Predicting deterioration in dengue using a low cost wearable for continuous clinical monitoring

Dengue is a major global health burden; 5–10% of symptomatic cases progress to severe disease with vascular leakage, shock, and potential death. Continuous monitoring of vital signs aids early identification of deterioration and guides supportive therapy. Existing triage tools (e.g., WHO criteria) are typically applied at a single time point and may miss the dynamic progression of dengue. Regular, repeated monitoring is challenging in many low- and middle-income settings due to resource constraints. Non-invasive wearables using photoplethysmography (PPG) offer low-power, safe, continuous monitoring and have shown clinical effectiveness in other contexts. Prior work suggests PPG features can predict shock in severe dengue. This study evaluates whether a low-cost PPG wearable combined with deep learning can predict real-time clinical deterioration and stratify risk among hospitalized dengue patients, potentially enabling safer ambulatory management and better prioritization of care.

The paper references the growing use of PPG-based wearables for pulse oximetry and broader cardiovascular monitoring, including applications in low-income surgical settings and remote monitoring (e.g., COVID-19). Prior dengue-specific work indicates PPG-derived indices can predict recurrent shock in intensive care. The study situates itself within literature on early warning scores like NEWS2 and disease-specific scores (mSOFA) for prognostication, as well as on machine learning approaches for time-series physiological data, including CNNs, LSTMs, and transformer-based models. It highlights the potential of multi-task learning to leverage related outcomes to improve prediction and the importance of AUPRC under class imbalance.

Design and setting: Prospective observational study at the Hospital for Tropical Diseases (HTD), Ho Chi Minh City, Vietnam, from January 2020 to October 2022. Target sample size was 250 based on feasibility; 250 enrolled, 153 included in analyses after data quality exclusions. Participants: Inclusion criteria: age ≥8 years, clinical diagnosis of dengue (NS1 positive), and hospitalization <28 h at enrollment. Exclusion: lack of consent/assent or clinical reasons at treating staff’s discretion. Recruitment occurred in ED, ICU (adult/paediatric), and general wards. Data collection: Continuous PPG signals recorded using a battery-operated wrist wearable with a transmissive finger probe (SmartCare Analytics, Oxford, UK), capturing red/infrared wavelengths at 100 Hz. Patients were instructed to wear up to 24 h post-enrollment. Clinical data collected at enrollment included demographics, presenting symptoms, vital signs (BP, pulse, respiratory rate, temperature, SpO2), and subsequent vital signs and treatments during 24 h as part of routine care (observations every 1–6 h according to clinical need). PPG data were stored onboard and downloaded after recording; data were also streamed to a mobile application. Outcomes: Primary—prediction of clinical shock (dengue shock syndrome [DSS] or recurrent shock) as a discrete event 2 h ahead, per WHO 2009 definitions, using 10-min PPG segments. Secondary—hourly patient illness severity during the initial 24 h derived from: (i) NEWS2 (aggregating derangements in RR, pulse, SpO2, SBP, consciousness, temperature), and (ii) dengue-specific mSOFA (validated locally). A binary low-risk state was defined as NEWS2<6 and mSOFA<6, without shock or significant bleeding. Signal processing and features: PPG segmented into overlapping fixed windows (10 min minimum sequence length; segments incremented by 1 min). Segments with insufficient labeled data were discarded. Signal quality indices included matching across multiple systolic wave detection algorithms and zero-crossing adaptations for PPG. Clinical features for multi-modal models: age, weight, height, presence/absence of headache, vomiting, diarrhoea, abdominal pain, hypertension, diabetes. Missing clinical observations used last observation carried forward. Modeling: Multiple time-series deep learning models were evaluated: Spatiotemporal/Temporal Fusion Transformer-based spatio-temporal fusion framework (STF), CNN, and CNN-LSTM. Spectrograms of 10-min PPG sequences were generated via Short-Time Fourier Transform and fed through convolutional and recurrent layers, with attention to model longer-term dependencies. Multi-task learning (MTL) was used to jointly predict dengue shock and illness severity to improve training efficiency and performance; single-task learning (STL) variants were also assessed. Multi-modal models integrated PPG with clinical/demographic features. Data partitioning and evaluation: Dataset partitioned with a 7:1 patient-level split to create a hold-out test set, maintaining class distributions; remaining data used for stratified 3-fold cross-validation for training/validation, ensuring patient-level separation. Final split: training n=132, hold-out test n=21. Binary classification metrics reported included AUROC, AUPRC, precision, recall, and F1, with emphasis on AUPRC due to class imbalance. CONSORT flow indicated 906 assessed, 250 enrolled, 153 analyzed.

- Data collected from 153 analyzed patients provided 1353 hours of PPG data; low-risk states constituted approximately 22.1% of labeled periods. - Multi-modal spatio-temporal fusion transformer (STF) with multi-task learning (MTL) predicted low-risk clinical states over a 2-hour horizon with strong discrimination in the hold-out setting: AUROC 0.83 ± 0.03; AUPRC 0.67 ± 0.12; F1 0.78 ± 0.06; precision 0.69 ± 0.02; recall 0.95 ± 0.03. The STL variant performed worse (AUROC 0.71 ± 0.08; AUPRC 0.58 ± 0.06; F1 0.68 ± 0.08). - Compared with CNN and CNN-LSTM baselines on a related task, STF models showed superior performance (e.g., AUPRC: STF 0.15 ± 0.03 vs CNN-LSTM 0.10 ± 0.04 and CNN 0.08 ± 0.01; AUROC: STF 0.72 ± 0.03 vs 0.53 ± 0.12 and 0.53 ± 0.05), highlighting the advantage of transformer-based architectures and MTL. - Sensitivity analyses using clinical features alone (time-invariant models) exhibited poorer discrimination than models incorporating PPG data, underscoring the value of continuous waveform information. - Prediction of dengue shock syndrome (DSS) had relatively modest discrimination, attributed to class imbalance and potential label/ground-truth biases, as well as rapid physiological normalization following treatment.

The study demonstrates that integrating continuous PPG waveforms with basic clinical information and deep learning enables real-time risk stratification of hospitalized dengue patients. Using NEWS2 and mSOFA as proxies for illness severity, the optimized multi-task spatio-temporal transformer substantially outperformed single-task and CNN/CNN-LSTM baselines, particularly for predicting low-risk states over a 2-hour horizon. Multi-task learning leveraged shared structure between severity scoring and shock prediction, improving overall model quality. However, performance for predicting dengue shock syndrome was lower, likely due to class imbalance, clinical label variability, and potential treatment effects that normalize physiologic signals. The findings suggest that cost-effective wearable monitoring can support triage and prioritization in resource-limited settings, potentially facilitating safe ambulatory care for low-risk patients and focused resources on higher-risk individuals.

Continuous, non-invasive PPG monitoring combined with deep learning can predict short-term illness severity in dengue during early hospitalization, achieving strong discrimination for identifying low-risk periods. This approach could enable cost-effective triage and better clinical prioritization in settings with limited monitoring capacity. Future work should improve shock prediction through addressing class imbalance, enhancing ground-truth labeling, and integrating additional hemodynamic sensors. Hardware miniaturization, improved signal quality filtering, and extended battery life could increase adherence and data completeness. Prospective interventional studies are warranted to assess clinical impact and safety in ambulatory or step-down settings.

- Data quality and completeness: Only 43% of data windows were of sufficient quality, leading to exclusion of many recordings (97/250 enrolled not analyzed). Movement artifacts, reduced waveform amplitudes, and decreased peripheral perfusion degraded PPG signal quality. - Hardware constraints: Finite battery life (~18 h) interrupted continuous monitoring; relatively large form factor reduced patient comfort and adherence. - Model limitations: Class imbalance for shock periods hindered training; lower discrimination for dengue shock syndrome. Potential label/ground-truth bias due to clinical interventions and variable monitoring frequency in LMIC settings. - Implementation context: Uncertainty among clinical teams about added value of an additional wearable, especially in ICU settings with concurrent monitors, may affect adoption and data collection. - Generalizability: Single-center study in Southern Vietnam; pragmatic observational design may limit broader applicability without external validation.