Online Poster Portal

  • Author
    Ryan Tiu
  • Discovery PI

    Isidro Salusky

  • Project Co-Author

    Jeffrey Feng, Tomas Ganz, Renata C. Pereira, Barbara Gales, Alex A.T. Bui, Isidro B. Salusky

  • Abstract Title

    Value of Machine Learning as a Predictor of Bone Pathology in Pediatric Chronic Kidney Disease-Mineral and Bone Disorder Patients Utilizing Biomarkers and Anthropometric Parameters

  • Discovery AOC Petal or Dual Degree Program

    Basic, Clinical, & Translational Research

  • Abstract

    Background: Bone histomorphometry remains the gold standard for evaluating bone turnover and mineralization at the cellular level and diagnosing renal osteodystrophy. As opposed to biochemical examination, bone biopsy is an invasive method with limited use due to lack of training in the technical expertise and tissue handling of bone biopsy. Thus, we developed a machine learning algorithm to predict bone turnover (BT) and mineralization (BM) based on serum biomarkers, height, and weight in pediatric CKD-MBD (PCKD-MBD) patients.

    Objective: Using machine learning (ML) to predict BT and BM based on biomarkers, height, and weight in (PCKD-MBD patients.

    Methods: This cross-sectional study includes 659 iliac crest bone biopsies from PCKD-MBD patients performed at UCLA from 1983-2023. Biomarkers (PTH (first and second generation IRMA Immutopics), phosphate, vitamin D, comprehensive metabolic panel, complete blood count, iron studies), height, and weight were measured at or near the time of biopsy. Bone histomorphometry was assessed using the OsteoMetrics system; all biopsies were read by the same pathologist (RP). ML models (logistic regression, random forest (RF), extreme gradient boosting) were used to predict BT and BM from biomarkers, height, and weight. SHapley Additive exPlanations were used to analyze feature importance.

    Results: On a holdout test set, RF achieved the highest area under the receiver operating curve of 0.687 (95% CI = 0.685-0.689) and 0.828 (95% CI = 0.826-0.829) for predicting BM and BT, respectively. The top features contributing to abnormal BM were higher weight and lower height, BMI Z-scores, calcium (Ca), creatinine (Cr), BUN, and phosphate Z-scores. Higher PTH, ALP, Cr, and fibroblast growth factor 23 (FGF23) and lower Ca were influential in predicting high BT. Iron deficiency was also a predictor of abnormal BM and BT. BM prediction errors were often due to low Ca despite normal BM, anthropometric parameters overshadowing other features, and low Cr unexpectedly predicting abnormal BM. The absence of strong biochemical correlates with mineralization lag time may account for weaker BM prediction. BT prediction errors often stemmed from elevated PTH and ALP despite low or normal BT.

    Discussion: This is the largest dataset of bone biopsies from PCKD-MBD patients. ML shows potential in predicting BT and BM from biomarkers. Low height as a predictor of abnormal BM is consistent with the growth retardation and bone deformities seen in CKD. Low or high hepcidin predicting normal or abnormal BM, respectively, along with the impact of elevated FGF23 and iron status, reflect the complex interplay between declining renal function, increased inflammation and anemia on bone and mineral metabolism.

    Conclusion: ML demonstrates potential in predicting BT and BM from biomarkers and anthropometric parameters for guiding CKD-MBD therapy in institutions unable to perform bone biopsies.