Transform Healthcare with Federated Data Lakes and AI

March 14, 2023
-
Public Sector Team
-

One of the most significant and difficult industries in the public sector is healthcare and healthcare management. Working with data and AI in this sector means handling, managing, and using the private and sensitive information of millions of people while at the same time developing new technologies and solutions. When it comes to data sharing and data-driven collaboration, which are crucial for advancing research and improving results, healthcare also encounters numerous challenges and restrictions.

The main data challenges in healthcare

One of the main challenges is data privacy. Healthcare data contains personal information that can reveal identities, diagnoses, treatments, and other confidential details. Sharing this data across different institutions or organizations can pose serious risks of data breaches, identity theft, discrimination, or misuse. Moreover, healthcare data is subject to strict regulations and ethical standards that limit its usage and distribution.

Another challenge is data availability. Healthcare data is often fragmented and siloed across different sources, such as hospitals, clinics, laboratories, pharmacies, or electronic health records (EHRs). This makes it difficult to access and integrate data from different locations and domains. Furthermore, healthcare data is often incomplete or inconsistent due to human errors or system failures.

These challenges hinder the potential of using artificial intelligence (AI) and machine learning (ML) in healthcare applications. AI and ML are powerful tools that can help analyze large amounts of data, discover patterns and insights, make predictions and recommendations, and automate tasks. However, AI and ML require access to sufficient and diverse data sets to train accurate and robust models that can generalize well to new situations.

Real World Federated Data Lake Examples

Federated data lakes are an emerging paradigm that aims to address these challenges by enabling collaborative learning without sharing raw data. A virtual data lakehouse allows multiple parties (e.g., hospitals) to jointly train a shared ML model by exchanging only model updates (e.g., gradients or parameters) instead of raw data. This way, a virtual data lakehouse preserves data privacy by keeping the data local at each party while still benefiting from the collective knowledge of all parties. Federated data access has many advantages for healthcare applications:

  • Improves the quality and diversity of data by aggregating information from different sources without compromising privacy or security.
  • Reduces the cost and complexity of data management by avoiding centralized storage or processing of large volumes of sensitive data.
  • Enhances the scalability and efficiency of learning by distributing computation across multiple devices or nodes instead of relying on a single server or cloud.
  • Empower innovation and collaboration by enabling cross-institutional or cross-domain learning without legal or ethical barriers.

Federated data-driven projects have already been applied [1] to various healthcare domains, such as medical imaging, remote health monitoring, genomics, and COVID-19 detection. Some examples are:

  • The ABIDE project used FL to train models on sensitive fMRI imaging data for identifying disease biomarkers.
  • The iPC [2] project used FL to train models on genomic data for personalized cancer treatment.
  • The COVID-Collab project [3] used FL to train models on smartphone sensor data for monitoring COVID-19 symptoms.

Challenges and how Blossom Sky helps to solve them

Federated data processing has its challenges. To overcome these challenges researchers and companies like DataBloom AI are developing novel techniques such as compression, aggregation, encryption, and automated data regulation. DataBloom’s flagship product, Blossom Sky, can solve or mitigate some of these challenges. Here are the most asked questions and our answers. 

Federated data processing requires frequent communication between parties to exchange model updates which can consume bandwidth resources especially when dealing with large models or datasets. 

That is true, and that’s why we developed in the first place our Virtual Data Lakehouse platform “Blossom Sky”. Blossom Sky organizes communication and minimizes the amount of transmitted data while ensuring that only approved data is used by participating parties. It features a comprehensive user interface that allows multiple parties to collaborate on the same project with changes being tracked and made transparent to the entire team. It can be thought of as the “Google Docs of AI”.

Data Federation involves heterogeneous parties that may have different types of devices (e.g., smartphones vs servers), datasets (e.g., size vs distribution), objectives (e.g., accuracy vs privacy), etc which can affect the convergence and performance of FL algorithms.

Blossom Sky uses Apache Wayang at this core. Apache Wayang is a cross-platform data processing system that aims to decouple the business logic of data analytics applications from concrete data processing platforms such as Apache Flink, Apache Spark, Tensorflow or any other data or AI framework. It is an API-first system designed to fully support cross-platform data processing and enables users to run data analytics over multiple data processing platforms, nodes or devices without changing the native code. This allows for greater flexibility and ease of use of different devices and datasets.

Federated data lakes are facing still security threats such as malicious parties that may tamper with model updates or infer private information from them using various attacks such as poisoning or inference.

This is true as for any AI / ML project, the outcome is only so good as the data behind are. There are several methods to defend against data poisoning attacks in federated data processing. One approach is to use an isolated forest algorithm to detect anomalies in the data. Another approach is to use a genetic algorithm during the participation stage of FL to find an optimal combination of data that avoids data poisoning attacks. DataBloom AI invests in researching mitigating approaches and develops prototypes with universities and early adopters which will be part of future releases of Blossom Sky.

A federated data lakehouse is a valid and emerging concept that transforms data-driven healthcare by enabling privacy-preserving collaborative data access and processing across multiple parties without sharing raw data. This way federated data can unlock new opportunities for innovation research and improvement in healthcare while respecting ethical legal and social values. 

[1]: The future of digital health with federated learning | npj Digital Medicine (nature.com)

[2]: iPC squares off against Paediatric Cancer | iPC Project EU

[3]: Overview ‹ Pandemic Response CoLab | MIT Media Lab

About Scalytics

The Scalytics Data Connect Platform is all about taking data collaboration and efficiency to the next level. Our platform tackles the big challenges of continuous data movement and timely data access to train AI effectively, bringing everything together in one easy-to-use system. It's built to work smoothly with a whole range of AI algorithms and models.

The cool part? Scalytics works hand-in-hand with top data frameworks like Databricks, Snowflake, Cloudera, and others, including Hadoop, Teradata, and Oracle. Plus, it's fully compatible with AI favorites like TensorFlow, Pandas, and PyTorch. We've made sure it fits right into your existing setup.

Scalytics is powered by Apache Wayang, and we're proud to support the project. You can check out their public GitHub repo right here. If you're enjoying our software, show your love and support - a star ⭐ would mean a lot!

If you need professional support from our team of industry leading experts, you can always reach out to us via Slack or Email.

Get started with Scalytics Connect today

Thank you! Our team will get in touch soon.
Oops! Something went wrong while submitting the form.