Machine Learning Study on the Relationship Between Heart Rate and Body Temperature in Pediatric Intensive Care Units

Academic Background

In the pediatric intensive care unit (PICU), heart rate (HR) and body temperature (BT) are crucial clinical indicators that reflect a patient’s physiological status. Although the relationship between HR and BT has been extensively studied in adults, research remains limited in children, especially in high-risk environments like the PICU. The physiological characteristics of pediatric patients differ significantly from those of adults, particularly in the 0 to 18 age range, where HR tends to decrease with age while changes in BT can influence HR. However, traditional linear models often show limitations when predicting HR, particularly underestimating or overestimating values across different BT ranges and age groups. Therefore, exploring the complex relationships between HR, BT, and age is vital for improving clinical decision-making in the PICU.

Paper Source

This paper was co-authored by Émilie Lu, Thanh-Dung Le, Philippe Jouvet, and Rita Noumeir from École de Technologie Supérieure, University of Luxembourg, and CHU Sainte-Justine Hospital. Published in 2024 in IEEE Transactions on Biomedical Engineering, the study received partial funding from the Natural Sciences and Engineering Research Council (NSERC) and Fonds de la Recherche en Santé du Québec (FRQS).

Research Workflow

1. Data Collection

The study utilized data from the PICU at CHU Sainte-Justine Hospital, spanning from August 2018 to October 2022. A total of 4,007 children aged 0 to 18 were included, with their HR and BT data collected during their stay in the PICU. BT data was recorded every 30 seconds, supplemented by manual measurements every 2 to 4 hours. HR data was continuously monitored via electrocardiogram (ECG) or pulse oximeter, recorded at one-second intervals. To ensure data accuracy, patients undergoing extracorporeal membrane oxygenation (ECMO), using pacemakers, or supported by Berlin Heart devices were excluded.

2. Data Preprocessing

Data preprocessing is a critical step to ensure model accuracy, including the following: - Excluding Movement Data: Utilized scales such as Cornell Assessment of Pediatric Delirium (CAPD), COMFORT-B, and FLACC to assess patient movement states and exclude data affected by patient motion or agitation. - Temperature Normalization: Adjusted axillary temperature data to align with rectal temperature readings for consistency. - Excluding Drug Effects: Excluded data during treatment with drugs affecting HR (e.g., beta-blockers, dopamine). - Heart Rate Median Calculation: Aggregated per-second HR data into per-minute medians to reduce data volume. - HR and BT Association: Associated HR and BT data within a 10-minute window to ensure synchronization. - Excluding Extreme Values: Removed extreme HR values below 30 bpm or above 240 bpm, and BT values below 30°C or above 43°C. - Temperature Grouping: Grouped BT data into 1°C intervals, ranging from 33°C to 40.9°C. - Single Observation Retention: Retained a single observation closest to the median value for each patient within each 1°C BT range to ensure balanced data.

3. Machine Learning Modeling

The study employed various machine learning (ML) and deep learning (DL) models to capture the relationships between HR, BT, and age, including: - Linear Regression Models: Simple linear regression (LR), multiple linear regression (MLR), and support vector machines (SVM) were used to validate traditional linear assumptions. - Quantile Regression (QR): Combined with gradient boosting machines (GBM), random forests (RF), recurrent neural networks (RNN), etc., to capture non-linear and complex patterns. - Gradient Boosting Machines (GBM): Iteratively optimized decision tree models to progressively reduce prediction errors, suitable for capturing non-linear relationships. - Neural Network Models: Included multilayer perceptrons (MLP) and long short-term memory networks (LSTM) for handling time-series data and complex relationships.

4. Model Evaluation

Model performance was assessed using R-squared (R²) and Mean Squared Error (MSE) for linear models, while quantile regression models were evaluated using Quantile Loss. Hyperparameter tuning was conducted via grid search and early stopping techniques to ensure optimal model performance.

Key Results

1. Relationship Between HR and Age

Results showed that HR decreases with age, consistent with Pediatric Advanced Life Support (PALS) data. For instance, newborns’ HR ranged from 85 to 205 bpm, while adolescents’ HR ranged from 60 to 100 bpm.

2. Relationship Between HR and BT

The study found that HR increases with rising BT. Linear models underestimated HR at lower BT ranges and overestimated it at higher BT ranges, especially in younger children. The QR combined with GBM model performed best, accurately capturing the non-linear relationship between HR and BT.

3. Model Performance

The GBM model exhibited the lowest Quantile Loss in quantile regression, indicating superior accuracy and robustness in predicting the HR-BT relationship. In contrast, traditional linear models (e.g., LR, SVM) performed poorly in explaining HR variability, with R² values ranging from 0.3145 to 0.3576.

Conclusions and Significance

This study demonstrates that the relationship between HR, BT, and age is not a simple linear one, and traditional linear models have limitations in the PICU environment. By introducing advanced ML techniques like quantile regression and gradient boosting machines, the study more accurately captured the complex dynamics among these physiological indicators. This not only provides a more reliable predictive tool for clinical decision-making but also opens new avenues for future research in pediatric critical care. Additionally, a user-friendly interface was developed to help clinicians predict HR in real-time based on a patient’s age and BT, optimizing treatment plans.

Research Highlights

1. Revealing Non-Linear Relationships: The study first revealed the non-linear relationship between HR and BT in the PICU environment, challenging traditional linear assumptions.
2. Application of Advanced Machine Learning Models: The introduction of advanced algorithms like quantile regression and gradient boosting machines significantly improved prediction accuracy and robustness.
3. Clinical Application Value: Developed an ML-based predictive tool capable of assisting real-time clinical decision-making, enhancing the quality of care for PICU patients.

Future Research Directions

Future studies could further validate the model’s generalizability by expanding datasets to include more patient populations and clinical scenarios. Additionally, exploring the impact of other physiological indicators (e.g., respiratory rate, blood pressure) on HR, as well as the role of gender differences in the HR-BT relationship, are worthwhile areas for further investigation.