Predicting Adverse Outcomes in Kidney Transplant Recipients Using an Interpretable Model Based on Shear-Wave Elastography

Figures

• 

  Fig. 1.. Comparison of SWE and pathology findings in transplanted renal tissue. a, b Gray scale images. c, d SWE images. Circles 1 and 2 indicate the measurement results from a single ROI in the cortex and medulla, respectively. e, f Pathology results of renal tissue in the transplanted kidney (renal biopsy slides were scanned at ×40 magnification, Masson’s trichrome stain).a, c, e A 67-year-old male patient with mild fibrosis, with an IFTA score of 0. The Emean results of 6 ROIs in the cortex and medulla were 13.25 kPa and 7.83 kPa, respectively. b, d, f A 56-year-old male patient with severe fibrosis, with an IFTA score of 2. The Emean results of 6 ROIs in the cortex and medulla were 20.20 kPa and 15.16 kPa, respectively. IFTA, interstitial fibrosis and tubular atrophy; ROI, region of interest.

• 

  Fig. 2.. Box plot of AUC for machine learning models with 20 repeats of nested tenfold cross validation. Comparison the AUC distribution of the five machine learning methods. The reference model included albumin, hemoglobin, eGFR and BUN (blue boxes). The model with SWE features included median medullary SWE and mean cortical SWE in addition to the reference model (red boxes). SVM, support vector machine; LR, logistic regression; H2O.DL, H2O’s auto-deep learning; RF, random forest; XGBoost, Extreme Gradient Boosting; BUN, blood uric nitrogen; eGFR, estimated glomerular filtration rate.

• 

  Fig. 3.. ROC curves of final XGBoost model in training set and validation set.

• 

  Fig. 4.. Model interpretation and visualization for the XGBoost prediction model. a Features importance ranking based on the XGBoost prediction model for the validation set. b The SHAP values of features in the XGBoost prediction model for the validation set. BUN, blood uric nitrogen; eGFR, estimated glomerular filtration rate.

• 

  Fig. 5.. The SHAP force plot for 2 individual patients of the validation set. a A 44-year-old female patient was predicted to have a 49.50% probability (0.02 log-odds ratio) of the composite outcome mainly due to high value of median medullary SWE, mean cortical SWE and low albumin. This patient experienced a 38.5% decline in eGFR from 76.62 to 47.13 mL/min/1.73 m2 decline with follow-up of 13.5 months. b A 49-year-old male patient was predicted to have a low risk with 11.61% probability (−2.03 log-odds ratio). This patient had a stable eGFR during a 22.5-month follow-up; the latest eGFR was 72.35 mL/min/1.73 m2. BUN, blood uric nitrogen; eGFR, estimated glomerular filtration rate.

• 

  Fig. 6.. SHAP dependence scatter plots for SWE features. a, b The SHAP dependence plot of median medullary SWE for the train set and validation set, respectively, the value of median medullary SWE over approximately 10 kPa was associated with a higher risk of graft deterioration. c, d The SHAP dependence plot of mean cortical SWE for the train set and validation set, respectively, the cutoff value of mean cortical SWE was around 16 kPa.