Detection of Non-Diabetic Kidney Disease in Diabetic Patients Using Machine Learning

Biopsy-confirmed non-diabetic kidney disease was present in 53% of the 1,136 diabetic patients studied. The dataset contains 103 parameters from electronic medical records used to train seven machine learning models, with the best achieving an AUC of 0.841 for detecting NDKD. This resource supports the development of computational diagnostic tools for patients where invasive kidney biopsy is not feasible.

Use Cases

• Training a binary classification model to predict non-diabetic kidney disease based on demographic and lab test data.
• Performing feature importance analysis (e.g., using SHAP) to identify key clinical indicators like hematuria or IgG levels.
• Benchmarking machine learning models (XGBoost, SVM, MLP) for clinical prediction tasks on a real-world patient cohort.
• Studying the relationship between diabetic retinopathy status and other clinical parameters in kidney disease diagnosis.

Strengths

• Contains biopsy-confirmed ground truth labels for 1,136 patients, providing a reliable clinical endpoint.
• Includes a large set of 103 clinical parameters from EMRs, covering demographics, physical exams, and laboratory tests.
• Model performance is documented with specific AUC metrics (e.g., 0.833 for XGBoost) and statistical comparisons.

Limitations

• The specific 103 column names and data types are not provided, limiting precise data understanding.
• The dataset appears to be a supplementary PDF file for a research article, not the raw structured data itself.
• Geographic and institutional origin of the patient cohort is not specified.

Provenance

Source

Data collected from diabetic patients who underwent kidney biopsy at a tertiary teaching hospital.

Collection Method

103 parameters extracted from electronic medical records.

Time Range

null

Freshness

Last updated on 2026-03 18.

Geography

null