Towards a Global Model for Diabetic Kidney Disease Screening using ATR-FTIR

Infrared (IR) spectroscopy of urine extracts coupled with Machine Learning (ML) methods has been proposed as a promising technique for Diabetic Kidney Disease (DKD) screening. However, for clinical translation, predictive models based on machine learning techniques necessitate substantial sets of samples for calibration and testing under various experimental parameters and populations, posing a critical barrier to the development of globally deployable tools. Here, we aim to assess the methodology's ability to establish DKD diagnostic models applicable across diverse populations, instruments and experimental conditions worldwide. Two datasets were compared. The Australian set included 155 DKD and 22 control samples from a 24-hour urine collection, with preconcentrated proteins measured on a Bruker spectrometer. The Spanish set, comprising 35 DKD and 26 control spot urine samples, was analyzed using a Perkin-Elmer spectrometer. Different ML methods were developed to identify DKD and microalbuminuria, aiming to compare their performance in terms of generalization and adaptation to different datasets. Models developed using Australian spectra successfully predicted Spanish samples, achieving AUROC values of 0.87 and 0.98 for DKD and microalbuminuria identification, respectively. Both values improved to 0.99 when a global model was calibrated and independently tested with a combined set integrating samples from both countries. Results evidence that the spectral markers found in the IR spectra, based on signals arising from albumin and other glycoproteins, have proven to be robust, minimizing the effects of population and instrument variability. Results exemplify the potential of developing global big-data spectroscopic datasets to facilitate the deployment of IR-based diagnostic methods in real-world settings.