Abstract
Sepsis is a critical medical condition where early intervention is paramount, yet current predictive models often rely on invasive laboratory data and high-power computing. This paper introduces SepAI, a TinyML-based system designed for low-power wearables that predicts sepsis onset using only six non-invasive vital signs. By employing a quantized Multi-head Temporal Convolutional Network (TCN) on an ARM Cortex-M33 microcontroller, SepAI achieves a median lead time of 9.8 hours with minimal energy consumption (2.68 mJ per inference), enabling continuous, proactive monitoring outside of intensive care settings.
The Challenge of Early Sepsis Detection
Sepsis claims approximately 11 million lives annually. Because every hour of delay in administering antibiotics increases mortality risk by roughly 0.42%, early detection is vital. However, most existing deep learning solutions for sepsis prediction require "high-fidelity" laboratory data (e.g., blood lactate levels) and run on power-hungry GPUs. This infrastructure is rarely available for continuous monitoring in general hospital wards or home environments.
SepAI addresses this gap by utilizing "digital biomarkers"—vitals like heart rate and temperature that can be captured via wearable sensors—and processing them entirely on-device to ensure privacy and low latency.
Methods: Multi-head TCN and TinyML Optimization
The core of SepAI is a Multi-head Temporal Convolutional Network (TCN). Unlike standard architectures that process all variables simultaneously, SepAI assigns a separate TCN "head" to each of the six input vital signs. This design allows the model to extract individual temporal features from noisy sensor streams before fusing them in a dense layer for final classification.
Key Architectural Features:
- Temporal Logic: The model employs dilated causal convolutions, enabling it to analyze a four-hour window of historical data without "leakage" from future time steps.
- Hardware-Software Co-design: To fit the constraints of an ARM Cortex-M33 (nRF5340), the model was fully quantized from 32-bit floating-point to 8-bit integers (INT8). This optimization achieved a 4x reduction in memory footprint with negligible loss in accuracy.
- Consensus Algorithm: To mitigate "alarm fatigue," the system triggers an alert only after $K$ consecutive positive classifications, filtering out transient noise.
Datasets and Training
The researchers utilized the HiRID dataset from Bern University Hospital. Unlike the commonly used MIMIC dataset, HiRID provides high-granularity data sampled every 2 minutes, which more accurately reflects the continuous data stream produced by wearable sensors. The model focuses on six features: Heart Rate (HR), Respiratory Rate (RR), Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP), Oxygen Saturation (SpO2), and Body Temperature.
Key Results
SepAI demonstrates that sophisticated clinical monitoring can be moved from the cloud to the edge:
- Early Warning: The system identifies sepsis a median of 9.8 hours before clinical onset.
- On-Device Efficiency: Each inference takes only 143 ms and consumes 2.68 mJ of energy.
- Memory Footprint: The entire model requires only 103 KB for weights and 23 KB for activations, making it ideal for resource-constrained microcontrollers.
Conclusion and Future Work
SepAI proves that digital biomarkers are sufficient for meaningful sepsis prediction when paired with optimized TinyML architectures. While the current specificity (0.48) reflects a trade-off for high lead time, the system provides a significant window for medical intervention. Future research will focus on validating SepAI in "the wild" using physical wearable prototypes equipped with PPG and IMU sensors to move toward real-world clinical trials.
Summary updated