Philips accelerates growth of AI-enabled healthcare options with an MLOps platform constructed on Amazon SageMaker

It is a joint weblog with AWS and Philips.

Philips is a well being know-how firm targeted on bettering folks’s lives via significant innovation. Since 2014, the corporate has been providing clients its Philips HealthSuite Platform, which orchestrates dozens of AWS providers that healthcare and life sciences firms use to enhance affected person care. It companions with healthcare suppliers, startups, universities, and different firms to develop know-how that helps docs make extra exact diagnoses and ship extra customized therapy for tens of millions of individuals worldwide.

One of many key drivers of Philips’ innovation technique is synthetic intelligence (AI), which allows the creation of good and customized services that may enhance well being outcomes, improve buyer expertise, and optimize operational effectivity.

Amazon SageMaker offers purpose-built instruments for machine studying operations (MLOps) to assist automate and standardize processes throughout the ML lifecycle. With SageMaker MLOps instruments, groups can simply practice, check, troubleshoot, deploy, and govern ML fashions at scale to spice up productiveness of knowledge scientists and ML engineers whereas sustaining mannequin efficiency in manufacturing.

On this publish, we describe how Philips partnered with AWS to develop AI ToolSuite—a scalable, safe, and compliant ML platform on SageMaker. This platform offers capabilities starting from experimentation, information annotation, coaching, mannequin deployments, and reusable templates. All these capabilities are constructed to assist a number of traces of enterprise innovate with velocity and agility whereas governing at scale with central controls. We define the important thing use circumstances that supplied necessities for the primary iteration of the platform, the core parts, and the outcomes achieved. We conclude by figuring out the continuing efforts to allow the platform with generative AI workloads and quickly onboard new customers and groups to undertake the platform.

Buyer context

Philips makes use of AI in varied domains, comparable to imaging, diagnostics, remedy, private well being, and related care. Some examples of AI-enabled options that Philips has developed over the previous years are:

• Philips SmartSpeed – An AI-based imaging know-how for MRI that makes use of a novel Compressed-SENSE based mostly deep studying AI algorithm to take velocity and picture high quality to the following degree for a big number of sufferers
• Philips eCareManager – A telehealth resolution that makes use of AI to help the distant care and administration of critically ailing sufferers in intensive care models, by utilizing superior analytics and scientific algorithms to course of the affected person information from a number of sources, and offering actionable insights, alerts, and suggestions for the care staff
• Philips Sonicare – A wise toothbrush that makes use of AI to research the brushing conduct and oral well being of customers, and supply real-time steerage and customized suggestions, comparable to optimum brushing time, strain, and protection, to enhance their dental hygiene and stop cavities and gum ailments.

For a few years, Philips has been pioneering the event of data-driven algorithms to gasoline its modern options throughout the healthcare continuum. Within the diagnostic imaging area, Philips developed a mess of ML functions for medical picture reconstruction and interpretation, workflow administration, and therapy optimization. Additionally in affected person monitoring, picture guided remedy, ultrasound and private well being groups have been creating ML algorithms and functions. Nevertheless, innovation was hampered as a result of utilizing fragmented AI growth environments throughout groups. These environments ranged from particular person laptops and desktops to various on-premises computational clusters and cloud-based infrastructure. This heterogeneity initially enabled totally different groups to maneuver quick of their early AI growth efforts, however is now holding again alternatives to scale and enhance effectivity of our AI growth processes.

It was evident {that a} elementary shift in the direction of a unified and standardized setting was crucial to really unleash the potential of data-driven endeavors at Philips.

Key AI/ML use circumstances and platform necessities

AI/ML-enabled propositions can remodel healthcare by automating administrative duties completed by clinicians. For instance:

• AI can analyze medical photographs to assist radiologists diagnose ailments sooner and extra precisely
• AI can predict future medical occasions by analyzing affected person information and bettering proactive care
• AI can advocate customized therapy tailor-made to sufferers’ wants
• AI can extract and construction data from scientific notes to make record-taking extra environment friendly
• AI interfaces can present affected person help for queries, reminders, and symptom checkers

Total, AI/ML guarantees diminished human error, time and price financial savings, optimized affected person experiences, and well timed, customized interventions.

One of many key necessities for the ML growth and deployment platform was the flexibility of the platform to help the continual iterative growth and deployment course of, as proven within the following determine.

The AI asset growth begins in a lab setting, the place the information is collected and curated, after which the fashions are skilled and validated. When the mannequin is prepared and authorised to be used, it’s deployed into the real-world manufacturing programs. As soon as deployed, mannequin efficiency is constantly monitored. The actual-world efficiency and suggestions are finally used for additional mannequin enhancements with full automation of the mannequin coaching and deployment.

The extra detailed AI ToolSuite necessities had been pushed by three instance use circumstances:

• Develop a pc imaginative and prescient utility geared toward object detection on the edge. The information science staff anticipated an AI-based automated picture annotation workflow to hurry up a time-consuming labeling course of.
• Allow an information science staff to handle a household of basic ML fashions for benchmarking statistics throughout a number of medical models. The mission required automation of mannequin deployment, experiment monitoring, mannequin monitoring, and extra management over all the course of finish to finish each for auditing and retraining sooner or later.
• Enhance the standard and time to marketplace for deep studying fashions in diagnostic medical imaging. The present computing infrastructure didn’t enable for working many experiments in parallel, which delayed mannequin growth. Additionally, for regulatory functions, it’s essential to allow full reproducibility of mannequin coaching for a number of years.

Non-functional necessities

Constructing a scalable and strong AI/ML platform requires cautious consideration of non-functional necessities. These necessities transcend the particular functionalities of the platform and give attention to guaranteeing the next:

• Scalability – The AI ToolSuite platform should be capable of scale Philips’s insights technology infrastructure extra successfully in order that the platform can deal with a rising quantity of knowledge, customers, and AI/ML workloads with out sacrificing efficiency. It must be designed to scale horizontally and vertically to fulfill rising calls for seamlessly whereas offering central useful resource administration.
• Efficiency – The platform should ship high-performance computing capabilities to effectively course of complicated AI/ML algorithms. SageMaker provides a variety of occasion varieties, together with cases with highly effective GPUs, which may considerably speed up mannequin coaching and inference duties. It additionally ought to reduce latency and response instances to supply real-time or near-real-time outcomes.
• Reliability – The platform should present a extremely dependable and strong AI infrastructure that spans throughout a number of Availability Zones. This multi-AZ structure ought to guarantee uninterrupted AI operations by distributing assets and workloads throughout distinct information facilities.
• Availability – The platform have to be obtainable 24/7, with minimal downtime for upkeep and upgrades. AI ToolSuite’s excessive availability ought to embody load balancing, fault-tolerant architectures, and proactive monitoring.
• Safety and Governance – The platform should make use of strong safety measures, encryption, entry controls, devoted roles, and authentication mechanisms with steady monitoring for uncommon actions and conducting safety audits.
• Knowledge Administration – Environment friendly information administration is essential for AI/ML platforms. Laws within the healthcare business name for particularly rigorous information governance. It ought to embody options like information versioning, information lineage, information governance, and information high quality assurance to make sure correct and dependable outcomes.
• Interoperability – The platform must be designed to combine simply with Philips’s inner information repositories, permitting seamless information trade and collaboration with third-party functions.
• Maintainability – The platform’s structure and code base must be effectively organized, modular, and maintainable. This permits Philips ML engineers and builders to supply updates, bug fixes, and future enhancements with out disrupting all the system.
• Useful resource optimization – The platform ought to monitor utilization reviews very carefully to verify computing assets are used effectively and allocate assets dynamically based mostly on demand. As well as, Philips ought to use AWS Billing and Value Administration instruments to verify groups obtain notifications when utilization passes the allotted threshold quantity.
• Monitoring and logging – The platform ought to use Amazon CloudWatch alerts for complete monitoring and logging capabilities, that are vital to trace system efficiency, determine bottlenecks, and troubleshoot points successfully.
• Compliance – The platform also can assist enhance regulatory compliance of AI-enabled propositions. Reproducibility and traceability have to be enabled mechanically by the end-to-end information processing pipelines, the place many obligatory documentation artifacts, comparable to information lineage reviews and mannequin playing cards, might be ready mechanically.
• Testing and validation – Rigorous testing and validation procedures have to be in place to make sure the accuracy and reliability of AI/ML fashions and stop unintended biases.

Answer overview

AI ToolSuite is an end-to-end, scalable, fast begin AI growth setting providing native SageMaker and related AI/ML providers with Philips HealthSuite safety and privateness guardrails and Philips ecosystem integrations. There are three personas with devoted units of entry permissions:

• Knowledge scientist – Put together information, and develop and practice fashions in a collaborative workspace
• ML engineer – Productionize ML functions with mannequin deployment, monitoring, and upkeep
• Knowledge science admin – Create a mission per staff request to supply devoted remoted environments with use case-specific templates

The platform growth spanned a number of launch cycles in an iterative cycle of uncover, design, construct, check, and deploy. As a result of uniqueness of some functions, the extension of the platform required embedding present customized parts like information shops or proprietary instruments for annotation.
The next determine illustrates the three-layer structure of AI ToolSuite, together with the bottom infrastructure as the primary layer, frequent ML parts because the second layer, and project-specific templates because the third layer.

Layer 1 incorporates the bottom infrastructure:

• A networking layer with parametrized entry to the web with excessive availability
• Self-service provisioning with infrastructure as code (IaC)
• An built-in growth setting (IDE) utilizing an Amazon SageMaker Studio area
• Platform roles (information science admin, information scientist)
• Artifacts storage
• Logging and monitoring for observability

Layer 2 incorporates frequent ML parts:

• Automated experiment monitoring for each job and pipeline
• A mannequin construct pipeline to launch a brand new mannequin construct replace
• A mannequin coaching pipeline comprised of mannequin coaching, analysis, registration
• A mannequin deploy pipeline to deploy the mannequin for closing testing and approval
• A mannequin registry to simply handle mannequin variations
• A mission position created particularly for a given use case, to be assigned to SageMaker Studio customers
• A picture repository for storing processing, coaching, and inference container photographs constructed for the mission
• A code repository to retailer code artifacts
• A mission Amazon Simple Storage Service (Amazon S3) bucket to retailer all mission information and artifacts

Layer 3 incorporates project-specific templates that may be created with customized parts as required by new initiatives. For instance:

• Template 1 – Features a element for information querying and historical past monitoring
• Template 2 – Features a element for information annotations with a customized annotation workflow to make use of proprietary annotation tooling
• Template 3 – Consists of parts for customized container photographs to customise each their growth setting and coaching routines, devoted HPC file system, and entry from a neighborhood IDE for customers

The next diagram highlights the important thing AWS providers spanning a number of AWS accounts for growth, staging, and manufacturing.

Within the following sections, we focus on the important thing capabilities of the platform enabled by AWS providers, together with SageMaker, AWS Service Catalog, CloudWatch, AWS Lambda, Amazon Elastic Container Registry (Amazon ECR), Amazon S3, AWS Identity and Access Management (IAM), and others.

Infrastructure as code

The platform makes use of IaC, which permits Philips to automate the provisioning and administration of infrastructure assets. This method can even assist reproducibility, scalability, model management, consistency, safety, and portability for growth, testing, or manufacturing.

Entry to AWS environments

SageMaker and related AI/ML providers are accessed with safety guardrails for information preparation, mannequin growth, coaching, annotation, and deployment.

Isolation and collaboration

The platform ensures information isolation by storing and processing individually, decreasing the chance of unauthorized entry or information breaches.

The platform facilitates staff collaboration, which is crucial in AI initiatives that sometimes contain cross-functional groups, together with information scientists, information science admins, and MLOps engineers.

Position-based entry management

Position-based entry management (RBAC) is crucial in managing permissions and simplifying entry administration by defining roles and permissions in a structured method. It makes it easy to handle permissions as groups and initiatives develop and entry management for various personas concerned in AWS AI/ML initiatives, comparable to the information science admin, information scientist, annotation admin, annotator, and MLOps engineer.

Entry to information shops

The platform permits SageMaker entry to information shops, which ensures that information might be effectively utilized for mannequin coaching and inference with out the necessity to duplicate or transfer information throughout totally different storage places, thereby optimizing useful resource utilization and decreasing prices.

Annotation utilizing Philips-specific annotation instruments

AWS provides a collection of AI and ML providers, comparable to SageMaker, Amazon SageMaker Ground Truth, and Amazon Cognito, that are absolutely built-in with Philips-specific in-house annotation instruments. This integration allows builders to coach and deploy ML fashions utilizing the annotated information inside the AWS setting.

ML templates

The AI ToolSuite platform provides templates in AWS for varied ML workflows. These templates are preconfigured infrastructure setups tailor-made to particular ML use circumstances and are accessible via providers like SageMaker project templates, AWS CloudFormation, and Service Catalog.

Integration with Philips GitHub

Integration with GitHub enhances effectivity by offering a centralized platform for model management, code evaluations, and automatic CI/CD (steady integration and steady deployment) pipelines, decreasing handbook duties and boosting productiveness.

Visible Studio Code integration

Integration with Visible Studio Code offers a unified setting for coding, debugging, and managing ML initiatives. This streamlines all the ML workflow, decreasing context switching and saving time. The combination additionally enhances collaboration amongst staff members by enabling them to work on SageMaker initiatives collectively inside a well-known growth setting, using model management programs, and sharing code and notebooks seamlessly.

Mannequin and information lineage and traceability for reproducibility and compliance

The platform offers versioning, which helps maintain observe of adjustments to the information scientist’s coaching and inference information over time, making it simpler to breed outcomes and perceive the evolution of the datasets.

The platform additionally allows SageMaker experiment monitoring, which permits end-users to log and observe all of the metadata related to their ML experiments, together with hyperparameters, enter information, code, and mannequin artifacts. These capabilities are important for demonstrating compliance with regulatory requirements and guaranteeing transparency and accountability in AI/ML workflows.

AI/ML specification report technology for regulatory compliance

AWS maintains compliance certifications for varied business requirements and rules. AI/ML specification reviews function important compliance documentation, showcasing adherence to regulatory necessities. These reviews doc the versioning of datasets, fashions, and code. Model management is crucial for sustaining information lineage, traceability, and reproducibility, all of that are vital for regulatory compliance and auditing.

Undertaking-level funds administration

Undertaking-level funds administration permits the group to set limits on spending, serving to to keep away from sudden prices and guaranteeing that the ML initiatives keep inside funds. With funds administration, the group can allocate particular budgets to particular person initiatives or groups, which helps groups determine useful resource inefficiencies or sudden price spikes early on. Along with funds administration, with the function to mechanically shut down idle notebooks, staff members keep away from paying for unused assets, additionally releasing worthwhile assets when they aren’t actively in use, making them obtainable for different duties or customers.

Outcomes

AI ToolSuite was designed and applied as an enterprise-wide platform for ML growth and deployment for information scientists throughout Philips. Various necessities from all enterprise models had been collected and thought of through the design and growth. Early within the mission, Philips recognized champions from the enterprise groups who supplied suggestions and helped consider the worth of the platform.

The next outcomes had been achieved:

• Consumer adoption is likely one of the key main indicators for Philips. Customers from a number of enterprise models had been skilled and onboarded to the platform, and that quantity is anticipated to develop in 2024.
• One other necessary metric is the effectivity for information science customers. With AI ToolSuite, new ML growth environments are deployed in lower than an hour as an alternative of a number of days.
• Knowledge science groups can entry a scalable, safe, cost-efficient, cloud-based compute infrastructure.
• Groups can run a number of mannequin coaching experiments in parallel, which considerably diminished the typical coaching time from weeks to 1–3 days.
• As a result of the setting deployment is absolutely automated, it requires just about no involvement of the cloud infrastructure engineers, which diminished operational prices.
• The usage of AI ToolSuite considerably enhanced the general maturity of knowledge and AI deliverables by selling using good ML practices, standardized workflows, and end-to-end reproducibility, which is vital for regulatory compliance within the healthcare business.

Wanting ahead with generative AI

As organizations race to undertake the following state-of-the-art in AI, it’s crucial to undertake new know-how within the context of the group’s safety and governance coverage. The structure of AI ToolSuite offers a wonderful blueprint for enabling entry to generative AI capabilities in AWS for various groups at Philips. Groups can use basis fashions made obtainable with Amazon SageMaker JumpStart, which offers an unlimited variety of open supply fashions from Hugging Face and different suppliers. With the required guardrails already in place when it comes to entry management, mission provisioning, and price controls, it is going to be seamless for groups to start out utilizing the generative AI capabilities inside SageMaker.

Moreover, entry to Amazon Bedrock, a totally managed API-driven service for generative AI, might be provisioned for particular person accounts based mostly on mission necessities, and the customers can entry Amazon Bedrock APIs both by way of the SageMaker pocket book interface or via their most well-liked IDE.

There are further concerns regarding the adoption of generative AI in a regulated setting, comparable to healthcare. Cautious consideration must be given to the worth created by generative AI functions towards the related dangers and prices. There’s additionally a must create a threat and authorized framework that governs the group’s use of generative AI applied sciences. Components comparable to information safety, bias and equity, and regulatory compliance have to be thought of as a part of such mechanisms.

Conclusion

Philips launched into a journey of harnessing the ability of data-driven algorithms to revolutionize healthcare options. Over time, innovation in diagnostic imaging has yielded a number of ML functions, from picture reconstruction to workflow administration and therapy optimization. Nevertheless, the varied vary of setups, from particular person laptops to on-premises clusters and cloud infrastructure, posed formidable challenges. Separate system administration, safety measures, help mechanisms, and information protocol inhibited a complete view of TCO and complex transitions between groups. The transition from analysis and growth to manufacturing was burdened by the dearth of lineage and reproducibility, making steady mannequin retraining tough.

As a part of the strategic collaboration between Philips and AWS, the AI ToolSuite platform was created to develop a scalable, safe, and compliant ML platform with SageMaker. This platform offers capabilities starting from experimentation, information annotation, coaching, mannequin deployments, and reusable templates. All these capabilities had been constructed iteratively over a number of cycles of uncover, design, construct, check, and deploy. This helped a number of enterprise models innovate with velocity and agility whereas governing at scale with central controls.

This journey serves as an inspiration for organizations trying to harness the ability of AI and ML to drive innovation and effectivity in healthcare, finally benefiting sufferers and care suppliers worldwide. As they proceed to construct upon this success, Philips stands poised to make even better strides in bettering well being outcomes via modern AI-enabled options.

To be taught extra about Philips innovation on AWS, go to Philips on AWS.

In regards to the authors

Frank Wartena is a program supervisor at Philips Innovation & Technique. He coordinates information & AI associated platform belongings in help of our Philips information & AI enabled propositions. He has broad expertise in synthetic intelligence, information science and interoperability. In his spare time, Frank enjoys working, studying and rowing, and spending time together with his household.

Irina Fedulova is a Principal Knowledge & AI Lead at Philips Innovation & Technique. She is driving strategic actions targeted on the instruments, platforms, and finest practices that velocity up and scale the event and productization of (Generative) AI-enabled options at Philips. Irina has a robust technical background in machine studying, cloud computing, and software program engineering. Outdoors work, she enjoys spending time along with her household, touring and studying.

Selvakumar Palaniyappan is a Product Proprietor at Philips Innovation & Technique, answerable for product administration for Philips HealthSuite AI & ML platform. He’s extremely skilled in technical product administration and software program engineering. He’s at present engaged on constructing a scalable and compliant AI and ML growth and deployment platform. Moreover, he’s spearheading its adoption by Philips’ information science groups to be able to develop AI-driven well being programs and options.

Adnan Elci is a Senior Cloud Infrastructure Architect at AWS Skilled Providers. He operates within the capability of a Tech Lead, overseeing varied operations for shoppers in Healthcare and Life Sciences, Finance, Aviation, and Manufacturing. His enthusiasm for automation is clear in his intensive involvement in designing, constructing and implementing enterprise degree buyer options inside the AWS setting. Past his skilled commitments, Adnan actively dedicates himself to volunteer work, striving to create a significant and constructive affect inside the group.

Hasan Poonawala is a Senior AI/ML Specialist Options Architect at AWS, Hasan helps clients design and deploy machine studying functions in manufacturing on AWS. He has over 12 years of labor expertise as an information scientist, machine studying practitioner, and software program developer. In his spare time, Hasan likes to discover nature and spend time with family and friends.

Sreoshi Roy is a Senior International Engagement Supervisor with AWS. Because the enterprise associate to the Healthcare & Life Sciences Clients, she comes with an unparalleled expertise in defining and delivering options for complicated enterprise issues. She helps her clients make strategic aims, outline and design cloud/ information methods and implement the scaled and strong resolution to fulfill their technical and enterprise aims. Past her skilled endeavors, her dedication lies in making a significant affect on folks’s lives by fostering empathy and selling inclusivity.

Wajahat Aziz is a frontrunner for AI/ML & HPC in AWS Healthcare and Life Sciences staff. Having served as a know-how chief in numerous roles with life science organizations, Wajahat leverages his expertise to assist healthcare and life sciences clients leverage AWS applied sciences for growing state-of-the-art ML and HPC options. His present areas of focus are early analysis, scientific trials and privateness preserving machine studying.

Wioletta Stobieniecka is a Knowledge Scientist at AWS Skilled Providers. All through her skilled profession, she has delivered a number of analytics-driven initiatives for various industries comparable to banking, insurance coverage, telco, and the general public sector. Her information of superior statistical strategies and machine studying is effectively mixed with a enterprise acumen. She brings latest AI developments to create worth for patrons.