GlucoPulse

GlucoPulse.ai

A personal health analytics project to analyze continuous glucose monitor (CGM) data using Snowflake, Snowflake Cortex AI, and Snowpark for Python.

1. Project Goal

As a person managing Type 2 Diabetes, I built this project to move beyond simple averages like HbA1c and analyze the real-world impact of **glycemic variability**.

The core hypothesis is that large, rapid glucose excursions are disproportionately damaging, and that a 3-month average (HbA1c) can be dangerously misleading. This project uses data analysis and machine learning to quantify, classify, and detect these high-variability events from over 96,000 personal CGM readings.

2. Technical Stack

• Data Warehouse: Snowflake
• Data Pipeline: Snowflake SQL (ELT Pattern)
• Core Table: RAW_READINGS (contains all Type 0, 1, and 6 records)
• NLP/AI: Snowflake Cortex AI (COMPLETE function)
• ML Modeling: Snowpark for Python (with scikit-learn)
• Source Control: Git / GitHub

3. Data Engineering Pipeline (ELT)

The source data from the LibreView export is not in a clean, queryable format. A multi-step ELT (Extract-Load-Transform) pipeline was built to clean and model the data for analysis.

Source Data Profile

The raw `.csv` export contains several data quality issues:

• Junk Header: A non-data title row is present on row 1.
• Misaligned Headers: The true column headers are on row 2.
• Conditional Data: Glucose readings are stored in two different columns (Historic Glucose mg/dL vs. Scan Glucose mg/dL) based on a Record Type (0 or 1).
• Sparse Data: Non-reading events (Record Type 6), which contain all lifestyle notes, are mixed with valid glucose readings.
• Fragmented Data: Lifestyle notes are spread across 10+ separate columns.

Loading Strategy (ELT)

A robust ELT pattern was used to ensure data integrity and auditability.

1. Extract/Load: The raw CSV is first uploaded to a Snowflake Stage (@RAW_FILE_STAGE).