosage forms. Machine learning tools learned the spectra features and a predictive algorithm was developed. The data from PADs and NIR probe different features of pharmaceutical dosage forms, increasing the predictive value of the screening tests when they are used together. In this study, we investigated how best to combine and weight these complimentary forms of data and compared several methods for expressing the confidence limits of the results.