Artificial Intelligence for Kidney Stone Spectra Analysis: Using Artificial Intelligence Algorithms for Quality Assurance in the Clinical Laboratory

Abstract

Objective: To determine if a set of artificial intelligence (AI) algorithms could be leveraged to interpret Fourier transform infrared spectroscopy (FTIR) spectra and detect potentially erroneous stone composition results reported in the laboratory information system by the clinical laboratory.

Background: Nephrolithiasis (kidney stones) is highly prevalent, causes significant pain, and costs billions of dollars annually to treat and prevent. Currently, FTIR is considered the reference method for clinical kidney stone constituent analysis. This process, however, involves human interpretation of spectra by a qualified technologist and is susceptible to human error.

Methods: This prospective validation study was conducted from October 29, 2020, to October 28, 2021, to test if the addition of AI algorithm overreads to FTIR spectra could improve the detection rate of technologist-misclassified FTIR spectra. The preceding year was used as a control period. Disagreement between the AI overread and technician interpretation was resolved by an independent human interpretation. The rate of verified human misclassifications that resulted in revised reported results was the primary end point.

Results: Spectra of 81,517 kidney stones were reviewed over the course of 1 year. The overall clinical concordance between the technologist and algorithm was 90.0% (73,388/81,517). The report revision rate during the AI implementation period was nearly 8 times higher than that during the control period (relative risk, 7.9; 95% CI, 4.1-15.2).

Conclusion: This study demonstrated that an AI quality assurance check of human spectra interpretation resulted in the identification of a significant increase in erroneously classified spectra by clinical laboratory technologists.

Figure 1

Workflow of Fourier Transformed Infrared (FTIR) spectra classification process

Figure 2

Example of the explainability of the model predictions using SHAP activation profiles. The first row illustrates 2 different reference spectra, pure magnesium ammonium phosphate (MAP, struvite) and ammonium urate (AU). The second row presents the SHAP activation profiles that contributed to the prediction with high confidence (>0.999 for both spectra). SHAP values were generated for the initial validation samples using 600 randomly selected training samples as the background. Fifteen IF features with the largest absolute SHAP values were selected and overlayed on the spectra signal identifying most important individual wavelengths to the prediction. The third row illustrates how the model predictions function on a spectra that combines the MAP and AU in a 60:40 ratio. The activation profiles for this combined element demonstrate how features from the pure spectra are combined into this mixed stone type.