Predictions of Total Serum Bilirubin From Transcutaneous Bilirubin Using Supervised Machine Learning Models

Varun Sundarrajan, Dhanam Venkatachalam Suresh, Maheshwari Doreraju, Suresh Viswanathan, Santhanakrishnan Ramakrishnan

Abstract

Aim: To predict total serum bilirubin (TsB) from transcutaneous bilirubin (TcB) using supervised machine learning algorithms, and to identify the best-performing model.

Materials and Methods: This was a single-center, cross-sectional, observational study. Neonates admitted to SKS Hospital & Postgraduate Medical Institute (Salem, Tamil Nadu, India) who underwent routine blood sampling for clinical indications were considered. Neonates who previously received phototherapy were excluded. TcB was measured from the ear lobes using the BilicareTM device simultaneously within 15 minutes of the clinically indicated blood sample. Supervised machine learning models including linear regression, ridge regression, lasso regression, polynomial regression, K-nearest neighbors, support vector regression, gradient boosting, random forest, and decision tree were used to train and test the TcB and TsB datasets, to assess the predictive metrics of TsB from TcB. Performance metrics such as the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and R2 were calculated. Regression equations of the best performing models have been reported.

Results: Of the 171 neonates, 101 (59%) were male, and 70 (41%) were female. Their median gestational age was 37 weeks (27 41 wk). Their mean birth weight was 2.592 kg (SD: 0.607 kg). The linear, lasso, and ridge regression models outperformed other regression models in accurately predicting TsB from TcB. MAE and RMSE were close to 1.0, and the R2 was 0.85 in these 3 models. The regression equations of the 3 best-performing models are: linear regression equation = 0.9410 × TcB + 0.0849; ridge regression equation = 0.9404 × TcB + 0.0911; and lasso regression equation = 0.8499 × TcB + 0.9790.

Conclusion: It is possible to accurately predict TsB from TcB using supervised machine learning models such as the linear, lasso, and ridge regression. When interpreting the predicted bilirubin values, physicians must use their clinical judgement to decide on the management plan.