Save Time with AI-Powered Alert Management: Categorization and Prioritization [Case Study]

Project Background

Empeek’s team developed a new product for a client that operates in the vendor management software market.

The primary idea behind the Product is to use Artificial Intelligence (AI) to recognize and classify important data from emails, texts, and other sources.

The goal is to minimize the time employees spend reviewing various alerts and notifications, allowing them to focus solely on critical matters.

Timeline: 6+ months

Team Size: 3-4

Save Time with AI-Powered Alert Management: Categorization and Prioritization [Case Study] 1

Tech Stack

Languages

Python, C#, React

Neural Network Model Types

PyTorch, ONNX

Framework

Transformers

Infrastructure

Google Colab, Microsoft Azure ML Studio

Start you AI-enabled Project today. Get free consultation.

Challenges and Requirements

Initially, there was no clear vision of how the system should work. The team quickly ruled out the usage of third-party services and Large Language Models (LLMs) due to performance concerns, cost considerations, and the potential for handling personal client information.

The project required a solution that could:

Accurately identify and extract relevant tokens (entities, keywords, etc.) from text data.
Classify alerts and notifications into appropriate categories.
Summarize lengthy texts for easy comprehension.
Ensure privacy and security by processing data locally without sending it to external services.
Provide high performance and cost-effectiveness.

Save Time with AI-Powered Alert Management: Categorization and Prioritization [Case Study] 2

Save Time with AI-Powered Alert Management: Categorization and Prioritization [Case Study] 3

Chosen Approach

After evaluating various options, we decided to use specific pre-trained models from the HuggingFace repository and implement the solution using Python and HuggingFace library.

To optimize model size and performance, the team converts the models to the ONNX format, allowing them to be used directly in .NET projects (with some limitations due to the lack of robust libraries).

Named Entity Recognition (NER)

NER models from HuggingFace are used to identify tokens (entities, keywords, etc.) in text data. These models are fine-tuned on pre-processed client data.

Classification

Sentence-transformer models from the Sentence-Transformers (SBERT) library are employed for alert classification. The SetFit technique is used for efficient model fine-tuning, even with small training datasets.

Summarization

A pre-trained summarization model is also incorporated, which can be fine-tuned if needed.

Categorization

Embedding vectors are used to measure semantic similarity, allowing for the best match between the given text and category/subcategory descriptions.

Training & Development Process

The training process is currently conducted on Google Colab, using a custom-built pipeline that fine-tunes the chosen HuggingFace base models with client data in a specific format. The proper models are then uploaded to Google Drive or Azure Storage.

The data preparation process involves:

Manual analysis and annotation of client data by support personnel and developers.
Thorough data validation to ensure consistency and accuracy.
Categorization of alerts into different types.

Empeek’s team initially had focused on achieving 80% accuracy for classifying the top 8 alert types and plans to continuously expand and improve the system.

Save Time with AI-Powered Alert Management: Categorization and Prioritization [Case Study] 4

Challenges and Lessons Learned

One of the biggest challenges was obtaining a sufficient amount of high-quality training data. Data preparation, including manual annotation and validation, was a time-consuming process.

Data Quality

To ensure data quality and consistency is crucial for model performance.

Personnel Involvement

Involving subject matter experts (support personnel, developers) in the data annotation process is essential.

All time improvements

Continuous iteration and improvement are necessary as new alert types and patterns emerge.

Future Plans

The team plans to expand the number of clients using the X version of the Product, offering automated alert classification, prioritization, and ticket creation to minimize human intervention.
The long-term goal is to release the beta version of the Product, incorporating feedback from alpha users to enhance functionality.
Ongoing models refinement and customization options will be available to cater to specific client needs and preferences for alert handling.

Save Time with AI-Powered Alert Management: Categorization and Prioritization [Case Study] 5

About Empeek

Engineering a better healthcare future.

Empeek is a custom healthcare software development company that helps healthtech startups and medical facilities create and leverage innovative, HIPAA/HITECH compliant technology solutions such as EMR and telemedicine systems, patient-centered crossplatform apps, AI-powered tools, IoT ecosystems, and others.

150+ Specialists
On the Market Since 2015
HIPAA & GDPR Compliant

Save Time with AI-Powered Alert Management: Categorization and Prioritization [Case Study] 6

(No Ratings Yet)

Written by:

Alex Shpachuk CEO

Alex Shpachuk is the owner and strategic partner of Empeek. His effective leadership and a visionary approach to the future of healthcare turned the company into a dynamic environment attracting the brightest minds with the common vision for product impact and service excellence. With over a decade of experience in software engineering and comprehensive knowledge of designing and deploying tailor-made solutions for healthcare providers, Alex channels his passion for software development and consulting into the written word.