ASN's Mission

To create a world without kidney diseases, the ASN Alliance for Kidney Health elevates care by educating and informing, driving breakthroughs and innovation, and advocating for policies that create transformative changes in kidney medicine throughout the world.

learn more

Contact ASN

1401 H St, NW, Ste 900, Washington, DC 20005


The Latest on X

Kidney Week

Please note that you are viewing an archived section from 2023 and some content may be unavailable. To unlock all content for 2023, please visit the archives.

Abstract: FR-PO009

Using Deep Learning to Determine Kidney Function Decline in Patients with ADPKD

Session Information

Category: Augmented Intelligence, Digital Health, and Data Science

  • 300 Augmented Intelligence, Digital Health, and Data Science


  • Jang, Hyun Bae, McGill University, Montreal, Quebec, Canada
  • Reinhold, Caroline, McGill University Health Centre, Montreal, Quebec, Canada
  • Najafian, Keyhan, University of Saskatchewan, Saskatoon, Saskatchewan, Canada
  • Maleki, Farhad, University of Calgary, Calgary, Alberta, Canada
  • Alam, Ahsan, McGill University Health Centre, Montreal, Quebec, Canada

Autosomal dominant polycystic kidney disease (ADPKD) is associated with progressive kidney cyst growth, eventually leading to kidney failure in many patients. The current risk stratification tool uses height-adjusted total kidney volume (TKV) by age. However, it does not include clinical or other imaging features that may contribute to kidney function decline. In this study, we developed a deep learning classifier that integrates clinical and imaging features to predict rapid estimated GFR (eGFR) decline in patients with ADPKD.


We included 120 patients with confirmed ADPKD with at least one MRI scan and at least 3 serial eGFR measures. An eGFR decline of 4 mL/min/1.73 m2 per year or greater was defined as rapid progression. We tripled the sample size by extracting the largest manually segmented 2D MRI slice and its two neighbouring slices for each patient and assigning the corresponding clinical features. Pixel-level and spatial-level transformations were applied to the images to enhance our deep model’s generalizability and robustness. The model consisted of three components: 1) EfficientNet-b2, 2) FuseNet, and 3) Classifier with the train:validation:test split of 63:17:20. The EfficientNet-b2 extracted 1000 features from each MRI slice. The FuseNet performed feature fusion on 15 clinically relevant features associated with ADPKD. The feature maps obtained from both models were concatenated and fed into the Classifier. Weighted-average F1 and AUC scores and the confusion matrix were used to assess the model’s performance.


The mean age of the study cohort was 46 years (SD 14), 54% were male, 95% were non-Black, hypertension prevalence was 79%, the median eGFR was 67 ml/min/1.73 m2 (IQR 45-99), height-adjusted TKV was 940 mL/m (IQR 542-1373), tolvaptan use was 53%, and the Mayo Imaging Class was 8%, 17%, 40%, 23%, 10%, and 2% for 1A, 1B, 1C, 1D, 1E and class 2, respectively. The weighted-average F1 and AUC scores and the true positive and true negative values were 0.88, 0.88, 0.88, and 0.88 on the validation set, respectively and 0.83, 0.83, 0.86, and 0.80 on the test set.


Our study demonstrates that a deep learning approach which integrates clinical information with MRI features can successfully classify patients at risk for rapid eGFR decline. Further validation in an external cohort is required.