Machine learning based clinical decision tool to predict acute kidney injury and survival in therapeutic hypothermia treated neonates

Abstract

Therapeutic hypothermia (TH) significantly reduces mortality and morbidities in neonates with Neonatal Encephalopathy (NE). NE may result in neonatal death and multisystem organ impairment, including acute kidney injury (AKI). Our study aimed to utilize machine learning (ML) methods to predict the outcome of TH-treated NE neonates developing AKI and death during TH. In this retrospective multinational study, 1149 TH-treated NE neonates and 801 controls were included. AKI was classified using KDIGO neonatal criteria based on serum creatinine measurements. The ML model incorporated gestational age, birth weight, postnatal age, and serum creatinine values. The algorithm used all these covariates to predict one of five outcomes: survival with/without AKI, mortality with/without AKI, and hospitalized non-NE controls. The XGBoost model achieved an AUC of 95% and an accuracy of 75.08% in predicting AKI and survival, surpassing other ML classifiers that demonstrated accuracy levels ranging from 54% to 65%. To our knowledge this is the first ML model trained on multicenter, multinational data specifically aimed at predicting neonates' AKI, death, and survival within the first three days. Our ML scoring systems' code and user interface are freely available ( https://github.com/NUBagciLab/Therapeutic-Hypothermia-Outcome-Classification , https://thprediction.streamlit.app/ ). This tool has potential to support neonatologists to personalize therapies, and to optimize pharmacotherapy for renally cleared drugs.

Fig. 1

Graphical abstract of the newly developed model. Our algorithm was built on using four types of data: gestational age, birth weight, postnatal age, and serum creatinine observations during TH as input and predicting one of five classes as an outcome: (1) TH-treated NE neonates who survived, did not have AKI, (2) TH-treated NE neonates who survived and had AKI, (3) TH-treated NE neonates who died, did not have AKI; (4) TH-treated NE neonates who died and had AKI, (5) neonates without NE who did not need TH.

Fig. 2

Calibration curve of our single classifier approach model. This plot shows the calibration of the predicted probabilities across five classes (Class 1 to Class 5). The x-axis represents the mean predicted probability, while the y-axis represents the fraction of positives (observed frequency). Each line corresponds to a different class, with the dashed diagonal line (“Perfectly calibrated”) indicating an ideal calibration where predicted probabilities perfectly match the observed outcomes. Deviation from the diagonal suggests miscalibration, where probabilities either underestimate or overestimate the actual outcomes.

Fig. 3

Confusion matrix of our developed model. The confusion matrix illustrates the percentage of cases correctly predicted by our model. The y-axis represents the true labels, while the x-axis represents the predicted labels. The bold boxes highlight the percentage of true labeled patients correctly predicted by the model. For instance, the model correctly predicted “TH-treated NE neonates survived without AKI” in 80.65% of cases.

Fig. 4

ROC of our model predicting clinical outcome in TH-treated neonates. ROC graphs for our model, cross-validated using ten-fold validation. During each of the ten folds, the model undergoes training and testing, and the Area Under the Curve (AUC) is computed for each iteration. The ten AUC values are combined to yield a singular performance metric through averaging. The model’s mean AUC of 95% over all 10 folds demonstrates its constant and excellent performance in accurately differentiating between classes.

Fig. 5

Serum Creatinine concentrations for different patient groups in our model. The graph displays the simplified median trend lines creatinine (sCr) observations for each label in our model over a period of 10 days. These are actual values from the model. The baseline sCr values for each labeled group are shown, along with their trends over time.

Fig. 6

Serum creatinine concentrations of neonates with AKI and without AKI. The graph illustrates the median creatinine (sCr) trends of neonates with AKI (both those who survived and those who died) and neonates without AKI. The neonates without AKI are further divided into three groups: hospitalized non-NE neonates, neonates who died, and neonates who survived.

Fig. 7

Serum creatinine concentrations for all TH-treated NE neonates and control neonates. The graph shows the median creatinine (sCr) trend lines for neonates who did not have AKI, divided into three groups. The creatinine trends for these groups overlapped during their NICU stay, and the median creatinine values were close to each other. This illustrates the difficulty in differentiating between these groups based solely on creatinine values.

Fig. 8

Serum creatinine concentrations for all TH-treated NE neonates who survived and who died in the dataset. The graph illustrates the median creatinine (sCr) levels for neonates who survived and those who died. The trend lines show that survived neonates had lower creatinine levels compared to those who died.

Fig. 9

Percentiles of serum creatinine concentrations based on postnatal age on the study dataset. The graph shows the percentiles of serum creatinine (sCr) concentrations based on postnatal age for the entire study dataset. The median centile lines illustrate the sCr values for both NE neonates and hospitalized non-NE neonates.

Fig. 10

Features’ importance of each classifier in the hierarchical classification model. The ranking of features (variables) plays a crucial role in predicting outcomes for a hierarchical classification model. Pink, green, pastel orange, and purple lines represent each of the outcome predictions. Input variables interact with each other for the outcome prediction. Creatinine, postnatal age interaction, and creatinine play important roles in algorithmic decision-making.

Fig. 11

Features’ importance of the single classifier model approach. Ranking the importance of features in predicting outcomes using a “single classifier model”. The relationship between creatinine and gestational age is crucial in the decision-making process of algorithms.