Enhancing\nConfidence\nin Microplastic Spectral Identification\nvia Conformal Prediction

Microplastics are\nan emerging pollutant of concern, with\nenvironmental\nobservations recorded across the world. Identifying the type of microplastic\nis challenging due to spectral similarities among the most common\npolymers, necessitating methods that can confidently distinguish plastic\nidentities. In practice, a researcher chooses the reference vibrational\nspectrum that is most like the unknown spectrum, where the likeness\nbetween the two spectra is expressed numerically as the hit quality\nindex (HQI). Despite the widespread use of HQI thresholds in the literature,\nacceptance of a spectral label often lacks any associated confidence.\nTo address this gap, we apply a machine-learning framework called\nconformal prediction to output a set of possible labels that contain\nthe true identity of the unknown spectrum with a user-defined probability\n(e.g., 90%). Microplastic reference libraries of environmentally aged\nand pristine polymeric materials, as well as unknown environmental\nplastic spectra, were employed to illustrate the benefits of this\napproach when used with two similarity metrics to compute HQI. We\npresent an adaptable workflow using our open-access code to ensure\nspectral matching confidence for the microplastic community, reducing\nmanual inspection of spectral matches and enhancing the robustness\nof quantification in the field.