Using Machine Learning to Detect Chronic Kidney Disease and Predict Creatinine Levels from Home

Introduction

Chronic Kidney Disease (CKD) is a serious health condition that can lead to kidney failure if not detected early. Traditionally, diagnosing CKD requires blood tests and medical supervision. However, advances in machine learning (ML) are making it possible to predict CKD using simple at-home measurements. This article explores how ML models can classify CKD and predict creatinine levels, helping individuals monitor their kidney health from home.

Data Collection and Source

The data used in this study comes from a publicly available dataset hosted by the University of California, Irvine (UCI). It includes 400 patients from a hospital in Tamil Nadu, India, over a two-month period. Among these patients, 250 were diagnosed with CKD, while 150 were not. The classification was done by nephrologists based on patient history, symptoms, and test results.

The dataset contains essential information like:

• Blood test results
• Comorbidity (other existing medical conditions)
• Demographic data (age, gender, race, etc.)
• Serum creatinine levels (a key indicator of kidney function)

Serum creatinine levels are crucial in calculating the estimated Glomerular Filtration Rate (eGFR), which determines kidney function. In addition, sex and race data were supplemented from other sources to improve the accuracy of the analysis.

Feature Selection: Grouping the Data for At-Home Use

To make kidney health monitoring accessible from home, we categorized features into three sets:

1. At-Home Features

These include demographic information (age, gender, etc.), existing medical conditions, and blood pressure readings, which can be easily measured at home or at a local pharmacy.

2. Monitoring Features

These include at-home features plus blood tests such as blood urea and blood glucose, which can be done at a clinic but are commonly available.

3. Laboratory Features

These include all 25 available features, including specialized tests requiring laboratory equipment.

By testing ML models with different feature sets, we assess whether at-home features alone can provide reliable CKD predictions.

Data Preprocessing: Preparing the Data for ML Models

Before training ML models, we preprocess the data to improve accuracy. The preprocessing steps include:

1. Handling Missing Data

Some patients in the dataset had missing serum creatinine readings. These cases were removed since they couldn't be used for training the regression models.

For missing numerical values, we used the k-nearest neighbors (k-NN) method instead of simple averaging. This method finds the five closest patients (k=5) with complete data and estimates missing values based on them.

2. Encoding Categorical Data

For features like gender or race, which are not numerical, we applied "one-hot encoding." This method converts categories into separate columns, each indicating whether a patient belongs to a category (1 for yes, 0 for no). Any missing categorical values were treated as a separate category.

3. Standardizing Data

To ensure fair comparisons, numerical features were standardized to have a mean of 0 and a standard deviation of 1.

4. Transforming Creatinine Levels

Since serum creatinine levels vary widely, we applied a logarithmic transformation to ensure better predictions and prevent negative values.

Machine Learning Algorithms Used

To classify CKD and predict creatinine levels, we used two powerful ML algorithms: Artificial Neural Networks (ANNs) and Random Forests (RFs).

1. Artificial Neural Networks (ANNs)

ANNs mimic the human brain using layers of connected nodes. They are excellent for recognizing complex patterns but require large amounts of high-quality data. One challenge with ANNs is that they act as "black boxes," meaning their decision-making process isn't always clear.

2. Random Forests (RFs)

RFs use multiple decision trees to make predictions. They combine multiple models to avoid overfitting and improve accuracy. Unlike ANNs, RFs are easier to interpret because we can see how each decision tree splits the data.

Evaluating Model Performance

To assess how well our models perform, we used various evaluation metrics.

For CKD Classification

• Accuracy: Measures overall correctness but may not be reliable if the dataset is imbalanced.
• True Positive Rate (TPR): Measures how well the model detects CKD cases.
• True Negative Rate (TNR): Measures how well the model identifies non-CKD cases.
• False Positive Rate (FPR): Measures how often the model wrongly classifies a non-CKD patient as having CKD.
• False Negative Rate (FNR): Measures how often CKD cases are missed, which is crucial in medical applications.

We also used ROC curves, which show how well a model differentiates between CKD and non-CKD cases. The Area Under the Curve (AUC) provides a single score to summarize model performance.

For Predicting Creatinine Levels

• Mean Squared Error (MSE): Measures how much predictions differ from actual values.
• R-Squared (R²): Measures how well the model explains variance in creatinine levels.
• Mean Absolute Error (MAE): Measures the average prediction error, useful for understanding real-world differences.

Training and Cross-Validation

Since we had a small dataset, we used k-fold cross-validation (k=10) to improve training. This technique splits data into 10 parts, training on 9 parts and testing on 1, then rotating through all parts. This prevents overfitting and provides a more reliable accuracy estimate.

Model Training Process

Here’s how we trained our models step by step:

1. Split data into 10 folds.
2. Reserve 20% of training data for validation.
3. Train models using Keras for ANNs and scikit-learn for RFs.
4. Fine-tune hyperparameters using random search and grid search.
5. Evaluate models on test data.
6. Repeat for all feature sets and algorithms.

Conclusion: Can We Detect CKD from Home?

The results of this study suggest that machine learning models, especially using features measurable at home, can help detect CKD and predict creatinine levels. While laboratory tests remain the gold standard, at-home monitoring using ML can serve as an early warning system, prompting individuals to seek medical attention sooner.

Future advancements in wearable health technology and home testing kits may further improve the accuracy and accessibility of CKD detection. This study demonstrates the potential of AI-driven healthcare solutions in transforming early disease detection and management.

Key Takeaways

• Machine learning can help classify CKD and predict creatinine levels using at-home measurements.
• Preprocessing steps like handling missing data and standardizing features improve model accuracy.
• Random Forests provide better interpretability, while ANNs are more powerful but require more data.
• Using only at-home features, ML models can provide an early warning system for CKD.
• As home health monitoring technology improves, AI-driven healthcare solutions will become even more effective.

By leveraging ML, individuals can take charge of their kidney health with simple at-home monitoring, leading to earlier intervention and better health outcomes.