AI-Powered Healthcare with MongoDB & Microsoft

Data Architecture and Flow

In this solution, MongoDB Atlas serves as the operational datastore for real-time AI applications, while Microsoft OneLake handles analytics for long-term trend analysis. This architecture enables the following features:

• Real-time processing of patient data and medical imaging.

• Integration between operational and analytical systems.

• Efficient data flow from transactional to analytical processing.

• Support for both millisecond-response operational queries and complex analytical workloads.

Figure 2: Real-time to analytics data pipeline

Predictive AI for Early Detection

Predictive AI can be used in healthcare to generate an accurate diagnosis from large datasets. Microsoft Fabric Data Science presents a robust platform to train and experiment with ML Models and manage MLOps cycles. This solution uses models for the following purposes:

1. BI-RADS prediction

   BI-RADS is an industry-standard mechanism to analyze mammogram findings. Healthcare specialists use BI-RADS to describe results from breast-imaging tests with a number from 0 to 6, with the possibility of malignancy increasing with the score.

   This solution uses the VGG16 deep convolutional neural network (CNN) to predict BI-RADS scores from images. The model is trained on mammogram images from a Kaggle dataset. Each images is grouped into folders that correspond to their BI-RADS.

   Fabric Data Science analyzes the performance of multiple models for this task and selects the best one. It trains the models, runs experiments, and manages multiple versions. The training images are directly uploaded to the Lakehouse in OneLake from the user's local machine by using the Lakehouse UI. Additionally, you can easily reference the images stored in Azure Blob Storage by using the wget or curl commands, using shortcuts, or by using a data pipeline. The solution stores the image metadata and final predictions in MongoDB Atlas.

2. Biopsy classification

   Classification or regression models can be used to classify tumors as malignant or benign. The random forest classifier model is trained on a Kaggle dataset that includes input parameters such as clump thickness, uniformity of cell size and shape, bare nuclei, and mitoses. The model can then predict whether a tumor is malignant or benign. In production, you can add more parameters to the dataset and train the model on these values to make more accurate predictions. During solution development, the random forest model had an accuracy rate of over 97%. The solution fetches the training dataset from MongoDB Atlas, and the prediction output is updated in MongoDB by using the MongoDB Spark Connector.

Fabric Data Science makes training and managing your models easy by automatically logging related parameters for each model and experiment.

Vector Search Implementation

This solution's intelligent query system relies on Vector Search, as shown in the following diagram.

Figure 3. Vector Search implementation process flow

1. Data preparation:

   • Azure OpenAI's text-embedding-ada-002 model processes clinical notes.

   • Data is converted into vector embeddings for high-dimensional space representation.

   • Vector embeddings are stored in MongoDB Atlas with optimized search indexes.

2. Query processing:

   • Natural language queries are converted into vector representations.

   • Semantic understanding enables complex medical queries.

   • Query vectors are matched against stored embeddings.

3. Document retrieval:

   • Returns relevant medical records based on semantic matching.

   • Enables intuitive access to patient information.

   • Atlas Vector Search executes similarity-based searches.

RAG-based Chatbot Architecture

The chatbot implementation leverages RAG architecture in the following contexts, as shown in the following diagram:

Figure 4. Blueprint for the chatbot architecture

1. Patient information retrieval:

   • Executes queries to fetch current patient details.

   • Retrieves structured patient data from MongoDB collections.

   • Provides immediate access to critical patient information.

2. Historical data processing:

   • Accesses 10-year patient history from MongoDB Atlas.

   • Decodes and summarizes historical data through Azure OpenAI LLM.

   • Implements thought chaining for context-aware responses.

3. Medical knowledge integration:

   • Uses vectorized medical documentation.

   • Performs real-time vector searches based on the query's context.

   • Integrates relevant medical literature and case studies.

Analytics and Visualization

This solution uses the following two visualization platforms for analytics.

First, MongoDB Atlas Charts provides native, real-time operational dashboards directly connected to MongoDB data. It enables immediate insights into critical healthcare metrics through intuitive visualizations without requiring data transformations or additional tools. The operational dashboard (Figure 5) demonstrates key metrics including patient numbers, appointment status, and clinic distribution.

Figure 5. Operational dashboard with Atlas Charts

Second, Power BI integration extends the analytics capabilities by enabling enterprise-wide data analysis and advanced visualizations. Through the MongoDB Atlas Connector, healthcare data can be combined with other enterprise sources in Microsoft OneLake. The geographical visualization dashboard (Figure 6) showcases this integration, displaying patient distribution and enabling sophisticated analytical capabilities.

Figure 6. PowerBI integration with MongoDB Atlas

Together, these platforms provide a complete analytics solution that handles both immediate operational needs and long-term analytical requirements.

This solution demonstrates how MongoDB Atlas can handle operational data, Vector Search capabilities, and analytics requirements while seamlessly integrating with Microsoft's AI and visualization tools. This architecture enables healthcare providers to leverage real-time operational insights and long-term analytical capabilities within a single system.