An artificial intelligence-based approach to identify volume status in patients with severe dengue using wearable PPG data

Dengue shock syndrome (DSS) is a serious complication of dengue infection which occurs when critical plasma leakage results in haemodynamic shock. Treatment is challenging as fluid therapy must balance the risk of hypoperfusion with volume overload. In this study, we investigate the potential utility of wearable photoplethysmography (PPG) to determine volume status in DSS. In this prospective observational study, we enrolled 250 adults and children with a clinical diagnosis of dengue admitted to the Hospital for Tropical Diseases, Ho Chi Minh City. PPG monitoring using a wearable device was applied for a 24-hour period. Clinical events were then matched to the PPG data by date and time. We predefined two clinical states for comparison: (1) the 2-hour period before a shock event was an “empty” volume state and (2) the 2-hour period between 1 and 3 hours after a fluid initiation event was a “full” volume state. PPG data were sampled from these states for analysis. Variability and waveform morphology features were extracted and analyzed using principal components analysis and random forest. Waveform images were used to develop a computer vision model. Of the 250 patients enrolled, 90 patients experienced the predefined outcomes, and had sufficient data for the analysis. Principal components analysis identified four principal components (PCs), from the 23 pulse wave features. Logistic regression using these PCs showed that the empty state is associated with PCs 1 (p = 0.016) and 4 (p = 0.036) with both PCs denoting increased sympathetic activity. Random forest showed that heart rate and the LF-HF ratio are the most important features. A computer vision model had a sensitivity of 0.81 and a specificity of 0.70 for the empty state. These results provide proof of concept that an artificial intelligence-based approach using continuous PPG monitoring can provide information on volume states in DSS.Author summary: Dengue is a globally important public health threat with an estimated 100 million symptomatic infections occurring each year. Dengue shock syndrome (DSS) is a complication of disease that occurs when progressive hypovolemia leads to circulatory collapse. Treatment of DSS is supportive, requiring careful fluid management over a 24-to-48-hour critical phase depending on patients’ volume status. However, a reliable way to assess volume status is lacking. In this study, we show that photoplethysmography (PPG) monitoring for 24 hours reflects volume status in patients with DSS. We predefined two opposite volume states - “empty” or “full” based on clinical events then compare their corresponding PPG signals. Using machine learning we showed that the empty state signal is characterized by a phenotype of sympathetic dominance. This is a novel finding that highlights the underlying pathophysiology of DSS. Furthermore, a computer vision model, using the PPG signal images as inputs, had a sensitivity of 81% in detecting the empty state. We expect that with refinement, artificial intelligence-based models applied to continuous PPG monitoring will enhance the way that acute DSS is treated, enabling dynamic and personalized fluid management.