Kidney Week

Abstract: FR-PO921

Development of Machine Learning-Based Biopsy Algorithm for Kidney: Multicenter-Based Model Development and Validation Study in South Korea

Session Information

• Glomerular Diseases: Potpourri
  October 25, 2024 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Glomerular Diseases

• 1402 Glomerular Diseases: Clinical, Outcomes, and Therapeutics

Authors

• Yun, Hae-Ryong, Yonsei University College of Medicine, Seodaemun-gu, Seoul, Korea (the Republic of)

Hae-Ryong Yun, MD, PhD

• Kim, Hyung Woo, Yonsei University College of Medicine, Seodaemun-gu, Seoul, Korea (the Republic of)

Hyung Woo Kim, MD

• Park, Jung Tak, Yonsei University College of Medicine, Seodaemun-gu, Seoul, Korea (the Republic of)

Jung Tak Park, MD, PhD

• Han, Seung Hyeok, Yonsei University College of Medicine, Seodaemun-gu, Seoul, Korea (the Republic of)

Seung Hyeok Han, MD, PhD

• Kang, Shin-Wook, Yonsei University College of Medicine, Seodaemun-gu, Seoul, Korea (the Republic of)

Shin-Wook Kang, MD, PhD

• Yoo, Tae-Hyun, Yonsei University College of Medicine, Seodaemun-gu, Seoul, Korea (the Republic of)

Tae-Hyun Yoo, MD, PhD

Background

Kidney biopsy has become an essential procedure for microscopic assessment of kidney disease. However, the decision of kidney biopsy can be inconsistent due to the complexity of interpreting patient data and varying clinician expertise. Therefore, based on three nephrologists' decisions, we developed three machine-learning (ML) algorithms, integrating them into a biopsy recommendation algorithm for kidney.

Methods

First, using 8,228 patients with proteinuria, hematuria, and/or unexplained kidney function abnormalities who visited outpatient clinics between 2010 and 2018, we trained a deep-learning algorithm to mimic nephrologists' decisions for kidney biopsy. Subsequently, we tested the performance of the biopsy algorithm on 760 patients who underwent kidney biopsy between 2020 and 2023 in Severance Hospital. The XGBoost was employed to develop the algorithm and the confusion matrix was used to evaluate the performance of the algorithm.

Results

Among the 8,228 patients, the number of cases in which nephrologists decided to conduct kidney biopsy was 1,290 (15.6%), 1,261 (15.3%), and 1,106 (13.4%), respectively. The area under the receiver operating characteristic curve (AUROC) for each ML-algorithm was 99.6%, 99.8%, and 97.0%, respectively. These results suggest that each algorithm closely emulates the decisions of corresponding nephrologists in determining the need for kidney biopsy, demonstrating proficiency in accurately replicating clinical decisions. The integrated final kidney biopsy algorithm, combining the three algorithms, achieved an AUROC of 99.5%. Finally, we applied the algorithm to the patients who underwent kidney biopsy. Among the 760 patients, 659 (86.7%) received a diagnosis of glomerular, tubule-interstitial, or inherited diseases. The AUROC for the algorithm was 86.2%. In addition, the positive predictive value of the algorithm was 100.0 %.

Conclusion

We developed an ML-algorithm to emulate nephrologists' decisions on kidney biopsy. We expected that this algorithm could support clinicians in strengthening the decision-making process and simplifying patient management workflows for kidney biopsy in primary care.