TL;DR
This paper introduces an interpretable machine learning pipeline using wearable sensors to predict sepsis mortality in low-resource settings. By leveraging continuous heart rate variability (HRV) data from wearables, the system outperforms traditional bedside monitoring for mortality prediction, offering a promising solution for improving sepsis care in underserved regions.
What Problem Does This Paper Address?
Sepsis is a critical global health issue with disproportionately high mortality rates in low- and middle-income countries (LMICs). The challenge in these regions is the scarcity of healthcare professionals and the high cost of essential monitoring equipment, hindering timely diagnosis and effective management. While machine learning (ML) shows promise for sepsis prediction, most studies focus on high-income country (HIC) settings and rely on traditional, expensive bedside monitors. This paper aims to bridge this gap by exploring the feasibility of using cost-effective wearable sensors and developing ML models for sepsis mortality prediction specifically tailored for LMIC environments.
Methods
The study employs a two-pronged approach: first, assessing the feasibility of wearable sensors for sepsis monitoring, and second, developing automated ML prediction models.
- Data Collection: Continuous electrocardiogram (ECG) data was collected for up to 24 hours using a wearable ePatch® sensor from 50 sepsis patients admitted to the Hospital for Tropical Diseases in Vietnam. Simultaneously, standard vital signs (temperature, pulse, blood pressure, oxygen saturation) were recorded using bedside monitors.
- Feature Extraction: From the ECG data, Heart Rate Variability (HRV) features were extracted, encompassing time-domain, frequency-domain, and non-linear parameters. Vital signs from bedside monitors were also collected.
- Machine Learning Models: The researchers trained and evaluated several state-of-the-art ML models, including Support Vector Machines with Recursive Feature Elimination (SVM-RFE), Gaussian Process Classification (GPC), and Gradient Boosting Decision Trees (XGBoost). For time-varying data, Long Short-Term Memory (LSTM) networks were utilized.
- Interpretability: A key aspect of the study was the focus on model interpretability. Techniques like feature importance analysis for GBDT and LIME (Local Interpretable Model-agnostic Explanations) for LSTMs were used to understand which features and at what time points were most influential in predicting mortality.
Datasets
The study utilized a newly collected dataset of 50 adult sepsis patients admitted to the Hospital for Tropical Diseases (HTD) in Ho Chi Minh City, Vietnam. After exclusions due to missing outcome data, the final dataset comprised 40 patients. This dataset represents a novel benchmark for wearable sensor data in sepsis prediction within an LMIC context.
Key Insights and Ideas
- Wearables Outperform Bedside Monitors: The most striking insight is that ML models trained on data from wearable sensors consistently outperformed those trained on traditional bedside monitor data for sepsis mortality prediction. This highlights the potential of accessible technology to improve care.
- HRV is a Powerful Predictor: Heart Rate Variability (HRV) features, particularly their time-varying dynamics, emerged as highly informative for predicting sepsis mortality. This suggests that subtle physiological changes captured by HRV are more sensitive indicators of deterioration than basic vital signs alone.
- Interpretability is Crucial for Adoption: The emphasis on interpretable ML models is vital. By understanding why a model makes a prediction, clinicians can build trust and integrate these tools into their decision-making process, especially in resource-limited settings where expert judgment is paramount.
- LMIC Context Matters: The study's focus on LMICs is critical. The findings demonstrate that solutions developed for HICs may not be directly applicable and that tailored approaches using affordable technology are necessary to address healthcare disparities.
Key Results
- The highest performance for mortality prediction was achieved using time-varying HRV features with LSTM models, resulting in an Area Under the Precision-Recall Curve (AUCPRC) of 0.83.
- When using static features, the SVM-RFE model with a combination of HRV and vital signs (SepHRVital) achieved the highest AUCPRC of 0.74 in the LOSO cross-validation scheme.
- Across various models and feature sets, wearable data generally led to better prediction metrics (e.g., PPV, F1-score, AUCROC, AUCPRC) compared to bedside monitor data.
- Interpretability analysis consistently identified HRV features, particularly HR(mean), as highly contributing factors for mortality prediction across different ML models.
Limitations and Future Work
The authors acknowledge several limitations:
- Small Sample Size: The dataset of 40 patients is relatively small, which can limit the generalizability of the findings.
- Lack of External Validation: The models were not externally validated on data from other clinical settings or countries.
- Imbalanced Dataset: The dataset had a higher proportion of survivors than non-survivors, which can affect the performance of certain metrics. The authors mitigated this by using metrics like AUCPRC, which are less sensitive to class imbalance.
Future work proposed by the authors includes:
- Clinical Deployment: Implementing the developed prediction system in clinical practice at the Hospital for Tropical Diseases for more efficient patient monitoring.
- Larger Datasets and Transfer Learning: Leveraging larger datasets from high-income settings and employing transfer learning techniques to improve the models' performance and generalizability.
- Real-time Monitoring: Developing real-time monitoring capabilities to provide immediate alerts to clinicians.
Summary updated