An integrated machine learning and hyperparameter optimization framework for noninvasive creatinine estimation using photoplethysmography signals

Frequent measurement of creatinine levels is vital for patients with chronic kidney disease. Traditional creatinine level measurement requires invasive blood test which has several disadvantages like discomfort, anxiety, panic, pain, risk of infection, etc. To address the issue, we propose a noninvasive machine learning (ML) model-based method to estimate creatinine level using photoplethysmography (PPG) signal. We obtained the PPG signal and gold-standard serum creatinine level of 404 patients from the Medical News Mart for Concentrated Care III (MIMIC III) database. In data preprocessing, we analyzed the PPG signal following several steps and created PPG feature set. We used multiple feature engineering methods to identify the most important features. We integrated Optuna, a hyperparameter optimization framework, with every ML model to get the optimal hyperparameters. We developed five ML models and compared their performance both with and without the application of Optuna. We found that Optuna significantly improves every model's performance. With Optuna, extreme gradient boosting (XGBoost) performed best among all five models. This XGBoost model had an accuracy of 85.2 %, an average k-fold cross validation score (k = 10) of 0.70, and a “receiver operating characteristic area under the curve” (ROC-AUC) score of 0.80. With the high performance exhibited by our developed model, the study can play a crucial role in the field of noninvasive creatinine estimation and diagnosis of chronic kidney disease. • Present machine learning (ML) model for estimating creatinine levels noninvasively using PPG signals. • Utilize the MIMIC III database, comprising PPG signals and gold-standard serum creatinine levels of patients. • Conduct data preprocessing on PPG, including feature extraction and feature engineering to identify important features. • Integrate Optuna, a hyperparameter optimization framework, to optimize models’ performance. • Employ various machine learning classification algorithms to select the best-performing model.