Classification of the urinary metabolome using machine learning and potential applications to diagnosing interstitial cystitis

Feng Tong; Muhammad Shahid; Peng Jin; Sungyong Jung; Won Hwa Kim; Jayoung Kim

doi:10.14440/bladder.2020.815

Classification of the urinary metabolome using machine learning and potential applications to diagnosing interstitial cystitis

Bladder (San Franc). 2020 Jun 2;7(2):e43. doi: 10.14440/bladder.2020.815. eCollection 2020.

Authors

Feng Tong¹, Muhammad Shahid², Peng Jin^{2

3}, Sungyong Jung⁴, Won Hwa Kim¹, Jayoung Kim^{2

5

6

7}

Affiliations

¹ Department of Computer Science, University of Texas at Arlington, Arlington TX 76019, USA.
² Departments of Surgery and Biomedical Sciences, Cedars-Sinai Medical Center, Los Angeles, CA 90048, USA.
³ Shengjing Hospital of China Medical University, Shenyang 110004, Liaoning, China.
⁴ Department of Electrical Engineering, University of Texas, Arlington, Arlington, TX 76019, USA.
⁵ Samuel Oschin Comprehensive Cancer Institute, Cedars-Sinai Medical Center, Los Angeles, CA 90048, USA.
⁶ University of California Los Angeles, Los Angeles, CA 90095, USA.
⁷ Department of Urology, Gachon University College of Medicine, Incheon 1 3 1 2 0, South Korea.

Abstract

With the advent of artificial intelligence (AI) in biostatistical analysis and modeling, machine learning can potentially be applied into developing diagnostic models for interstitial cystitis (IC). In the current clinical setting, urologists are dependent on cystoscopy and questionnaire-based decisions to diagnose IC. This is a result of a lack of objective diagnostic molecular biomarkers. The purpose of this study was to develop a machine learning-based method for diagnosing IC and assess its performance using metabolomics profiles obtained from a prior study. To develop the machine learning algorithm, two classification methods, support vector machine (SVM) and logistic regression (LR), set at various parameters, were applied to 43 IC patients and 16 healthy controls. There were 3 measures used in this study, accuracy, precision (positive predictive value), and recall (sensitivity). Individual precision and recall (PR) curves were drafted. Since the sample size was relatively small, complicated deep learning could not be done. We achieved a 76%-86% accuracy with leave-one-out cross validation depending on the method and parameters set. The highest accuracy achieved was 86.4% using SVM with a polynomial kernel degree set to 5, but a larger area under the curve (AUC) from the PR curve was achieved using LR with a l ₁-norm regularizer. The AUC was greater than 0.9 in its ability to discriminate IC patients from controls, suggesting that the algorithm works well in identifying IC, even when there is a class distribution imbalance between the IC and control samples. This finding provides further insight into utilizing previously identified urinary metabolic biomarkers in developing machine learning algorithms that can be applied in the clinical setting.

Keywords: artificial algorithm; biomarker; interstitial cystitis; machine learning; metabolomics; urine.