Advancing Smart Hospital Buildings: Using Granular Data and AI for Outcome-Based BIM

JACQUES LEVY-BENCHETON: Good morning, everyone, and thank you for joining us today. We hope you are enjoying your time at AU in San Diego as much as we are. My colleagues, Mathieu, Nabil, and I are thrilled to be here marking our first AU in this vibrant city. Mathieu and I first met at AU 2019 in Las Vegas, and despite being both French, it was this event that brought us together. Since then, we have been collaborating closely on the topics we are presenting today. With Nabil joining us through a valuable introduction by Laurent Praden from the French Autodesk team. Together, we have been developing our own data and AI platforms.

My name is Jacques Levy-Bencheton. I have been with the firm since 1992. Over 30 years. As a partner and BIM manager, I lead the integration of new technologies, particularly the development of BIM processes, which have been central to our projects since 2005. Today, all our architectural projects are designed, developed, and built using digital models with most following a full BIM process based on Autodesk Construction Cloud and our data and AI platforms.

I'm involved in a few European BIM organization, among them the European Architecture Executive Council, led by Autodesk and comprising some of the most famous architectural practices in Europe.

About our company. Founded in 1981 by architects Jerome Brunet and Eric Saunier. Brunet, Saunier, and Partners has been a leading force in public procurement for over 40 years, with health care projects making up 90% of our current activities. We also work in other fields like universities, high schools, offices, and so forth.

20 of our hospital buildings are currently either in operation or under construction in France and neighboring European countries. And the gross area of these hospital projects hovers between 400,000 and 1,000,000 square feet, and the latest one reached 2,000,000 square feet. We are likewise involved in new hospital design competition in non-European countries.

Our offices are located both in Paris, France and Berne, Switzerland, and we have about 50 architects in total. Our architecture. About 20 years of hospital design has led us to acquire extensive expertise in the health care field. We have developed a unique perspective, both aesthetic and scientific, on changes in health care culture and practice.

And our firm has developed likewise a unique and minimalist architectural style, combining an abstraction of form with an extreme flexibility in order to maximize the capacity of our hospital buildings to evolve. And this is currently a major issue in hospital design.

On this slide, you can get a glimpse of some of our achievements Among them, the Jules Bordet Institute in Brussels, Belgium. 850,000 square feet. Another one is the Gustave Julliard Building, University Hospital of Geneva in Switzerland, 540,000 square feet And maybe another one is the Trauma and Cancer Center in Helsinki, Finland, a building of 750,000 square feet

So the learning objectives related to this presentation first, to identify the functionalities and benefits of a common data environment for collaboration and coordination of digital models in complex design. Second, to implement granular data management within BIM workflows for enhanced efficiency in complex design projects. And the last one, to define AI's roles in automatic design processes to enhance efficiency, particularly in hospital settings.

These different learning objectives we are going to develop are now applied to our latest hospital project, the Saint-Ouen Great Paris North University Hospital. Its code names indicates that the project is located in Saint-Ouen, a suburb North of Paris. And if you follow the Paris Olympic games, it will be very close to where the Olympic village was located.

In fact, it constitutes significant new infrastructure for the whole of Greater North Paris, hence its very large size. It's a project we are developing with Renzo Piano Building workshop, the Parisian workshop with Ingerop Engineering and [INAUDIBLE], which is the economist. This hospital building will replace two existing hospitals due to be closed permanently.

Some key figures of this project. It's a sizeable building of 1,800,000 square feet-- 850 feet in length, 310 feet in width, and 90 feet in height. 20 departments and 1,200 beds. About the design team. 20 different companies and more than 130 users.

It's fair to say that a typical construction project generates a colossal amount of data, and this project at its inception has already generated more than 1,200,000 individual pieces of data. The building includes more than 8,000 rooms, and our data dictionary contains 260 parameters. Obviously, not all these 260 parameters are relevant for each room type. On average, about 150 parameters have actual values filled in for each room type. Mathieu will explain later on what I mean by data dictionary. And the budget of this hospital project reaches $600 million, but it's far from over.

Technological advancement. So I would like to take you through the evolution of hospital project management by BSA showcasing our technological advancements from 2005 to 2024. Until 2005, project management primarily relied on CAD. And back then, data was locked within specific software, making it difficult to access and share information efficiently.

Between 2005 and 2012, we leave the shift towards building information modeling, and BIM allows us for the creation of precise digital representations of buildings, incorporating both geometric data and contextual and functional information. However, during this period, the data remained largely confined within the software.

From 2012 to 2020, we experienced significant progress with the integration of BIM 260. It was the beginning of our common data environment. It enabled data to be integrated within the software while making it accessible and shareable, and this facilitated better collaboration among various stakeholders involved in our project.

Since 2020 and moving forward to 2024, the focus has been on leveraging the hospital BIM data and artificial intelligence with the introduction of our own data platform and AI platform. And these technologies allow for the management of granular data that isn't tied to any specific software or product. And this shift provides increased flexibility and agility in project management, enabling optimization and innovation in hospital infrastructure.

In summary, the evolution of hospital project management tools at BSA highlights a trend towards greater data integration, accessibility, and sophistication, leveraging advanced technologies like database and artificial intelligence.

Some words about the common data environment. Through our journey towards optimizing project management, the introduction of a common data environment has been groundbreaking. Acting as a single source of truth, it has drastically reduced errors and increased efficiency across all our hospital projects.

What exactly is a common data environment? It's a central repository where all lifecycle project information is stored. It supports all the tools needed to efficiently implement the project's workflow. And it's not just limited to digital models created in a BIM environment. It also includes documentation, graphical models, and non-graphical information.

Why do we use a common data environment? To avoid errors, redundancy, rework, missed deadlines, cost overruns, and litigation, and to increase quality control and overall project success.

Let's now have a look at our common data environment today. Our common data environment is the combination of three interconnected platforms, and it evolves constantly. How have we organized this common data environment? The first platform is the Autodesk Cloud platform. The second one is the data platform, and the third one is an AI platform.

These platforms each have a bidirectional link between them. That means the three platforms are interconnected in order to allow the links of information, granular data, files, digital models from one platform to another. And using outcomes-based BIM, we leveraged these connected platforms to make sure our BIM processes line up with specific project goals which help us optimize results and track success more effectively.

About the first platform, The Autodesk Construction Cloud platform, the Autodesk Construction Cloud platform has had a significant impact on our hospital project. It's a complete suite with key modules like Autodesk Docs, which is the basis of the following modules-- Ready to Work Sharing, Design Collaboration, Model Coordinate, and many other modules.

About Docs and Ready to Work sharing, as I said previously, they act as a central hub where all our project documents and Revit models are stored and updated. And this ensures everyone accesses the latest files, cutting down on confusion and errors. And by centralizing our data, we streamline workflows and keep all team members in sync.

Next, in the design collaboration module, effective collaboration is, of course, crucial for any project success, and this module provides tools that boost communication among team members. It allows us real-time updates and task tracking, so everyone knows what needs to be done and when. And this transparency improves our efficiency and reduces missed deadlines.

And finally, we have the Model Coordinate module. It includes powerful tools for coordinating various constructions aspects like clash detection, which helps identify and resolve design conflicts before they become costly on site issues. And by visualizing on coordinate module our synthesis Revit model, which integrates all the engineers' Revit models, we can plan better and avoid potential problems, again thanks to the Model Coordinate tools.

In conclusion, the Autodesk Construction Cloud platform has significantly improved our hospital project by providing a unified and efficient way to manage all constructions aspects. And by leveraging this technology, we ensure our hospital project are completed on time and to the highest quality standard.

About our data platform, we have used Autodesk Construction Cloud for a few years and identified some lags in managing hospital programs and room data sheet. So we needed a database that could reliably manage all hospital data and provide access to all project stakeholders, even those not working directly on a Revit model.

Our first goal was to manage the client program from the initial stage through to the advanced phases, and even during the construction phase with the contractors. And these tools has proven incredibly useful for the client, even during the operational phase of the hospital [INAUDIBLE].

The second goal was to handle technical data sheet for each room type. We created a parameter dictionary to establish correspondences with any programmer working on our different hospital project. We establish a bidirectional link to align database data with different digital models. This way, data can be integrated, modified, and extracted from other source.

And the third goal involves managing the biomedical equipment data sheets for typical rooms and their layouts. And this module allows us to integrate list of biomedical equipment and hospital furnitures for these rooms. It facilitates the use of this layout during biomedical space planning meetings with official hospital users.

In conclusion, this platforms allows us to achieve greater reliability and precision in hospital data management by enabling all project stakeholders to manage their own data on a unified platform based on assigned permissions. And it also ensures precise traceability of modification made by stakeholders, including the client, throughout all project phases. And this enables us to justify design and construction changes.

About our AI platform, once we implemented our data platform, and given that we perfectly understand and manage the client's requirements through managing biomedical equipment and hospital furnitures data by room type, we didn't hesitate to take the next step. I mean, developing an artificial intelligence capable of automatically doing biomedical space planning, integrating hospital equipment and furniture items in their rooms the client asked us to be installed.

There are more than 700 room types in which we need to deliver the layout of biomedical space planning in the North Hospital project. The goal is for this artificial intelligence to gather the necessary information from the equipment sheet of a room type on our data platform, and to propose biomedical space planning options within the room.

Once the most relevant option is validated, the AI opens a Revit model and places the BIM object in the room, of course, selecting the appropriate object from our Revit family library. Our artificial intelligence platform integrates a rating system for the different proposals made for each room, and the users must rate this proposal to allow the AI to learn from the expertise of our architects in this specific field.

The principle of biomedical space planning in selected room types required that once this work is validated during appropriate meetings with the future users, it is deployed and it populates the rooms across the entire project. Our algorithm will enable us to do this for our North Hospital project, which, as I mentioned earlier contains more than 8,000 rooms.

So the benefits for us are huge. I invite you to read these different advantages. These benefits can be broken down into two main categories-- improvement in architectural quality on this slide, and improvement in financial efficiency on this one.

In conclusion, integrating AI biomedical space planning in our hospital project has not only enhanced our architectural precision and creativity, but also significantly improved our financial efficiency. Now, I would like to hand over to Mathieu who will deal deeper into how these advancements are being applied in our latest project. Thank you very much.

MATHIEU LALANNE: Yes. Thank you, Jacques for your presentation. Your vision is awesome. Create an ecosystem based on data. In this context at db lab we worked on a part of this system, the data platform, a dedicated platform built to connect and extract [INAUDIBLE] data from BIM models and store and consolidate data with other data. I will explain all of you what we did.

First I introduced myself. My name is Mathieu Lalanne. I'm the CEO of db lab, a small French company based in Paris and Marseilles. We are specialized in web-based software development for IEC. As a previous architect at db lab we have a real culture of the construction industry, so we work with different kind of sectors like hospitals-- and by example, [INAUDIBLE], or architects like BSA.

With [INAUDIBLE], we are data focused on our different subjects. And because we love using web technologies in our software developments, we are specialized in Autodesk Platform Services-- APS, a wonderful approach to help professionals and developers by proposing tools to communicate with our technologies. We are a certified partner and service provider.

This is what I will talk about today-- how APS help us to build a web software dedicated in data management. Here are the different steps. I will talk about the data flux with Data Hub, connectors, picking data, data storage within DataSets, and much more.

As explained, Jacques, hospital project needs to realize complex missions. Complexity is increased due to building sizes. And because of the hospital size, we have a very large amount of data. Millions of parameters. This is why it's so difficult to manage this with simple Excel files. These missions are control and verify hospital programs from rooms data. control and verify hospital equipment needs. Control and verify door specifications.

These missions has to publish a lot of documents in Excel format or in PDF format. It take a lot of time to do by classical tools, Excel or Revit, et cetera. Jacques came to us to create a specific platform with the objectives of centralizing, controlling, and analyzing specific data to set a mission of the architects. But he also came to us to create a tool with capabilities to automatically create documents that they need to provide to their partners and clients at each mission step.

Here in this diagram, you can see the global concept we built. It's a centralized platform with inputs and outputs. For the inputs, we are connected to Autodesk platform like ACC and BIM 360 to access Revit files. Otherwise, Excel files are also input with an internal uploading system. Outputs are the published files for the different missions, and we built an API to be used by the AI platform.

Thanks to IPS APIs, we have built a platform that is fully integrated with ASIC and BIM 360 platform, whether in the USA or Europe. As you can see on the right side, we have implemented several APIs from the Autodesk catalog. Each API has its own capabilities. To enable user authentication, access to different projects in the ACC, documentary navigation, data extraction from Revit models and more than visualization.

Here we authenticates in the platform. Each user uses his own Autodesk account. Then we access to his assigned project in ACC. You can access to the three folders, files inside, and of course, users' data model.

Here, a project dashboard with data from ACC like user list or project information. In each project, we develop a module with the capability to pass folders and files to find the good Revit file to be connected. In each file, we access the specific data to be used for the architect missions. Here by example, the list of rooms present in the BIM model.

In this hospital project, we have more than 8,000 rooms. That's a lot. We don't need to open and visualize the BIM models to access to the data. This is a capabilities of two API from APIs. This is we will see next.

We call granular data the specific data we need to extract from BIM models. This can be room, like in the image, or doors or windows. Stairs and equipment, by example. But by using the Autodesk API, we don't need to open or visualize the BIM file. We access directly the raw data sent by Autodesk servers. Here, an extract of JSON raw data.

Two API can be used for this. The first one is the first API developed by Autodesk, Model Derivative API, built to translate native file from Revit or IFC in a unique format usable for web platform. This API can also grab data from an object list.

This API is based on the full model, so it's fine to pass the data, but it's not so efficient. So Autodesk proposed a new API completely based on granular data since this year. It used the GraphQL method of requests, so you can grab only the data you need. It's powerful and efficient, of course.

Once the data is connected, isolated, and extracted, the data is stored in a data warehouse. It's an environment where the data is prepared before it's used. We have structured the data with data models. The data model is designed to organize data and give it a context.

So we describe all the data we will store with raw data from BIM models with dictionaries. It's a way to define what kind of object properties we want to store before with summary data to increase performance and with metadata to manage data with complementary information like username, dates, commentaries, and version of data.

We store all of this data in a database. We use MongoDB Database Design for big data with the capability to integrate JSON elements. To proceed data by user, we built interfaces for business tasks. With Dashboard Table Sheets, the data can be compared, analyzed, verified, traceable, edited, [INAUDIBLE] intelligence.

As we saw before, BSA forces these different mission needs to reach different documents. The platform has the capability to create PDF documents for the data rooms' sheets or to create Excel files to be used by the partners.

In conclusion, DBS at the data platform is a way to ensure we have quality data. Quality data is also fair data. The data can be findable, accessible, interoperable, reusable. It's optimized for multiple business tasks. And of course, it's the first step to prepare AI processes. Thank you very much. And Nabil now will talk about AI. Thank you.

NABIL SADEG: Thank you, Mathieu and Jacques, for the introduction. As you've seen, thanks to the DBS app solution, we now have access to information about equipment and rooms. This allows us to explore automating the placement of medical equipment in hospitals using artificial intelligence.

Before we proceed, let me take a moment to introduce myself. My name is Nabil Sadeg, and I serve as the Chief Technology Officer at Zedsoft, a UK-based software development company specialized in artificial intelligence and health care. My expertise lies in high performance graphics, particularly in real-time computation and optimisation. I've also worked extensively in health care software development, leveraging artificial intelligence for the discovery of new biomarkers, as well as in the field of cybersecurity.

Now that you know a bit more about me, let's break down the challenge we are addressing. At first glance, this might seem like a complex problem with many unknowns and variables. But let's simplify it. At its core, this is a space optimization challenge. Our goal is to maximize space efficiency around and between the equipment while maintaining a fully functional layout.

The first step is to define our constraints to build a clear framework. So let's take a closer look at them. We have five main constraints. First, adaptability. Brunet-Saunier already has an effective pipeline that has worked for a long time. Our aim is to improve it, not disrupt it. Second, integration. The system must seamlessly fit into their existing workspace, tools, and workflows. We want them to adopt it gradually without interfering with current operations.

Third, scalability. The system should be primarily used during working hours and may scale up or down based on demand. It should not consume resources when not in use. Fourth, lightweight. We want to avoid any need for Brunet-Saunier to invest in new hardware or manage additional infrastructure. Lastly, user-friendliness. It must be simple enough that anyone in the office can use it without any prior technical knowledge.

With these constraints in mind, it's time to make key architectural decisions. Each of these constraints must be considered and addressed through our technical decisions. First, platform compatibility. Brunet-Saunier uses Revit and Autodesk Construction Cloud, so our system must obviously integrate with those platforms.

Next, cloud infrastructure. To reduce dependence on-premise devices, we'll use a cloud-based solution. We'll also use a serverless architecture in the cloud, ensuring both scalability and efficiency. Finally, we'll develop a progressive web app to ensure a simple, responsive system that is accessible across devices and easy to update.

Now, let's talk about our data sources. On one side, we have the DBS app API, which provides a list of objects to be placed in each room. On the other, we have the Revit API, which gives us room geometry and access to object libraries. This data is then processed by our computational model.

The computation process is implemented in two distinct phases. First, we have implemented a rule-based space optimization algorithm. This initial implementation is simpler and faster to use and to develop in the artificial intelligence approach. It helps us lay the foundation of the pipeline while allowing us to explore and understand the data in greater depth.

Then we move on to the AI model integration. Here, the artificial intelligence is incorporated into the pipeline, utilizing the insights and improvements gained from the rule-based approach to significantly enhance the system's efficiency and performance.

Our initial algorithm provided great insights, but it also revealed some challenges. The algorithmic approach has highlighted two key issues. First, the objects in the library had discrepancies. We worked with architects to normalize metadata and fix object orientations.

Second, the data set was too small. To address this, we implemented two solutions. First, a feedback loop for continuous data gathering. Secondly, Brunet-Saunier typically works with room types which serve as templates. These are furnished and reviewed by stakeholders before being replicated across similar rooms.

While this approach is efficient for manual furnishing, it does not generate enough data for our needs. To address this, we shifted our focus to working with individual rooms, which increase the data set tenfold.

One important aspect to keep in mind is simplicity. The system must be intuitive enough for users to operate without any prior technical knowledge. We've adopted the four-step user process. First, users access a dashboard to browse, search, and filter jobs. Next, they can create a job, selecting rooms from the interface or Revit.

Once the artificial intelligence generates solution, the user reviews them and chooses one. The system applies the selected one in Revit. Finally, every solution is rated to provide quality data for the AI improvement.

Now let's focus on the feedback step. For every room added to the job placement, the system generates three placement proposals. Each proposal is rated by the architects on a scale from 1 to 5. Additional feedback can be provided, which helps pre-classify future outputs.

This is crucial because it enables us to significantly expand the data set over time with high quality data that has been reviewed and approved by architects. Therefore, this feedback is critical for training the AI and improving its accuracy.

Now let's take a quick look at the system's technical architecture. On the left, you can see what runs on the user's machine with minimized local components to simplify maintenance. The rest of the architecture runs on the cloud. Our main API, which manages jobs, tracks progress, and collects reviews, is serverless and scales automatically.

Job computations are also serverless, running a Docker container when needed and returning the solution before shutting down. With this architecture, we can scale the entire system from zero to full capacity.

Not that the architecture is in place, we're ready to move forward with the artificial intelligence development. So at this stage, we have a fully developed and deployed pipeline using the rule-based algorithm. The user interface is in place. The data has been normalized. We are now in a good position to develop the artificial intelligence model.

Our approach here is to train a model that learns to autoregressively predict sequences of objects to place in a room. For each element in the sequence, the model considers the room layout using an orthographic projection, the objects already present, and their associated properties. And finally, it predicts the object's class, size, location, orientation, and whether or not the object needs to be mirrored.

Because we are working with highly specific room types, our model uses specific object labels. For instance, if there are three types of beds, we use three distinct class labels rather than a unified one. This might make the size prediction seem irrelevant, as we're not using it to reverse search for the specific bed. However, we use the size and orientation prediction to deduce or, at time, guess the correct orientation. This is important because the objects in our data set were modeled in different coordinate frames, and some degree of guessing is required.

Additionally, we have multiple placement models, each designed for a specific category of rooms. Since rooms can be complex with various inputs, it is more efficient to train one model per room category, requiring less data. Each room is sorted by a sorting algorithm that determines the category based on the equipment to be placed.

Once categorized, the room is processed using the appropriate model. This approach not only allows for more specific models and simplified computations, but also enables parallel processing. If different room types are part of the same job, computation can be run simultaneously.

With this setup, we now have a robust architecture, high quality data and a user friendly interface. The architecture is highly modular, which allows us to extend it further. Future development could include implementing a more comprehensive library management system, automating door placements, and optimizing lighting in rooms.

Finally, we can now extend our models and automation, bringing us closer to outcome-based BIM where automation works in the background to deliver results autonomously. Thank you very much for your attention.