Dengue fever poses a significant global health burden, with 5-10% of symptomatic individuals developing severe disease characterized by vascular leakage, potentially leading to cardiovascular collapse and death. Close monitoring of vital signs is crucial for identifying high-risk patients and guiding supportive therapies like fluid resuscitation. However, consistent monitoring can be challenging in resource-limited settings due to equipment limitations, training gaps, and staff shortages. Non-invasive wearable devices, utilizing photoplethysmography (PPG), offer a cost-effective solution for continuous individualised monitoring. PPG, a safe and low-power modality, measures arterial waveform distortion through skin tissue. Growing evidence supports its clinical effectiveness in various applications, including surgery and remote patient monitoring. This study aimed to leverage PPG signals from a wearable device to predict dengue deterioration in a real-time system, assisting in clinical prioritization and guiding fluid therapy.

Existing clinical tools, such as those from the WHO, are used for dengue triage and hospitalization criteria. However, these single-point evaluations may not capture the dynamic nature of the disease. Regular monitoring allows for better clinical prioritization, improving outcomes and healthcare effectiveness. Wearables are beneficial in resource-limited settings, providing cost-effective individualized monitoring. Previous research has shown that PPG signals are predictive of shock in patients with severe dengue admitted to intensive care, and this study builds upon that evidence by exploring the application of this technology in a broader hospital setting.

This prospective observational study was conducted at the Hospital for Tropical Diseases in Ho Chi Minh City, Vietnam, enrolling 153 patients (age 8+) with a clinical dengue diagnosis and hospitalization less than 28 hours. Patients wore a battery-operated wrist wearable (SmartCare Analytics) continuously recording PPG signals (100 Hz) for up to 24 hours. Clinical data (age, sex, presenting symptoms, vital signs, treatment) were also collected. The primary outcome was the prediction of dengue shock syndrome (DSS) 2 hours in advance, based on WHO criteria. Secondary outcomes included hourly illness severity scores using NEWS2 and mSOFA scores. A low clinical risk was defined as NEWS2 <6 and mSOFA <6, without shock or significant bleeding. The dataset was split into training (n=132) and hold-out test sets (n=21). Several deep-learning models were used, including a Spatiotemporal Fusion Transformer (STF) with multi-task learning (MTL) and single-task learning (STL), convolutional neural networks (CNN) with and without long short-term memory (LSTM). Model performance was evaluated using AUROC, AUPRC, F1 score, precision, and recall.

The optimized STF model with MTL predicted low-risk states 2 hours ahead with an AUROC of 0.83 and AUPRC of 0.67. However, predicting DSS showed significantly lower performance. Sensitivity analyses using clinical features alone showed poorer discrimination than models incorporating PPG data. The study also noted that approximately 43% of the collected data was of sufficient quality for analysis due to factors including movement artifacts and noise, and the wearable device's battery life.

This study demonstrates the potential of a low-cost PPG wearable for real-time risk stratification in hospitalized dengue patients. The integration of PPG data with clinical information significantly improved predictive performance compared to using clinical data alone. The relatively lower performance in predicting DSS might be due to class imbalance in the data and potential bias in DSS diagnosis. While the model shows promise, the limitations related to data quality and device usability highlight areas for improvement.

This study demonstrates the feasibility and potential of using a low-cost, non-invasive PPG wearable for continuous monitoring of dengue patients, enabling real-time risk stratification. While limitations related to data quality and device design exist, the findings suggest avenues for improving the technology and its clinical implementation. Future research should focus on addressing these limitations, such as improving hardware design for better comfort and signal quality, and exploring larger, more diverse datasets to enhance model performance.

The study was limited by the relatively small sample size, the exclusion of a substantial portion of collected data due to signal quality issues and the wearable device's limited battery life, potentially impacting the generalizability of the findings. Additionally, the relatively lower performance in predicting dengue shock syndrome could be attributed to class imbalance and potential diagnostic bias. The device's form factor also impacted patient adherence and comfort.