Operationalizing Healthcare Models in Production with Innovaccer’s ML Platform

Shubham Gupta & Sparsh Dutta
Mon 25 Sep 2023
Share this blog:

The advancements and research in big data and AI have led healthcare to increasingly adopt machine learning and predictive algorithms in their workflows. This field offers numerous opportunities for data scientists to explore ideas, conduct experiments, and develop new knowledge and models.

However, it is well-known that only a small fraction, approximately 10%, of these models actually make it to production where they can impact real-life business problems and processes. This lopsided ratio is difficult to change due to regulations, guardrails, biases, and a lack of mature testing frameworks. Nevertheless, data science teams face a new set of challenges when dealing with the models that do reach production. These challenges include:

  1. Acquiring, cleaning, and performing quality checks on large amounts of data: "Why is my model execution failing or producing strange output?"
  2. Tracking and versioning for experiments and model training: "I think the model built last quarter was better. Can I retrospectively run the current data with the previous version of the model?"
  3. Setting up standardized deployment and monitoring pipelines for production-ready models: "How can I ensure the models run reliably in production?" Also, "I have new data scientists joining the team. How quickly can they familiarize themselves with the models currently in production?"

In the past, when software engineering went through maturity cycles to address scalability, reliability, and maintainability, DevOps emerged as a framework of solutions. DevOps brought a set of practices for developing, testing, deploying, and operating large-scale software systems. It resulted in shorter development cycles, increased deployment velocity, and auditable, dependable system releases.

Likewise, MLOps acts as a framework of best practices and solutions at the intersection of DevOps, data engineering, and machine learning. While it shares similarities with DevOps in concept, MLOps differs in execution. ML systems are inherently experimental and involve more complex components to build and operate.

How Do We Handle MLOps at Innovaccer?

Integrating machine learning into administrative applications and clinical workflows in the healthcare industry presents unique challenges, including data privacy, data drift, regulatory compliance, and inherent bias. At Innovaccer, we recognize these challenges and have customized our MLOps approach to address each of them. Our goal is to ensure ease of model deployment, model interpretability, bias mitigation, and seamless integration in multi-client environments. We have developed an end-to-end ML platform that standardizes the key aspects of managing machine learning/deep learning models during the production phase.

Our platform's primary objective is to enable data scientists to concentrate on model development, without worrying about allocating infra resources, deploying and monitoring models, scheduling jobs, or engaging in redundant engineering activities.

Fig. 1: Our MLOps platform at Innovaccer performs over 80% of the tasks during the post-model build phase, freeing data scientists to optimize their workflows.

To simplify the MLOps process, we introduced a deployment template that data scientists can use to populate their model training and inference codes, as well as standardize their pipelines. This template reduces the burden of MLOps, simplifies replication in multiple client environments, and significantly reduces deployment time and cost.

Fig. 2: The process flow illustrates Innovaccer's MLOps workflow and task segregation between data scientists and MLOps engineering.

Innovaccer's Machine Learning Platform: Tools and Frameworks

  1. Data Quality Pre-checker with Great Expectations: We use Great Expectations to meticulously verify data quality before it enters our ETL pipeline. This ensures that only reliable and consistent data flows through the system, setting the stage for accurate model training.
  2. Feature Store Powered by FeatureVault: Innovaccer's Feature Store serves as the centralized hub for storing, sharing, and managing machine learning features. This promotes seamless collaboration and maintains feature consistency across diverse projects. It is powered by our proprietary in-house tool, FeatureVault, which enhances the Databricks Feature Store. FeatureVault acts as an intermediate engine for building healthcare features, with feature management handled by the Databricks Feature Store.
  3. Standardized ML Template Powered by Kedro: Our standardized ML template powered by Kedro provides a battle-tested blueprint for deploying machine learning models into production. With streamlined workflows, this template ensures consistent and reliable model deployment, accelerating time to market.
  4. Efficient Model Management and Versioning with MLflow: Our Model Management component, MLflow, ensures organized and efficient tracking of machine learning models and provides version control capabilities.
  5. Real-time Model Monitoring and Alerts with Nannyml: The Nannyml monitoring framework enables us to keep a watchful eye on deployed models for any model drifts, providing real-time model monitoring and alerts.
  6. Automated Pipeline Orchestration using Airflow: We leverage Airflow to seamlessly automate the flow of data and processes. This fusion of components enhances efficiency, ensures repeatability, and reduces manual intervention in our ML workflows.
  7. Consistent CI/CD Using Gitlab: Our standardized CI/CD process ensures uniform, continuous integration and deployment for models and analytics. This reliability simplifies the deployment and management of models across the board.
  8. Robust Infrastructure Management with Databricks: Databricks handles the heavy lifting of infrastructure management, providing a robust and scalable environment for running complex ML workflows. Its unified analytics platform accelerates data processing and supports our diverse computational needs.
  9. Comprehensive Model Insights with Evidently.ai: Evidently.ai enriches our platform by providing in-depth insights into model performance and data quality through visualization and analysis. This integration empowers data-driven decisions and fosters continuous model improvement.
  10. Real-time Notifications via Slack: Slack ensures that stakeholders are promptly informed about pipeline statuses, model performance, and critical events. This real-time communication streamlines decision-making and enables rapid responses.

Innovaccer’s ML Platform: Accelerating Innovation, Enhancing Collaboration, and Improving Patient Outcomes

  1. Accelerated Model Deployment and Iteration

Our ML Platform expedites model deployment and iteration by integrating ML with DevOps practices. Streamlined deployment pipelines, automated updates, and containerization techniques ensure that insights from advanced models transition swiftly into real-world healthcare scenarios. By streamlining the process, we significantly reduce the time-to-production from three weeks to just three days.

Fig. 3: The Innovaccer ML Platform reduces engineering overhead for model deployment by 80%.

  1. Enhanced Collaboration Between Data Scientists and Operations Team

Our ML platform fosters collaboration between data scientists and engineering teams, combining DevOps principles and ML deployment best practices. This collaboration aligns innovative model designs with operational realities, resulting in feasible and efficient solutions for the healthcare ecosystem.

  1. Improved Patient Outcomes Through Data-Driven Decisions

The Innovaccer ML Platform directly impacts patient outcomes by enabling data-driven decisions. Using MLOps, we extract actionable insights from vast healthcare datasets, providing evidence-based information to medical professionals. This empowers them to make informed decisions, enhance patient care, diagnose illnesses, predict health trends, and personalize treatment plans. Our solutions optimize interventions and care programs, leading to improved patient outcomes.

Innovaccer's Commitment to AI-Enabled Healthcare Innovation

Innovaccer is a leader in data-driven healthcare and understands that MLOps is an ongoing journey that must adapt to the evolving technological needs of the healthcare industry. Our ML Platform, which includes MLOps, exemplifies our commitment to pushing the boundaries of AI-enabled healthcare innovation while maintaining the highest standards of data privacy and clinical efficacy.

We have also launched our AI suite, Sara, which offers self-conversational tools to improve patient outcomes, strengthen financial performance, and engage providers and consumers. If you want to accelerate AI adoption in your HIT systems, schedule a demo with our experts to start your AI-led healthcare journey.


See the industry’s most advanced decision support system

Get an overview of the most advanced Health Cloud built to help healthcare care as one.
Tags: Healthcare Models
Shubham Gupta & Sparsh Dutta

Accelerate Your Digital Transformation with the Innovaccer Health Cloud

Request a free demo of the #1 healthcare data platform to know how you can generate millions in savings just like our superhero customers.

errorhi there

errorhi there

errorhi there