Biosensor Population Analysis & Microbial Correlation Pipeline

Project Scope: Longitudinal Validation of 45-Subject Prototype Sensor Cohort

1. Project Overview

In this project, I developed an end-to-end data science pipeline to validate the performance of a prototype wearable biosensor. My goal was to move beyond basic accuracy reporting and investigate how a user’s unique biological profile—specifically their BMI, inflammatory activity, and gut microbiome—impacts sensor reliability.

2. Methodology & Technical Implementation

A. Data Integrity & Ingestion

I processed raw time-series data for a cohort of 49 potential subjects. During my initial audit, I identified that four subjects (24, 25, 37, and 40) had incomplete data records or missing clinical metadata. To ensure the integrity of my statistical analysis, I excluded these individuals, resulting in a robust final cohort of 45 subjects.

B. Signal Conditioning

To handle the high-frequency "jitter" found in the prototype hardware, I implemented:

• Gaussian Smoothing: I applied a Gaussian filter () to the raw sensor data to extract the true physiological signal.
• Temporal Synchronization: I aligned the prototype data with the gold-standard reference (Dexcom G6).

C. Advanced Engineering Metrics

I implemented four specialized metrics to quantify the sensor's performance:

• MARD (Mean Absolute Relative Difference): The primary metric for clinical accuracy.
• SNR (Signal-to-Noise Ratio): Quantifying signal quality against residual hardware noise.
• Cross-Correlation Lag: Determining the exact response time of the prototype.
• Hysteresis Analysis: Evaluating error bias during rising vs. falling glucose trends.

3. Clinical & Microbial Discovery

The core value of this pipeline is the integration of diverse datasets, including physical markers (bio.csv) and microbiome profiles (microbes.csv).

The "Microbe Driver" Model

I built a Random Forest Regressor to identify correlations between the gut microbiome and sensor error.

By treating the counts of hundreds of bacterial species as features, I was able to rank which microbes "predict" a less accurate sensor.

Key Findings:

• Microbial Predictors: I identified that Streptococcus salivarius and Bacteroides thetaiotaomicron were among the top 10 features correlating with sensor drift.
• Physiological Impact: My analysis revealed a clear correlation between BMI and MARD, suggesting that subcutaneous tissue thickness is a significant covariate for this device.

4. Final Cohort Results

• Mean Population MARD: 25.49%
• Mean SNR: 49.14 dB
• Cohort Size: 45 Valid Subjects

5. Pipeline Structure

• src/preprocessing.py: Signal smoothing and time-alignment.
• src/metrics.py: Implementation of MARD, SNR, Lag, and Hysteresis.
• src/clinical_analysis.py: "Left-Join" merging of sensor results with BMI and inflammatory markers.
• src/microbe_analysis.py: Machine Learning driver analysis using Scikit-learn.
• main.py: One-click execution of the entire end-to-end pipeline.

---

Final Project Status: Complete

I have generated the following artifacts in the results/ directory:

• batch_summary_report.csv: Individual-level engineering KPIs.
• final_integrated_clinical_report.csv: Integrated master dataset for further analysis.