Abstract:
Sepsis, defined by Sepsis-3 as life-threatening organ dysfunction caused by a dysregulated
host response to infection, imposes profound physiological and socio-economic burdens.
Globally, it is associated with roughly 50 million cases and 11 million deaths per year,
prolonged ICU stays, and substantial healthcare costs and productivity loss. Because
organ failure can evolve rapidly, timely detection is critical yet remains challenging in
practice due to label noise, missingness, and severe class imbalance in clinical data.
This work investigates lightweight machine-learning models for early sepsis detection
using seven non-invasive vital signs (heart rate, temperature, systolic/diastolic/mean
arterial pressure, oxygen saturation (SpO2), and respiratory rate). We took hourly data,
addressed missingness and imbalance with a simple preprocessing pipeline, and used
SHAP and LIME to provide model insight. We developed a low-cost point-of-care
prototype (Raspberry Pi and commodity sensors with a Python/Django backend) that
demonstrates feasibility for real-time prediction as well as telemedicine monitoring. On
the PhysioNet Challenge dataset, the system achieved a utility score of ≈ 46%, which is
an improvement of approximately 10% from the state-of-the art works. Our model also
achieved an accuracy of 86.49%, and an AUROC of 0.94. For external validation on the
MIMIC-IV demo cohort, our best model (XGBoost) attained 40% recall for the sepsis
class, substantially outperforming a partial qSOFA baseline at 6% recall. These results
indicate that simple vital-sign models can materially improve sensitivity over rule-based
screening while remaining deployable on affordable hardware; future work should pursue
validation on full-scale cohorts and prospective evaluation in real-world ICU workflows.