AI Model Accuracy – Training the Blood Demand Prediction Model

🐞 The Bug/Hurdle

One major hurdle I faced while building the Blood Demand Prediction System was handling MongoDB time-series data efficiently for Prophet model training.

Issues Encountered:

• Inconsistent timestamps: Some records had missing or incorrect timestamps.
• Slow queries: Aggregation queries for time-series data were too slow, affecting model retraining.
• Data format issues: Prophet requires a ds (datetime) and y (value) column, but MongoDB stores dates in different formats.

🔧 How We Fixed It

1️⃣ Standardizing Timestamps

✅ Ensured all timestamps were stored in ISODate format in MongoDB.
✅ Used MongoDB’s

$dateFromString

in aggregation queries to fix incorrect formats.

2️⃣ Optimizing Queries

✅ Indexed the timestamp field to speed up filtering.
✅ Used MongoDB’s "$match" before "$group" to reduce the amount of scanned data.

3️⃣ Preprocessing Before Prophet Training

Transformed MongoDB documents into a clean DataFrame before passing them to Prophet:

import pandas as pd

# Fetch data from MongoDB
blood_data = list(mongo_collection.find({}, {"_id": 0, "timestamp": 1, "demand": 1}))

# Convert to DataFrame and rename columns
df = pd.DataFrame(blood_data)
df.rename(columns={"timestamp": "ds", "demand": "y"}, inplace=True)

# Ensure correct datetime format
df["ds"] = pd.to_datetime(df["ds"])