A comprehensive survey of artificial intelligence adoption in European laboratory medicine: current utilization and prospects (2025)

Adequacy of digital equipment (Table2)

A quarter of the respondents reported having no Wi-Fi network available. Approximately half of the laboratory reported having good or very good internet connection with minimal interruptions. The remainder experience slow speeds or occasional to frequent disconnections, with five laboratories lacking any internet connection whatsoever.

A total of 88.7 % of participants are allowed to use third-party programs on their work computers only with restrictions, limited privileges or not at all.

The majority of laboratories rate the quality of their internet connection, as well as their software and hardware, between 4 and 5 on a Likert scale where 1 indicates very dissatisfied and 6 indicates very satisfied.

Approximately half of the respondents do not use cloud storage, while another quarter (n=49) uses these services daily. Nearly half of this group (n=22) employs a cloud service provided by their corporation. Additionally, 70 % of respondents older than 60 years claimed not to use cloud services at all. The primary reasons for using cloud services, as reported by respondents, are storage, backup and file sharing. A considerable proportion of laboratories (17.9 %; n=35) utilize cloud services for research collaborations, with the vast majority of these (n=29) associated with tertiary care hospitals.

Access to health data (Figure2)

Only half of the participants have access to other data beyond those within their own LIS, while 17 % (n=33) need to request permission from different departments within the hospital. A third of the laboratories lack access to such data due to technical or regulatory boundaries; this group includes the majority of private healthcare centers, privately owned laboratories, and a third of tertiary care hospital laboratories. Meanwhile, 28 and 8 % of laboratories can use laboratory data without restriction for in-house and outside facility use, respectively, while the remaining laboratories are hindered by regulations, technical restrictions, or a lack of knowledge.

Laboratory data management and analysis (Table3)

Only 15.9 % of survey participants deem the speed and volume of data extraction within their local setting to be very or extremely adequate, meeting most or all their needs, while 42 % view this procedure/service not to be very adequate or inadequate. Most laboratories use spreadsheet software such as Microsoft Excel, IBM SPSS, or distinct modules within their LIS for data analysis. However, only a few laboratories have access to more advanced data science software like R or Python environments. Of the 28 and 18 laboratories able to use R or Python, respectively, 19 (68 %) and 12 (67 %) are located in tertiary hospitals.

Artificial intelligence advancements (Figure3)

Half of the laboratories reported not having any personnel with IT skills able to aid in AI processes, while 11 % stated having one or more computer engineers or scientists on staff. Primary care and privately owned laboratories have the highest percentage of employing computer engineers or scientists at 23 and 20 %, respectively. Conversely, secondary and tertiary care hospital laboratories report the lowest percentages, at 4.5 and 8.4 %, respectively .

60 % of survey participants rated their lab team’s knowledge on AI as very low or non-existent, while 4.6 % claimed their team possesses very good or expert AI knowledge. “Very little knowledge” was the most reported level among all types of workplace organizations. When splitting the answers by age group, “very little knowledge” was the most frequent answer among participants aged 30 to 45 and 40–60 years, while “some basic knowledge” was the most frequent answer in the group over 60 years old.

38.5 % of primary care, 58.8 % of private healthcare centers, 70 % of privately owned laboratories, 18.1 % of secondary care hospital laboratories, and 41.2 % of tertiary care hospital laboratories report having the possibility of interdisciplinary AI project collaborations, respectively.

Of the 50 laboratories that reported having ongoing AI projects, either in collaboration with external resources, IVD industry or as in-house development, 35 (70 %) are from tertiary care hospitals. Meanwhile, 71.8 % of respondents indicated that they do not currently have such projects ongoing.

Personal perspective (Table4)

The majority of survey participants believe that lack of skills, difficulties in understanding overly mathematical aspects, and ethical or regulatory issues will be the most prominent obstacles to the use of AI in medical laboratories. In addition to the predefined answer options for this question, several participants provided additional answers as free text (see footnote of Table 4).

Approximately 80 % of laboratories believe that AI will be most heavily involved in the fields of hematology and clinical biochemistry, while around 40 % anticipate improvements in coagulation, immunology and molecular biology and around 30 % in microbiology. When asked about the processes most likely to be shifted towards AI, most survey participants believe that result interpretation and diagnostic accuracy are the areas most likely to experience change. However, nearly all answer options for this question were ticked by at least half of the respondents, with the exception being the pre- and post-analytical phases and error reduction.

More than a quarter of survey participants reported having previously enrolled in an AI course, predominantly online, while nearly all expressed interest in future AI training, preferably at a basic or intermediate level. Nearly half of the participants have teaching responsibilities, with 70 % of them including AI topics in their course, primarily from tertiary care hospital laboratories. Only 6 % reported that one or more of their colleagues teach AI topics.

Data, data, data

FAIR (Findable, Accessible, Interoperable, and Reusable) data is the foundation of effective ML modelling [17], with “Accessibility” being a crucial component of this acronym. While most laboratories do have access to data within their own information system, as previously mentioned, accurate interpretation of these data typically requires the inclusion of accompanying clinical data from individual patients. This information is usually not provided alongside the lab order but can be found in the Electronic Health Record (EHR) system, either as structured data or in free text.

About half of the survey participants indicated that they do not have access to such data or that access was subject to obtaining permission from the. This result aligns with findings from Bellini etal. [18], who conducted a survey on AI and big data utilization in Italian clinical laboratories in 2022. They found that 65 % of laboratories could not acquire health data from sources other than their own LIS [18]. Private healthcare centers and privately owned laboratories appear particularly hindered in accessing data. Although access to the LIS data may be possible, in most laboratories (78.5 %), it requires users with advanced knowledge or assistance from the IT department to extract them. Moreover, only 15.9 % of survey participants found these data met their needs in terms of speed and volume of extraction. Furthermore, even when laboratories aim to use data from the LIS for in-house purposes, only 28 % report being able to do so without any restriction.

The efforts of IT departments and hospital management to protect sensitive medical data in the patient’s best interest are understandable. Similarly, data protection within a healthcare setting is crucial, as mandated by the European General Data Policy Regulations (GDPR) [19]. However, these restrictive regulations have resulted in a system in which accessing connected data from different IT systems in an unfiltered form is near to impossible. The administrative processes required to access even parts of such data can take several months, if not longer. Given the rapid pace of advancements in the field of AI, such delays are not only inconvenient but also prohibitive towards improving healthcare.

Considering the importance of high-quality data, the situation revealed by our survey among European laboratories is alarming. In the absence of distinct processes tailored for AI implementation or data acquisition, it appears that established, unspecific processes within the localand/or (inter)national healthcare setting(s) are being inappropriately and insufficiently applied to the specific needs of laboratories (e.g. general process to tender and acquire hardware for a healthcare facility is being used for AI services).

Appropriateness of equipment and data analysis

After acquiring unfiltered and high-quality structured data, effective utilization of ML in laboratory settings requires certain technical resources. While it is not typical for analytical laboratories to develop and fine-tune or even train large-scale in-house ML models due to the substantial computational resources required, universities and research centers might pursue such endeavors. However, for practical and clinical purposes, connectivity and access to cloud-based ML models remain paramount. Ensuring robust digital infrastructure, including reliable high-speed internet and Wi-Fi networks, is essential as these ML models are generally accessed and maintained via cloud services. Additionally, laboratories benefit from having adequate computational equipment, such as high-capacity computers with multi-thread processing capabilities, using Graphical Processing Units (GPUs) and/or Tensor Processing Units (TPUs), along with necessary programming environments (e.g., R or Python), to facilitate data analysis and research conducted by data scientists. This setup supports the integration of ML into clinical workflows, enhancing diagnostic processes and overall laboratory efficiency. It is also important to recognize that implementing ML requires specialized expertise, which is not easily found. While the availability of skilled professionals in ML is crucial to harness the full potential of these technologies in laboratory medicine, this resource is only scarcely available [20]. In addition, as such ML models may influence medical decisions, they may be regarded as medical devices and thereby fall under the according regulations, requiring thorough validation before approval for clinical use [4].

In our survey we discovered that a quarter of the responding laboratories lack a Wi-Fi network, and nearly half described their internet connection as very slow, slow or average in terms of speed and stability. Additionally, nearly all respondents (88.7 %) reported that they are not allowed to use third-party programs on their work computers or can only do so with significant restrictions or limited privileges.

Cloud computing has become a standard practice for ML accessing and integrating LLM in business processes, using platforms such as Amazon Web Services, Google Cloud Platform, Microsoft Azure, IBM Cloud and others. These systems are scalable in terms of storage and computational power and can be accessed on demand [21]. Essential prerequisites for this approach include data security, a fast and stable internet connection and of course access to cloud services. Despite the advantages of cloud computing, 51 % of survey respondents reported not using cloud services, with 27 % indicating that such use is prohibited by their organization. Among those who may access cloud services, the majority use them primarily for data storage and backup, as well as for collaborations. Only 3.6 % (n=7) use these services to develop algorithms.

In addition, most laboratories reported having access to basic data analysis software such as Microsoft Excel or more advanced statistical software like SPSS. However, only 24.1 % are equipped to use programming environments such as R or Python. Considering the limited access to data and the moderate technical equipment, it becomes evident that programming and training ML models locally or on dedicated cloud services, using a vast amount of high-quality data is rather the exception than the rule.

Knowledge/expertise

While AI has become a part of our daily lives, some basic knowledge of its benefits, risks and potential applications may be presumed. Serbaya etal. reported that 74 % of 361 physicians and nurses believed they understood the basic computational principles of AI [22]. Their study highlighted that healthcare workers generally possess good awareness and an optimistic attitude towards AI, but also harbor concern about potential job displacement in the future. However, these findings do not align with our survey results; only 10.8 % claimed to have good or very good knowledge on AI within their team. Moreover, 21.6 % indicated employing at least one data scientist, computer engineer, or computer scientist, with the highest representation in privately ownedlaboratories and the lowest in secondary and tertiary hospital laboratories. Interestingly, the oldest age group among our survey participants reported having “some basic knowledge”, in contrast to the younger groups, who most frequently reported having “very little knowledge”.

A strikingly high number of laboratories (47.2 %) reported having no team members with IT skills. Meanwhile, 42.6 % indicated the possibility of engaging in interdisciplinary AI projects. However, only 25.6 % of laboratories reported ongoing AI projects, again with privately owned laboratories being the most active and secondary care hospital laboratories the least active. This gap between potential and active projects may be caused from the inadequate environment described earlier.

Personal perspective

Besides image recognition, AI algorithms are well-suited to perform tasks involving complex data analysis, pattern recognition, and interpretation, making them particularly suited for areas with large datasets such as genomics, transcriptomics and proteomics. They are also effective in fields requiring image recognition, such as hematology, microbiology, and urine sediment analysis, or those involving complex patterns with a plethora of interpretation options, like clinical biochemistry [16, 23, 24]. Additionally, recent scientific publications have increasingly focused on using ML models to predict laboratory results from other clinically available data or additional diagnostic information 25], [26], [27.

When inquiring about the lab processes most likely to be affected by AI in the near future, we observed no definitive trend, except that the preanalytical phase is deemed the least likely to be involved across all age groups. This perception may partly arise because many laboratory specialists do not consider the test selection as part of the pre-analytical process. Nonetheless, AI algorithms are already in place to assist in detecting preanalytical errors, such as misidentification and issues with clotted or hemolytic/icteric/lipemic samples [28].

Laboratory specialists among all age groups displayed a strong interest in familiarizing themselves with AI, according to our survey results. This is in line with the results of a survey, conducted by Adler etal., who also saw a strong urge to gain knowledge and train more on digital skills [29]. We found that 28 % of respondents have already participated in an AI course, mostly online, with no significant differences among age groups. Furthermore, 89.7 % expressed interest in participating in such courses, with a preference for intermediate-level training. While advanced courses were less popular overall, younger participants were more inclined to attend them compared to the older age groups. These findings contrast with a 2021 survey by Allen etal. among 1,427 radiologists, which showed that while 33.5 % were using AI in clinical practice, 80 % of those who did not responded that they “see no benefit” and 72 % had no current plans to integrate AI solutions into their practices [6]. In contrast to broader surveys, a more targeted study by Jafri etal. revealed that all 13 participants undergoing a semi-structured interview expressed hesitancy toward integrating AI in clinical laboratories [30]. Conversely, a 2019 survey among physicians indicated that 83.4 % of the participants considered AI as beneficial in the medical field [31]. Our survey revealed that a significant subset of respondents (n=95; 48.7 %) have teaching responsibilities, with 34.7 % of these including AI topics in their courses. Notably, the majority of these answers originated from participants associated with tertiary care hospital laboratories both within and outside universities, which handle 1–10 million tests annually.

Limitations

One limitation of our study is the disproportionate number of survey respondents from Turkey (n=48; 23.5 %) compared to other countries. To avoid misrepresentation, we chose not to perform country-specific evaluations, owing to the limited number of responses per country. This decision was made toensure that our findings remained representative for Europe and unbiased across different national contexts.

We cannot exclude multiple answers per laboratory, since we did not record IP addresses due to the data protection regulation. We need to trust that participants followed the guideline in the invitation letter.

Another bias may emerge from the fact that people, interested in the topic of the survey are more likely to fill it in.