Exploring approaches to medical AI deployment in Norwegian hospitals: In-house development and commercial procurement
Research paper, IRIS 2024 (Uddevalla, Sweden)
This paper explores the deployment of artificial intelligence (AI) in the Norwegian hospitals, focusing on two main approaches: in-house development and commercial procurement. In-house development involves hospitals creating their own AI technologies using internal expertise and data. Commercial procurement, on the other hand, entails acquiring certified, commercial AI applications. The study examines the advantages and challenges of each approach through interviews with a variety of stakeholders in the Norwegian healthcare sector involved with AI deployments for hospitals. Preliminary findings suggest that both approaches have their distinct characteristics, and this paper aims to present a nuanced view of the current landscape to assist healthcare decision-makers in choosing the approach that best fits their needs. We also propose initial recommendations for guidelines and standardisation of best practices that could support the successful deployment of both approaches. The study is still in its early stages and ongoing, with six interviews completed and plans for further data collection.
In-house, Procurement, Artificial Intelligence, Hospitals
In het kort
Dit artikel vergelijkt twee manieren waarop Noorse ziekenhuizen AI in gebruik nemen: zelf ontwikkelen (in-house) of een gecertificeerd commercieel product inkopen. Op basis van kwalitatieve interviews met betrokkenen uit de Noorse gezondheidszorg beschrijven we hoe zij beide aanpakken ervaren.
Over in-house ontwikkeling vertellen de informanten dat AI-modellen uit onderzoeksprojecten zelden de kliniek bereiken. Onderzoekers worden afgerekend op publicaties, niet op bruikbare producten, en stappen na afloop van een project vaak over naar het volgende onderzoek. De technology transfer offices die onderzoeksresultaten naar producten moeten brengen, hebben daarvoor te weinig geld. CE-markering en de Medical Device Regulation worden ervaren als lang, duur en ongeschikt voor kleine partijen, ook wanneer de AI alleen beslissingsondersteuning biedt. Samenwerking met grote leveranciers die al in het ziekenhuis aanwezig zijn, zoals PACS-leveranciers, wordt gezien als de meest haalbare route naar de kliniek. Tegelijk blijft in-house ontwikkeling volgens de informanten nodig voor multimodale AI die meerdere databronnen combineert, omdat commerciële partijen geen toegang hebben tot die data.
Inkoop van commerciële toepassingen heeft als voordeel dat de certificering al rond is en de prestaties al deels gevalideerd zijn. Het ziekenhuis hoeft alleen nog lokaal te valideren en heeft minder mensen, data en rekenkracht nodig. Nadelen zijn de beperkte mogelijkheden om een gecertificeerd product aan de lokale werkwijze aan te passen en het statische karakter van de producten. Opvallend is dat ook bij inkoop technische kennis nodig blijft om te beoordelen welke producten goed zijn. Verder is er veel meer geld beschikbaar voor onderzoek dan voor implementatie, waardoor projecten “het geld volgen” richting onderzoek. Of een ziekenhuis een academisch profiel heeft, lijkt mee te bepalen welke aanpak het kiest.
De informanten zijn het erover eens dat beide trajecten sneller gaan naarmate ziekenhuizen er meer ervaring mee opdoen, en dat ziekenhuizen die ervaringen met elkaar moeten delen. Zij pleiten voor nationale richtlijnen, mede omdat regels per ziekenhuis en per functionaris gegevensbescherming anders worden uitgelegd. Het artikel spreekt geen voorkeur uit, maar wil beslissers helpen de aanpak te kiezen die bij hun situatie past. De studie was op dit moment nog in uitvoering en verdere interviews stonden gepland.
1 Introduction
The promise of artificial intelligence (AI)-based technologies for medical purposes (medical AI) includes potentials to automate certain tasks currently undertaken by healthcare professionals, serve as clinical decision support, improving clinical work efficiency and diagnostic outcomes (Kaul et al., 2020, p. 807). Currently, the most profound developments within medical AI are based on Machine Learning (ML) or Deep Learning (DL) technologies for image analysis and diagnostics. Such developments position the medical areas of radiology, pathology, dermatology and ophthalmology at the forefront of the ongoing advancements (Rajpurkar et al., 2022, p. 31; Wang et al., 2019, p. 293). However, despite these prospects, the deployment of AI into real-world hospital settings is still limited and in its early stages.
The present situation, with the growing development and availability in medical AI technologies on one hand and limited AI deployments on the other, presents us with an opportunity to examine emerging approaches for AI deployment in hospital settings. This paper investigates two distinct approaches observed within the Norwegian healthcare services: the in-house development of AI technologies and the procurement of commercially available AI applications. In-house development refers to hospitals developing their own AI applications with internal competence and the use of proprietary data to train the AI. These projects typically start out as research projects with the aim of developing AI to be deployed in the hospitals’ clinical settings. The procurement process involves the acquisition of existing, commercially available, CE marked1 AI applications.
Each approach comes with its own distinct characteristics, reflecting the complexity of adopting medical AI in hospital contexts, which are highly regulated and high-stakes environments (see, for example, Higgins & Madai, 2020; Dicuonzo et al., 2023). This variety of factors may influence the decision to develop AI solutions in-house or to acquire them through commercial procurement. Through interviews with a diverse group of stakeholders in the Norwegian healthcare system, this study seeks to shed light on their perspectives regarding these two approaches for AI deployment in hospital settings, and to answer the following research question: How do stakeholders perceive the in-house approach and the procurement approach to AI deployment in healthcare?
Our aim is not to endorse one approach over the other but to provide a nuanced view of the current landscape, shedding light on the distinct characteristics associated with each approach. By doing so, we aim to support healthcare decision-makers in making informed choices that best suit their context. Moreover, considering the early stage of AI deployment in hospitals, this paper proposes initial recommendations for guidelines and standardisation of best practices that could support the successful implementation of both approaches.
2 Methods
In this paper we present preliminary results from research in progress. The main purpose of the study is to explore stakeholder perceptions on the two approaches to AI deployment: in-house development and procurement of already developed and commercially available products. The study is designed as an interview-based study, and the preliminary findings are based on data from 6 qualitative semi-structured interviews with a variety of experts in the Norwegian healthcare system. As the research is ongoing, additional interviews are planned and will be conducted in April and May 2024.
The informants are selected because they are established actors in the field of AI and healthcare and have experience from working with either in-house development projects, procurement projects or both in relation to hospitals. One of the five interviewees did not participate in a project but was interviewed because of having expertise on AI in healthcare in general, and a clinical background providing a user perspective. The roles of the informants are displayed in Table 1. More interviews with informants with similar profiles are planned, but not yet conducted.
| Informants | |
|---|---|
| 1 | Relevant management function at Regional Health Authority |
| 2 | Part of deployment team at a procurement project |
| 3 | Involved with research and development of several projects |
| 4 | Clinical background with familiarity in both approaches |
| 5 | Involved with research, development and procurement projects |
| 6 | Part of deployment team at a procurement project |
The interviews were semi-structured and conducted based on an interview guideline. The informants were asked about their background, their role in relevant past and current projects, their experiences with either or both approaches to AI deployment in hospitals, their views on the approaches’ different characteristics, as well as their views on how the situation could be improved.
Interviews lasted around 1 hour each and were conducted in presence (1 interview) and online (5 interviews). All interviews were audio recorded with consent, and fully transcribed. We analysed the interview transcripts following a thematic analysis approach, in which patterns of recurring topics are identified and coded throughout the transcripts. We followed Braun and Clarke’s (2006, pp. 86–93) six-step method: initial familiarising by re-reading the transcripts; noting down initial lower-level codes; searching for themes in these codes; reviewing; defining themes; producing the analysis. This analytical process enabled us to identify recurring themes such as “regulatory challenges”, “transferability”, “financial/budget limitations” and “incentives of stakeholders”.
The study adhered to national guidelines for research ethics and received approval from SIKT2. All informants were informed about the study’s scope and objectives and provided their written consent prior to participating. To protect confidentiality and ensure privacy, all data was anonymised, with personal identifiers removed or altered without compromising the integrity of the data.
3 Study context
The Norwegian healthcare system is public, and governance is distributed into four Regional Health Authorities. Multiple hospitals in the same geographical region are often grouped together into Health Trusts, which are governed by these Regional Health Authorities. As mentioned in the introduction, the number of realised AI deployments in real-world hospital settings in Norway is currently limited. According to a report from the Norwegian Directorate of Health (Helsedirektoratet et al., 2022), each of the four Regional Health Authorities are utilising medical AI in their hospitals. However, only two out of the five projects mentioned in the report seem to concern use in actual hospital clinical practice and not just as pilot projects. Most of the projects referred to in the report are either research projects, innovation projects or trial projects testing out specific technologies (Helsedirektoratet et al., 2022, pp. 11–14). In august of 2023, a large Health Trust in Norway made the headlines in several national newspapers as the first Health Trust taking AI into clinical use for detecting bone fractures in the image diagnostic department of their four hospitals (Sundby, 2023).
4 Results
In this chapter we present the findings from the analysis of the interview data. We focus on the two approaches, in-house and procurement, based on the informants’ views. We first outline the in-house development approach, then the commercial procurement approach. We conclude with a section where we identify possible improvements to the AI deployment processes as expressed by our informants.
4.1 In-house development
The in-house approach involves the development of the AI technology using internal expertise and resources (Chen & Decary, 2020). As described in the introduction, many of these projects remain in the development phase as of today. According to the informants, a crucial reason for why AI models developed within research projects seldom reach actual use in clinical settings is related to the initial aim of the project. Research projects are concerned with doing research and commercialising the final research results are typically not a part of their plans, as was argued by one of the informants: “[…] You’ve probably heard of the term ‘publish or perish’, right? You have to publish. If you’re a researcher, that’s what you’re measured on. […] So if you ask a researcher to spend one, maybe two, maybe three years of their career developing an application that can be used in a clinic, he or she won’t have any publications for those years. That’s very damaging to a research career. And this is a big problem because many researchers, when they have finished their research and you basically have an idea or code that can be implemented, they move on to another research project.”
For the purposes of transferring research results into products, there are Technology Transfer Offices (TTOs)3. Their purpose is supporting the process of transferring research results into commercial products for use in clinical practice. However, two of the informants perceived these TTOs as ineffective due to limited budgets to succeed with such transfers: “You can contact one of these technological transformation offices, the TTOs. […] They have limited funding. I mean, we tried to do that once. […] I think we got like a couple of hundred thousand in Norwegian kroner funding. That’s not nearly enough to do anything in terms of getting something approved. So I think the funding part is a big challenge. I think there has to be more funding available to take something from a research project to a conventional product.”
Thus, as an attempt to solve the issue of not being able to commercialise research results due to the “publish or perish” issue, which shapes the prioritisations of the researchers, two of the informants said that they had or planned to employ dedicated personnel that could do the commercialisation job for them.
Two of the informants argued that the most viable solution for an in-house developed AI solution to reach the market is to partner up with larger commercial companies (e.g., PACS vendors like Philips, Siemens or Sectra) that are already part of the hospitals’ infrastructure and have established “pipelines” for deployment: “You have to sort of have collaboration with a company that has financial interests in this, or you can develop your own company and you can establish collaborations with the sort of pipelines that are already established in the clinic. […] You have to develop your product to work on these platforms that are already installed and established in the clinic.” However, even to enter such partnerships, the research result still has to be converted into a product that is CE marked and compliant with Medical Device Regulations (MDR) beforehand.
Another factor challenging the process of going from innovation to approved commercial product is how time-consuming and costly such efforts are: “Having a research group or a researcher going from an idea or an algorithm that is performing very well to actually having that AI tool CE marked and FDA approved, that’s very, very long. And it’s time consuming. It’s tedious. You have to have some sort of financial investors or companies behind you to actually do that work.”
Another informant highlighted the difficulties caused by this certification process being the same regardless of the technology’s intended use and risks: “So much documentation is needed, and it doesn’t really matter what the AI tool is doing. It doesn’t differentiate. […] Even if it’s only for decision support and the clinician is still making the decision in the end. It is not built to scale. And it’s not built for small start-up companies, it’s built for big, big companies who have a lot of lawyers and a lot of experience to go through with this.”
Finally, as the situation is today where there are weak mechanisms for transferring research results into products for use in clinical settings, one of the informants explained that after being involved in both in-house development projects for several years and in recent years looking more and more into the potential of implementing commercial products, the informant was more inclined to think that, currently, the procurement approach would be the most common: “95% of the solutions that we will start to use in the hospital are CE marked commercial AI tools that come from the industry”. However, the informant underscored that, in the future, the in-house development approach will be needed to enable the development of more multimodal AI technology that builds on multiple data sources. Such complex models will be difficult to develop for commercial actors as they do not have access to this multitude of data sources: “There is so much data in healthcare in various modalities and shapes. And what we now see in the commercial market are the low-hanging fruits, the tools that have been developed based on one image modality. So, for instance, based on only CT images or only X-ray images of the bone fractures. […] But they don’t have competence or access to other types of data that we have. We have a lot of different clinical registries in Norway. […] We have all the EHR data, data from the patient records, text data, all types of omics data. With these new technologies that we have now […], generative AI, we actually see the potential to develop these multimodal models that can be based on both images and text and maybe blood samples all at once. And if we want to take advantage of all that data, at the moment there is only one alternative and that’s to actually develop the solutions yourself. Because the commercial industry doesn’t have access to this data at the moment, so that’s of course another big reason why we should also focus on in-house development.” The informant also points out that, due to this limited data access, there is much unused data that could only be exploited by technologies that were developed in-house: “So, for instance, we have a spine registry. […] We develop a decision support tool for spine surgery, trying to select which patients will actually benefit from a surgery and which patients will not benefit. That’s also not possible to do for a commercial vendor alone. So, those are the two big reasons why we should also focus on in-house development to try to exploit the unused data that is not available for industry.”
4.2 Procurement of commercial applications
The second approach is the procurement of commercial applications that are available on the market. This procurement typically happens through platforms which offer a variety of AI applications, as an informant explained: “There are a lot of AI tools that are available in the clinic right now. And you can look at these platforms that I mentioned, like Siemens, Philips, and Sectra. I think in radiology, there’s about 60 or 70 AI applications that you can buy.”
As described by several informants, procurement of existing AI applications offers several advantages. Firstly, these products have already obtained CE marking and adhere to MDR and other relevant regulations, ensuring compliance and safety. Additionally, the algorithmic performance of these applications has already, to a certain extent, been researched and validated. This means that hospitals only need to conduct local validation processes on their own patient data (e.g., x-rays) to confirm applicability. Another benefit highlighted by informants is the reduced need for resources. Unlike the extensive demands of in-house development, which encompasses compute resources, data access, and development competence, procurement is a more resource-efficient alternative requiring less investment. This might make procurement a better option for hospitals with less resources and/or lower budgets for research and development: “We are in kind of a desperate situation. We need to take new technologies into our clinic to be able to cope with everything that’s thrown against us, with limited resources. So, to do that, we don’t have time and we don’t have resources to develop applications ourselves. To do so, you need a huge material, and there are many hurdles doing that, as you know. But we know that there are applications that are CE marked and are available in the market to be implemented. So, we chose not to develop things ourselves, but to use products that are trained and static, available in the market today.”
The procurement approach does, however, present its own challenges according to the informants. For example, the possibilities for customisation, i.e. tailoring the product to the users’ own workflows, are limited as CE marked products may only be modified to a certain degree. Vendors value feedback and work to customise their application to fit the specific context of the organisation, but they must do so within these restrictions. A more limited ability to tailor the product to the local context may also result in a more complicated deployment process, for example by requiring more training for employees. Two of the informants with experience from AI procurement did point out the AI vendors’ willingness to adjust the technology according to the customers’ needs: “The vendors are very interested in feedback because they’re working continuously on developing these applications. And because you know, these applications they are pretty new in the market and they are continuously improving, and I think at least for the X application that I am most familiar with, it’s upgraded two, three times a year […].”
Interestingly, one informant with experience in both approaches noted that the procurement approach does not mean that there is no need for technical/development competence: “We have to do development and one of the reasons why is because then we get the in-house competence on AI, and with that competence we can learn to evaluate what the best CE marked tools are. We need this competence […] to learn to start to use the best CE marked tools as well.”
Another informant claimed that in-house development projects tend to develop applications which they later compare to commercial applications developed for the same examination area just to find out that its performance is more or less the same. The informant found this approach less fruitful and causing a potential waste of money, asking the question of why one would develop something that already exist. This was also the reason why they believed that for them, at the time being, it was more beneficial to follow the approach of procuring and implementing commercial applications. At the same time, the informant did note that what they procured was static applications not learning from their examinations. However, the informant did not necessarily see this as a weakness, doubting it would be worthwhile to train the application on their own data while having to spend a lot more resources: “If we trained an application on our own patients, it would probably be a little bit better. But you know, the process of getting a CE marking of such an application, it’s a long way.” While perceiving the procurement approach as the best solution for them at the time being, the informant did argue that “… if nothing works at your institution, then you have to develop something yourself, or somebody has to.”
An important distinction between the two approaches may also be related to the kind of hospital behind the project, whether it is a university hospital with a strong research profile, or a primary hospital devoted to patient treatment. As one of the informants explained, there is a difference in the incentives of these two kinds of hospital organisations: “I think it’s because they have special interest in research instead of what is our main interest which is to have benefits in our daily routine.”
Finally, one of the informants representing the procurement approach argued that the reason why there are more research projects than actual deployment projects could be because it is easier to get financial support for the former than the latter: “[…] to get funding for development and doing research, there is so much more money available. […] To get funding for implementation, there’s almost no money available at all”. The informant continued by referring to how the Regional Health Authorities’ budgets for research funding were approximately ten times higher than for innovation projects. Thus, the informant viewed the Norwegian system of prioritising research above innovation as an obstacle for the procurement approach. This prioritisation was further viewed upon as a paradox, as the informant heard political discourse about the need for innovation and changes in the current ways of working, while not talking about the need of more research: “I think that [the available funding] might be a factor for why people are following this research and development track. It’s actually because they follow the money.”
4.3 Overview of both approaches
An overview of the two approaches and their characteristics as described by the informants can be found in Table 2.
| In-house development | Commercial procurement | |
|---|---|---|
| Initial project aims/interests | Academic interests with intention of clinical benefit later | Clinical interests |
| Human resources | Much competence in different disciplines is needed to develop proprietary AI applications | Less technical competence is needed, although some is still required for assessment of commercial applications |
| Material resources | More resource-intensive in terms of budget, data, etc. | Costs less, but there are also fewer funding opportunities |
| Regulations | CE marking is perceived as a complicated process, and the regulations make no distinction between different kinds of AI technologies | The technology is already CE marked/MDR compliant, but may still need efforts to comply to deployment-related regulations (e.g. data privacy laws) |
| Time | The process takes longer, mainly due to regulatory challenges (CE marking, MDR compliance, deployment-related regulations such as privacy laws) | The time to reaching implementation is shorter, only testing/validation and compliance to deployment-related regulations is needed |
| Transferability/scaling | Harder to scale/transfer to other hospitals. Application platforms may offer a solution there as they are established in the clinical context | Transfer to other hospitals is much easier, although there is still a need for guidance and standardisation of the deployment process |
| Data | While the technology can be trained on local data, informants disagree on whether this indeed provides higher accuracy. This approach may have additional benefits related to otherwise unused data, as well as multimodal AI that can build on multiple data sources | The technology cannot be trained on local data, so it needs to be tested with local data before deployment. The kinds and multitude of data that commercial vendors have access to may be more limited. |
4.4 Informants’ needs and recommendations for the future
In the previous sections, we described the characteristics of each approach. Based on their experiences, the informants also voiced their views on how the landscape could/should change in order to improve the process of deploying AI in hospitals. For instance one informant said: “I think that the main issue is actually that this is new to us. We don’t have experience with going through these processes. […] When we get more experience, these processes will go much faster. […] All of the steps in the process take more time than what they could have taken. And what we see is, as soon as we have done these things, one, two or three times, […] then we see that this is actually to a large extent possible to standardise. […] So I think that this is something when we do it over and over again, we build up competence, we introduce these standards, then it will go much faster already in one or two years from now.”
This sentiment was shared by informants involved with both approaches, as each approach comes with its own challenges to overcome. To share the lessons learned in that process between hospitals was a recommendation the informants all thought to be a good idea. One informant who had been involved with a completed project expressed a willingness to share what they learned and to create and foster a mutually beneficial sharing culture among the Norwegian hospitals: “I hope what happens now is that if we share our experience, we can also contact others and they will share with us the same way. So we’re trying to invest in something here and trying to establish a culture between hospitals in Norway where we try to help each other and share.”
Another informant argued there is a need for standardisation and development of guidelines on a national level as the regulations in themselves are not the only issue, every hospital may also have their own interpretation of these rules: “It’s not only the rules that are the problem, it’s also the interpretation of the rules. And the interpretation of the rules is different between the different hospitals. Different data protection officers interpret the rules in different manners so if this is standardised at the national level, and that we get guidance, guidelines on this, then I think it will help a lot. So for sure it’s possible to do something at the national level to kind of build a roadmap for this to make it more standardized.”
5 Discussion
In this paper, we have addressed the following research question: How do stakeholders perceive the in-house approach and the procurement approach to AI deployment in healthcare? In this section, we discuss our findings about how the stakeholders in our study perceive the two AI deployment approaches in Norwegian hospitals.
Our findings show that stakeholders describe the in-house approach as resource-intensive and misaligned with the commercialisation processes required for clinical application, in part due to the academic focus on publication. Some stakeholders expressed a need to manage this issue by employing personnel specifically dedicated to commercialisation or by working with larger companies. The research politics involved may influence the preferred approach dependent on the type of hospital, with more research-oriented university hospitals for example favouring in-house development.
On the commercial procurement approach, informants highlight the regulatory compliance of these technologies as a significant advantage, allowing for quicker deployment into clinical practice as well as easier transferability between hospitals. However, the static nature of these procured technologies and their limited adaptability to new data or local needs presents a paradox: while they offer more immediate benefits, they may not be able to fully realise the potential of clinical AI in the long-term in the way that in-house developed technologies might. In addition, the informants highlight that there is a discrepancy in financial resources, with less funding being spent on projects aimed at implementation of procured technologies than on research and development projects.
In conclusion, this paper has explored how a variety of stakeholders involved with AI deployment in hospitals perceive two predominant approaches: in-house development and commercial procurement. Our findings described and characterised the two approaches, and we highlighted the nuanced considerations that healthcare decision-makers may have to navigate in choosing how to approach AI deployment for their institution. Finally, we recommend that information exchange is used for the standardisation of processes and development of guidelines for both approaches.
References
Footnotes
Products with a CE mark indicate that they conform to EU standards regarding safety, health, and environmental protection. This marking is essential for products to be sold within the European market. Specifically, AI applications designed for human healthcare, such as for diagnosis, treatment, or monitoring, fall under the Class II category of the Medical Device Regulations (MDR). This means that the CE mark must be obtained through ‘notified body approval’, unlike other products where a manufacturer’s self-declaration might be enough to obtain the CE mark (Malvehy et al., 2022, p. 361; European Union, n.d.).↩︎
The Norwegian Agency for Shared Services in Education and Research (SIKT)↩︎
TTOs are typically owned by Regional Health Authorities, universities, university hospitals, and research institutes.↩︎