Predicting Hemodynamic Shock from Thermal Images using Machine Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Critical Care Medicine

Background:

• Hemodynamic shock detection is crucial for preventing organ failure and mortality.
• Thermal imaging offers a non-invasive method to assess body surface temperature and detect perfusion disturbances.
• Early and accurate shock detection in pediatric intensive care units remains a challenge.

Purpose of the Study:

• To automate the early detection and prediction of hemodynamic shock using machine learning on thermal images.
• To develop a non-invasive, non-contact decision support system for shock monitoring.
• To evaluate the performance of a machine learning model in predicting shock at various time points.

Main Methods:

• Utilized thermal images from a pediatric intensive care unit.
• Employed Histogram of Oriented Gradient features for machine learning-based region-of-interest segmentation, achieving 96% agreement with expert analysis.
• Developed a generalized linear mixed-effects model using segmented center-to-periphery temperature difference and pulse rate for longitudinal shock prediction.

Main Results:

• The machine learning model achieved a mean area under the receiver operating characteristic curve (AUC) of 75% for shock classification at 0 hours.
• The model demonstrated predictive capabilities with AUCs of 77% at 3 hours and 69% at 12 hours.
• The segmentation method showed high agreement with human expert identification of relevant regions.

Conclusions:

• The developed machine learning model effectively detects and predicts hemodynamic shock using non-invasive thermal imaging.
• This approach offers an affordable, non-contact, and tele-diagnostic decision support system for reliable shock management.
• The findings highlight the potential of thermal imaging combined with AI for proactive critical care in pediatric patients.