Predicting deterioration of early sepsis patients using wearables

Sepsis is a severe condition that can lead to septic shock and multiple organ failure if not promptly recognized and treated. Early detection of deterioration is challenging, as continuous monitoring is uncommon in nursing wards. However, the cardiovascular and respiratory system can be monitored through wearables able to capture photoplethysmography (PPG) and diaphragm electromyography (EMG). Continuously measured PPG and EMG waveforms were obtained at the ED of the UMCG. A Matlab algorithm was used to filter and analyze 12-hour measurements, extract features, and determine trends such as the number of slope changes and overall slope. Five supervised Machine Learning models were developed with five-fold cross-validation. We included 189 early septic patients of which 25 deteriorated, defined as death, Intensive Care Unit admission, or a rise in SOFA score by at least 2 points within 48 hours after acute admission. Using PPG trend features, we demonstrated effective predictive modeling. The K-nearest neighbor (KNN) model displayed good performance: the model’s accuracy was 75%, sensitivity 73%, specificity 78%, and AUROC 0.84. This study demonstrates the potential of using PPG trend features for early prediction of patient deterioration. Further research is needed to evaluate the generalization of the models with a separate test set.