SepSentinel

A wearable multimodal platform for early sepsis detection using continuous biosensor data and machine learning.

What is SepSentinel?

SepSentinel is a wearable patch system that monitors physiological signals and measures interstitial fluid biomarkers in situ via microneedle-integrated electrochemical sensors. No fluid is extracted or transported — the sensors contact ISF directly within the skin. This repository contains the software prototype: data pipeline, ML models, and monitoring dashboard.

Hardware Concept

Microneedle Patch (inserted in skin)        Physiological Sensors
         |                                         |
Electrochemical sensors measure               HR, RR, Temp, SpO2
analytes in ISF (no fluid extraction)              |
         |                                         |
  pH     (potentiometric — TBD)                    |
  Lactate (amperometric — TBD)                     |
  IL-6   (E-AB, three-electrode, SWV)              |
         |                                         |
         +---------- Potentiostat -----------------+
                         |
                  Analog Front End / ADC
                         |
                   Microcontroller
                (signal processing + calibration)
                         |
                      Bluetooth
                         |
                       Phone
                         |
                   ML Inference
                         |
                 Sepsis Risk Score

The IL-6 sensor uses electrochemical aptamer-based (E-AB) sensing with a three-electrode system (WE/RE/CE) in direct ISF contact. Lactate and pH sensor hardware remains modular; likely electrochemical but final implementations may differ. The software pipeline receives processed biomarker values regardless of sensing modality.

Input Signals

7 features measured approximately every 5 minutes:

Signal	Unit	Normal Range	Source
Heart Rate	bpm	60-100	PPG sensor
Respiratory Rate	breaths/min	12-20	Impedance/accelerometer
Temperature	C	36.1-37.2	Thermistor
SpO2	%	95-100	Pulse oximetry
pH	pH units	7.35-7.45	ISF, potentiometric (TBD)
Lactate	mmol/L	0.5-2.0	ISF, amperometric (TBD)
IL-6	pg/mL	0-7	ISF, E-AB sensor (SWV)

Two-Model Architecture

The system uses two ML models in series:

Model A: Electrochemical Signal Calibration

Converts raw sensor outputs into biomarker concentrations. Each analyte has a distinct sensing mode:

• IL-6: SWV waveform -> concentration (E-AB sensor)
• Lactate: Amperometric current -> concentration (TBD)
• pH: Potentiometric voltage -> pH value (TBD)

Status: Awaiting experimental calibration data. Synthetic data will be generated for pipeline development.

Model B: Sepsis Risk Prediction

Predicts continuous sepsis probability from biomarker concentrations + physiological signals.

Model	Type	Input	Status
Random Forest	Flat baseline	Flattened feature vector	Implemented
XGBoost	Flat baseline	Flattened feature vector	Implemented
GRU	Sequential	(batch, timesteps, n_features)	Implemented
TCN	Sequential	(batch, timesteps, n_features)	Implemented
Transformer	Sequential	(batch, timesteps, n_features)	Implemented

Input feature count is dynamic (staged development):

• Stage 1: HR, SpO2, Temp, RR (4 features - PhysioNet Challenge)
• Stage 2: + Lactate, pH (6 features)
• Stage 3: + IL-6 (7 features - requires Model A)

The encoder is separated from the prediction head, enabling future dual-branch architectures.

Output

• Probability of sepsis: 0.0 to 1.0
• Risk score: 0-100%
• Risk category: Low / Medium / High

Project Structure

sepsentinel/
    config/
        signals.py            # Signal definitions, stages, column mappings
        thresholds.py         # Alert thresholds (WARNING / CRITICAL)
    data/
        synthetic.py          # Synthetic data generator (flat + episodes)
        physionet.py          # PhysioNet/CinC 2019 Challenge loader
        sequences.py          # Sliding window / tensor construction
        preprocessing.py      # Normalization, imputation
        mimic.py              # MIMIC-IV loader (Module 9)
    model_a/                  # Electrochemical signal -> concentration
        base.py               # CalibrationModel ABC
        synthetic_data.py     # Synthetic calibration data (future)
    model_b/                  # Sepsis risk prediction
        base.py               # SepsisModel ABC + SequenceEncoder ABC
        registry.py           # Model factory
        random_forest.py      # RF baseline
        xgboost_model.py      # XGBoost baseline
        gru.py                # GRU sequential model
        tcn.py                # TCN (causal dilated convolutions)
        transformer.py        # Transformer (causal self-attention)
        training.py           # Training loop (early stopping, checkpointing)
        evaluation.py         # Test-set metrics and comparison
    dashboard/
        app.py                # Streamlit web dashboard
        components.py         # Reusable UI components
    hardware/
        bluetooth.py          # BLE data reception (Module 10)
    alerts.py                 # Alert checking logic
    visualization.py          # Matplotlib plots
    simulation.py             # 7-signal patient simulation
train_stage1.py               # Train & compare all Stage 1 models
main.py                       # CLI entry point
app.py                        # Streamlit Cloud entry point
data/                         # Datasets
models/                       # Saved model artifacts
results/                      # Evaluation plots

How to Run

Requirements

• Python 3.10+
• pip install -r requirements.txt

Quick Start

pip install -r requirements.txt
python main.py

Menu Options

1. Simulate a worsening patient - 7-signal simulation with plots
2. Enter signal values manually - type values, get risk score
3. Train Random Forest - train on synthetic 7-feature data
4. Launch Dashboard - opens Streamlit dashboard
5. Exit

Running the Dashboard Directly

streamlit run sepsentinel/dashboard/app.py

Alert Thresholds

Signal	Warning	Critical
Heart Rate	>100 or <50 bpm	>120 or <40 bpm
Respiratory Rate	>22 or <10 br/min	>30 or <8 br/min
Temperature	>38.0 or <35.5 C	>39.0 or <35.0 C
SpO2	<94%	<90%
pH	<=7.35	<=7.25
Lactate	>=2.0 mmol/L	>=4.0 mmol/L
IL-6	>=7 pg/mL	>=50 pg/mL
Risk Score	>=30%	>=60%

Data Strategy

Phase 1: PhysioNet/CinC 2019 Sepsis Challenge dataset for Model B prototyping (Stage 1-2 features).

Phase 2: MIMIC-IV Clinical Database for larger-scale validation and custom cohort construction.

Phase 3: Integrate Model A outputs once experimental calibration data are available.

See DATASETS.md for full dataset strategy.

Roadmap

Completed

• v1.0-v1.1 - Proof of concept (3 biomarkers, RF, Streamlit dashboard)
• Module 5 - Architecture refactor (7 signals, model interfaces, new package structure)
• Module 6 - Time-series data pipeline (episode generator, sliding windows, preprocessing)
• Stage 1 Model B - 5 models (RF, XGBoost, GRU, TCN, Transformer) trained and compared on PhysioNet Challenge data. Transformer best at AUROC 0.793. See RESULTS.md.

Planned

• Module 8 - Dashboard v2 (live temporal plots, model selector, alert history)
• Module 9 - MIMIC-IV integration
• Module 10 - Hardware integration (Bluetooth, real-time inference, multi-patient)
• Model A - Electrochemical signal calibration (awaiting experimental data)

References

• SepAI: "SepAl: Sepsis Alerts on Low Power Wearables With Digital Biomarkers and On-Device Tiny Machine Learning" - temporal learning and feature fusion architecture
• PhysioNet/CinC 2019 Sepsis Challenge: Reyna et al.
• MIMIC-IV: Johnson et al., PhysioNet