Clinical Decision Support System (CDSS) is an important technology for the healthcare industry that analyzes data to help healthcare professionals make decisions related to patient care. The market size for the global clinical decision support system appears poised for expansion, with one study predicting a compound annual growth rate (CAGR) of 10.4%, from 2022 to 2030, to $10.7 billion.

For any health organization that wants to build a CDSS system, one key block is to locate and extract the medical entities that are present in the clinical notes, medical journals, discharge summaries, etc. Along with entity extraction, the other key components of the CDSS system are capturing the temporal relationships, subjects, and certainty assessments.

At Google Cloud, we know how critical it is for the healthcare industry to build CDSS systems, so we worked with Apollo 24|7, the largest multi-channel digital healthcare platform in India, to build the key blocks of their CDSS solution.

We helped them to parse the discharge summaries and prescriptions to extract the medical entities. These entities can then be used to build a recommendation engine that would help doctors with the “Next Best Action” recommendation for medicines, lab tests, etc.

Let’s take a sneak peek at Apollo 24|7’s entity extraction solutions, and the various Google AI technologies that were tested to form the technology stack.

Datasets Used

To perform our experiments on entity extraction, we used two types of datasets.

1. i2b2 Dataset – i2b2 is an open-source clinical data warehousing and analytics research platform that provides annotated deidentified patient discharge summaries made available to the community for research purposes. This dataset was primarily used for training and validation of the models.
2. Apollo 24|7’s Dataset – De-identified doctor’s notes from Apollo24|7 were used for testing. Doctors annotated them to label the entities and offset values.

Experimentation and choosing the right approach — Four models put to test

For entity extraction, both Google Cloud products and open-source approaches were explored. Below are the details:

1. Healthcare Natural Language API: This is a no-code approach that provides machine learning solutions for deriving insights from medical text. Using this, we parsed unstructured medical text and then generated a structured data representation of the medical knowledge entities stored in the data for downstream analysis and automation. The process includes:

• Extract information about medical concepts like diseases, medications, medical devices, procedures, and their clinically relevant attributes;
• Map medical concepts to standard medical vocabularies such as RxNorm, ICD-10, MeSH, and SNOMED CT (US users only);
• Derive medical insights from text and integrate them with data analytics products in Google Cloud.

The advantage of using this approach is that it not only extracts a wide range of entity types like MED_DOSE, MED_DURATION, LAB_UNIT, LAB_VALUE, etc, but also captures functional features such as temporal relationships, subjects, and certainty assessments, along with the confidence scores. Since it is available on Google Cloud, this offers long-term product support. It is also the only fully-managed NLP service among all the approaches tested and hence, it requires the least effort to implement and manage.

But one thing to keep in mind is that since the Healthcare NL API offers natural language models that are pre-trained, it currently cannot be used for custom entity extraction models trained using custom annotated medical text or to extract custom entities. This has to be done via AutoML Entity Extraction for Healthcare, another Google Cloud service for custom model development. Custom model development is important for adapting the pre-trained models to new languages or region-specific natural language processing, such as medical terms whose use may be more prevalent in India than in other regions

2. Vertex AutoML Entity Extraction for Healthcare: This is a low-code approach that’s already available on Google Cloud. We used AutoML Entity Extraction to build and deploy custom machine learning models that analyzed documents, categorized them, and identified entities within them. This custom machine learning model was trained on the annotated dataset provided by the Apollo 24|7 team.

The advantage of AutoML Entity Extraction is that it gives the option to train on a new dataset. However, one of the prerequisites to keep in mind is that it needs a little pre-processing to capture the input data in the required JSONL format. Since this is an AutoML model just for Entity Extraction, it does not extract relationships, certainty assessments, etc.

3. BERT-based Models on Vertex AI: Vertex AI is Google Cloud’s fully managed unified AI platform to build, deploy, and scale ML models faster, with pre-trained and custom tooling. We experimented with multiple custom approaches based on pre-trained BERT-based models, which have shown state-of-the-art performance in many natural language tasks. To gain better contextual understanding of medical terms and procedures, these BERT-based approaches are explicitly trained on medical domain data. Our experiments were based on BioClinical BERT, BioLink BERT, Blue BERT trained on Pubmed dataset, and Blue BERT trained on Pubmed + MIMIC datasets.

The major advantage of these BERT-based models is that they can be finetuned on any Entity Recognition task with minimal efforts.

However, since this is a custom approach, it requires some technical expertise. Additionally, it does not extract relationships, certainty assessments, etc. This is one of the main limitations of using BERT-based models.

4. ScispaCy on Vertex AI: We used Vertex AI to perform experiments based on ScispaCy, which is a Python package containing spaCy models for processing biomedical, scientific or clinical text.

Along with Entity Extraction, Scispacy on Vertex AI provides additional components like Abbreviation Detector, Entity Linking, etc. However, when compared to other models, it was less precise, with too many junk phrases, like “Admission Date,” captured as entities.

“Exploring multiple approaches and understanding the pros/cons of each approach helped us to decide the one that would fit our business requirements.” according to Abdussamad M, Engineering Lead at Apollo 24|7.

Evaluation Strategy

In order to match the parsed entity with the test data labels, we used extensive matching logic that comprised of the below four methods:

1. Exact Match – Exact match captures entities where the model output and the entities in the test dataset match. Here, the offset values of the entities have also been considered. For example, the entity “gastrointestinal infection” that is present as-is in both the model output and the test label will be considered an “Exact Match.”
2. Match-Score Logic – We used a scoring logic for matching the entities. For each word in the test data labels, every word in the model output is matched along with the offset. A score is calculated between the entities and based on the threshold, it is considered as a match.
3. Partial Match – In this matching logic, entities like “hypertension” and “hypertensive” are matched based on the Fuzzy logic.
4. UMLS Abbreviation Lookup – We also observed that the medical text had some abbreviations, like AP meaning abdominal pain. These were first expanded by doing a lookup on the respective UMLS (Unified Medical Language System) tables and then passed to the individual entity extraction models.

Performance Metrics

We used precision and recall metrics to compare the outcomes of different models/experiments.

Precision (also called positive predictive value) is the fraction of relevant instances among the retrieved instances, while recall (also known as sensitivity) is the fraction of relevant instances that were retrieved.

The below example shows how to calculate these metrics for a given sample.

Example sample: “Krish has fever, headache and feels uncomfortable”

Expected Entities: [“fever”, “headache”]

Model Output: [“fever”, “feels”, “uncomfortable”]

Thus,

Experimentation Results
The following table captures the results of the above experiments on Apollo24|7’s internal datasets.

Finally, the Blue BERT model trained on the Pubmed dataset had the best performance metrics with a 81% improvement on Apollo 24|7’s baseline mode with the Healthcare Natural Language API providing the context, relationships, and codes. This performance could be further improved by implementing an ensemble of these two models.

“With the Blue BERT model giving the best performance for entity extraction on Vertex AI and the Healthcare NL API being able to extract the relationships, certainty assessments etc, we finally decided to go with an ensemble of these 2 approaches,“ Abdussamad added.

Fast track end-to-end deployment with Google Cloud AI Services (AIS)

Google AIS (Professional Services Organization) helped Apollo24|7 to build the key blocks of the CDSS system.

The partnership between Google Cloud and Apollo 24|7 is just one of the latest examples of how we’re providing AI-powered solutions to solve complex problems to help organizations drive the desired outcomes. To learn more about Google Cloud’s AI services, visit our AI & ML Products page, and to learn more about Google Cloud solutions for health care, explore our Google Cloud Healthcare Data Engine page.

Acknowledgements

We’d like to give special thanks to Nitin Aggarwal, Gopala Dhar and Kartik Chaudhary for their support and guidance throughout the project. We are also thankful to Manisha Yadav, Santosh Gadgei and Vasantha Kumar for implementing the GCP infrastructure. We are grateful to the Apollo team (Chaitanya Bharadwaj, Abdussamad GM, Lavish M, Dinesh Singamsetty, Anmol Singh and Prithwiraj) and our partner team from HCL/Wipro (Durga Tulluru and Praful Turanur) who partnered with us in delivering this successful project. Special thanks to the Cloud Healthcare NLP API team (Donny Cheung, Amirhossein Simjour, and Kalyan Pamarthy).

Documents like payslips and W2s are crucial to processes such as employment and income verification for mortgage loans, personal loans, personal finance, and benefits processing. Unfortunately, efficiently extracting data from these documents at scale can be challenging and time-consuming, with many organizations relying on manual examination of documents or automated approaches that don’t adequately capture the document data needed for given tasks. Google Cloud built Document AI to remove these barriers, empowering customers to deploy powerful machine learning models to more quickly process documents, save money, and discover insights. We’re excited to expand Document AI’s capabilities with the recent release of improved pre-trained models for W2s and payslips, built on Document AI Workbench.

Pre-trained models let developers focus on core application logic and leave the complex task of information extraction from the documents to Google’s AI technology. In many cases, the primary driver for automated data extraction is operational efficiency and cost savings, but Document AI can also open new possibilities. For example, a financial services company might use Document AI to enable fully self-serve loan applications on mobile devices, helping the organization to differentiate itself with simple, fast customer experiences.

We’ve heard from customers that more granular entity extraction from W2 and payslip documents is particularly important, with organizations requiring support for a wider variety of layouts and formats. The recent launch of the stable release of these pretrained models addresses these requests.

Here is what is new with W2 parser:

• The parser improves accuracy and entity specificity thanks to the ability to break down long entities such as addresses into fine-grained sub-entities like StreetAddressOrPostalBox, AdditionalStreetAddressOrPostalBox, City, State, and ZIP code.
• It can handle a wider variation of W2 forms, including multi-copies (2,3,4-ups) issued by various payroll vendors. The model is not limited to specific tax years, which means it should be able to process W2 for 2022 or beyond provided there are not significant changes to the format.
• It introduces eight new entities for Box 12 that represent both codes and values, enriching understanding of the various taxable and non-taxable components of the W2 recipient’s income.

Here is what is new with Payslip parser:

• Bonus, commissions, holiday, overtime, regular pay, and vacation are now part of earning_item/earning_this_period and earning_item/earning_ytd. The parser captures types of earnings beyond those categories, and maps them to their respective earning rates, hours, and pay (both for the period and year-to-date). This helps in building a more detailed understanding of the components of the payslip recipient’s income
• The parser now returns year-to-date and current-period taxes and deductions.
• Direct deposits are linked to corresponding bank account numbers.
• The parser now returns page numbers, state and federal tax exemptions, and filing statuses.

While these parsers have become more useful out of the box, with this release, the ability to uptrain makes them easy to modify as new needs arise. Uptraining lets developers further improve the accuracy of these models and extract additional fields with minimal development work. It also lets developers customize existing parsers to support new document types that are similar. For example, the parser is trained on U.S. data and could be uptrained to create a payslip parser for the U.K.

We’re pleased that parsers are already making a difference for customers. Bryan Jackson, CTO at lending automation firm Gateless, said, “High accuracy data extraction is critical to the success of our Smart Underwrite solution, and Document AI provided better results than competitors. Using the latest W2 & Payslip pretrained parsers, we saw a 48% increase in performance on pay stubs and a 15% performance improvement in W2s. The ability to easily uptrain models as new document variations are introduced ensures we continue to deliver optimal outcomes for our customers.”

Additional pre-trained models available as release candidates include parsers for 1040, 1099R, 1120, and 1120S documents. Check for details here. To learn more, talk to a Google Cloud sales executive about how Document AI can help your business, and check out our Document AI breakout session from Google Cloud Next ’22.

Presently, every government agency has to take a hard look at their data capabilities and decide whether their current infrastructure supports their workflow. For many, it doesn’t. Most data systems are developed with a strict set of parameters in mind before implementation, which can limit flexibility and long-term use. Particularly during a crisis, flexible “living systems” offer tremendous advantages as they’re able to change capacity rapidly. Building living data systems with the cloud in mind allows organizations to respond to a changing world with confidence.

Last summer, the Government Business Council conducted a survey of government employees to understand the impacts of data efficiency on government operations. The report Built to Last: A Survey on Organizational Data Efficiency in Times of Crisis offers key insights into organizational efficacy and whether organizations can adapt to a crisis at speed. It also highlights differences between traditional data systems and living data systems. 

Data needs to be readily available

When the pandemic first hit, many agencies needed to create or transition their systems to allow employees to work remotely. This change tested the limits of existing data systems. Even after finding a cloud service provider, agencies encountered the challenges of migrating their data to the cloud. 

Government organizations had decades of data stored in paper records. Most have been working to transfer these records to a digital format, but the process has been slow. They are also faced with collecting sizable amounts of data in real time from their ongoing services,  which involves interfacing with the public, external vendors, or third-party institutions. 

Building the cloud into a flexible data system can solve both issues. Old records can be digitized and given an easy-to-access home for those who need them. Incoming data, both internal and external, can be made accessible as well. Migrating data to the cloud also doubles as a way to create backups of raw data, adding an extra layer of security. Most importantly, building in the cloud unlocked the capacity to scale when demand rises. 

Data should be updated in real-time

One of the key takeaways from the Government Business Council report is the fact that agencies are better able to adapt at speed when data efficiencies are higher. 74% of organizations with pandemic related functions reported a moderate to severe impact to their jobs at the onset of the pandemic. Of those organizations, the ones reporting their data efficiency as “very good” have largely already recovered. That adaptability directly affects an agency’s ability to make informed decisions during a time of a crisis.

Having a real-time data solution in place lets agencies make near real-time decisions. A great example of this from early in the pandemic is vaccine distribution. Google Cloud supported multiple states, such as the State of Wyoming, in distributing vaccines efficiently while handling challenges such as reaching rural populations. Data systems that gathered real-time patient data made a difference in the number of vaccines distributed. Knowing population data and patient risk factors enabled quick and effective decision-making.

A global pandemic is far from the only crisis that needs effective data analytics. Natural disasters, food deserts, public health issues, and more can all be handled more efficiently by having real-time data at hand. Effective data analytics systems are the digital equal of “having your ear to the ground” in each community. They provide valuable insights into what people  need.

Data needs to be accessible and easy to use

Making data easy to work with and understand sets phenomenal data systems apart from functional ones. Having data in the cloud is a great first step, but agencies need to be able to easily access and quickly use the data to accomplish their goals. This is where traditional data systems fail most often. Traditional IT systems and data strategies are designed for a specific purpose, usually identified before development and implementation begin. That means that when the data living in those systems needs to be used differently, adapting to new requirements can be difficult. 

Data can often feel “locked” in traditional systems; the data is there, but there’s no way to get to it or work with it in a way that meets the needs of a crisis. Flexible data systems address this by allowing for greater accessibility. Google Cloud, for example, has customizable tools, such as Contact Center AI and Document AI, which let agencies work with data in ever-changing ways. This also produces greater data transparency since data sets can be worked with and accessed more easily.

Governments need to respond to the changing needs of their constituents in emergencies. While traditional data systems can handle slowly shifting demands on the system, they do not serve agencies well in a crisis. When urgency, accuracy, and accessibility all matter, flexible systems rise to the challenge. The pandemic has pushed agencies to adapt in real time, and many have realized they need a system that adapts with them.

Google Cloud has a suite of tools to create integrated data ecosystems. These ecosystems can scale with increasing demand, meet dynamic development needs, and adapt to a changing landscape. Data-first decision-making is a core tenet of “living data systems.” Google Cloud data systems have handled everything from administering vaccines to detecting fraud. In each of these applications, a core tenet of data-first decision making was implemented at scale. 

For more insights on how flexible data systems help the public sector, download the full report “Built to Last: A Survey on Organizational Data Efficiency in Times of Crisis.”

There’s an explosion in the number of digital images generated and used for both personal and business needs. On e-commerce platforms and online marketplaces for example, product images and visuals heavily influence the consumer’s perception, decision making and ultimately conversion rates. In addition, there’s been a rapid shift towards user-generated visual content for ecommerce — think seller-generated product imagery, host-generated rental property photos and influencer-generated social media content. The challenge? These user-generated images are often captured using mobile cameras and vary greatly in terms of their size, quality, compression ratios and resolution, making it difficult for companies to provide consistent high-quality product images on their platforms.

That’s exactly the problem that Let’s Enhance, a computer vision startup with teams across the US and Ukraine, set out to solve with AI. The Let’s Enhance platform improves the quality of any user-generated photo automatically using AI-based features to enhance images through automatic upscaling, pixelation and blur fixes, color and low-light correction, and removing compression artifacts — all with a single click and no professional equipment or photo-editing chops.

“Let’s Enhance.io is designed to be a simple platform that brings AI-powered visual technologies to everyone — from marketers and entrepreneurs to photographers and designers,” said Sofi Shvets, CEO and Co-founder of Let’s Enhance.

To date, Let’s Enhance has processed more than 100M photos for millions of customers worldwide for use-cases ranging from digital art galleries, real estate agencies, digital printing, ecommerce and online marketplaces. With the introduction of Claid.ai, their new API to automatically enhance and optimize user-generated content at scale for digital marketplaces, they needed to process millions of images every month and manage sudden peaks in user demand.

However, building and deploying an AI-enabled service at scale for global use is a huge technical challenge that spans model building, training, inference serving and resource scaling. It demands an infrastructure that’s easy to manage and monitor, can deliver real-time performance to end customers wherever they are and can scale as user-demand peaks, all while optimizing costs.

To support their growing user-base, Let’s Enhance chose to deploy their AI-powered platform to production on Google Cloud and NVIDIA. But before diving into their solution, let’s take a close look at the technical challenges they faced in meeting their business goals.

Architecting a solution to fuel the next wave of growth and innovation

To deliver the high-quality enhanced images that end-customers see, Let’s Enhance products are powered by cutting-edge deep neural networks (DNNs) that are both compute- and memory-intensive. While building and training these DNN models in itself is a complex, iterative process, application performance — when processing new user-requests, for instance — is crucial to delivering a quality end user experience and reducing total deployment costs.

A single inference or processing request i.e., from user-generated image, which can vary widely in size, at the input to the AI-enhanced output image, requires combining multiple DNN models within an end-to-end pipeline. The key inference performance metrics to optimize for included latency (the time it takes from providing an input image to the enhanced image being available) and throughput ( the number of images that can be processed per second)..

Together, Google Cloud and NVIDIA technologies provided all the elements that the Let’s Enhance team needed to set themselves up for growth and scale. There were three main components in the solution stack:

• Compute resources: A2 VMs, powered by NVIDIA A100 Tensor Core GPUs
• Infrastructure management: Google Kubernetes Engine (GKE)
• Inference serving: NVIDIA Triton Inference Server

Improved throughput and lower costs with NVIDIA A100 Multi-Instance GPUs (MIG) on Google Cloud

To meet the computational requirements of the DNN models and deliver real-time inference performance to their end-users, Let’s Enhance chose Google Cloud A2 VMs powered by NVIDIA A100 Tensor Core GPUs as their compute infrastructure. A100 GPU’s Multi-Instance GPU (MIG) capability offered them the unique ability to partition a single GPU into two independent instances and simultaneously process two user-requests at a time, making it possible to service a higher volume of user requests while reducing the total costs of deployment.

The A100 MIG instances delivered a 40% average throughput improvement compared to the NVIDIA V100 GPUs, with an increase of up to 80% for certain image enhancement pipelines using the same number of nodes for deployment. With improved performance from the same sized node pools, Let’s Enhance observed a 34% cost savings using MIG-enabled A100 GPUs.

Simplified infrastructure management with GKE

To offer a guaranteed quality-of-service (QoS) to their customers and manage user demand, Let’s Enhance needed to provision, manage and scale underlying compute resources while keeping utilization high and costs low.

GKE offers industry-leading capabilities for training and inference such as support for 15,000 nodes per cluster, auto-provisioning, auto-scaling and various machine types (e.g. CPU, GPU and on-demand, spot), making it the perfect choice for Let’s Enhance.

With support for NVIDIA GPUs and NVIDIA GPU sharing capabilities, GKE can provision multiple A100 MIG instances to process user requests in parallel and maximize utilization. As the compute required for the deployed ML pipelines increases (e.g., a sudden surge in inference requests to service), GKE can automatically scale to additional node-pools with MIG partitions — offering finer granularity to provision the right-sized GPU acceleration for workloads of all sizes.

“Our total averaged throughput varied between 10 and 80 images / sec. Thanks to support for NVIDIA A100 MIG and auto scaling mechanisms in GKE we can now scale that up to 150 images/sec and more, based on user-demand and GPU availability,” said Vlad Pranskevičius, Co-founder and CTO, Let’s Enhance.

In addition, GKE’s support for dynamic scheduling, automated maintenance and upgrades, high availability, job API, customizability and fault tolerance simplified managing a production deployment environment, allowing the Let’s Enhance team to focus on building advanced ML pipelines.

High-performance inference serving with NVIDIA Triton Inference Server

To optimize performance and simplify the deployment of their DNN models onto the GKE-managed node pools of NVIDIA A100 MIG instances, Let’s Enhance chose the open-source NVIDIA Triton Inference Server, which deploys, runs and scales AI models from any framework onto any GPU- or CPU-based infrastructure.

NVIDIA Triton’s support for multiple frameworks enabled the Let’s Enhance team to serve models trained in both TensorFlow and PyTorch, eliminating the need to set up and maintain multiple serving solutions for different framework backends. In addition, Triton’s ensemble model and shared memory features helped maximize performance and minimize data transfer overhead for their end-to-end image processing pipeline, which consists of multiple models and large amounts of raw image data transferring between them.

“We saw over 20% performance improvement with Triton vs. custom inference serving code, and were also able to fix several errors during deployment due to Triton’s self-healing features”, said Vlad. “Triton is packed with cool features, and as I was watching its progress from the very beginning, the pace of product development is just amazing.”

To further enhance end-to-end inference performance, the team is adopting NVIDIA TensorRT, an SDK to optimize trained models for deployment with the highest throughput and lowest latency while preserving the accuracy of predictions.

“With the subset of our models which we converted from Tensorflow to NVIDIA TensorRT we observed a speedup of anywhere between 10% and 42%, depending on the model, which is impressive,” said Vlad. “For memory-intensive models like ours, we were also able to achieve predictable memory consumption, which is another great benefit of using TensorRT, helping prevent any unexpected memory-related issues during production deployment.”

Teamwork makes the dream work

By using Google Cloud and NVIDIA, Let’s Enhance addressed the real-time inference serving and infrastructure management challenges of taking their AI-enabled service to production at scale in a secure Google Cloud infrastructure.

GKE, NVIDIA AI software and A2 VMs powered by NVIDIA A100 GPUs brought together all the elements they needed to deliver the desired user experience and build a solution that can dynamically scale their end-to-end pipelines based on user demand.

The close collaboration with the Google Cloud and NVIDIA team, every step of the way, also helped Let’s Enhance get their solution to market fast. “The Google Cloud and NVIDIA teams are incredible to work with. They are highly professional, always responsive and genuinely care about our success,” said Vlad. “We were able to get technical advice and recommendations directly from Google Cloud product and engineering teams, and also have a channel to share our feedback about Google Cloud products, which was really important to us.”

With a solid foundation and solution architecture that can meet their growing business needs, Let’s Enhance is marching towards their goal of bringing AI-enhanced digital images to everyone. “Our vision is to help businesses effectively manage user-generated content and increase conversion rates with next-gen AI tools,” said Sofi. ”Our next step towards that is to expand our offerings to completely replace manual work for photo preparation, including as well as cover pre-stages as image quality assessment and moderation.”

To learn more about the Let’s Enhance products and services, check out their blog here.