Our survey provides detailed insights into German surgeons’ perspectives on the use of AI in CDM in general and for patients with AAP. This must be compared with current literature to derive recommendations for fruitful action.

Our participants perceived the greatest potential for AI in CDM to support organizational tasks such as urgency triage, resource management, or handling workload (Figs. 2 and 7). A representative survey conducted by Deloitte GmbH revealed that 47% of the German population expect the use of AI in medicine will result in increased time for doctor-patient interactions by optimizing medical processes [11]. Both patients and our participants appear to agree with current scientific approaches in this matter. For instance, Farahmand et al. demonstrated that ML algorithms can perform accurate urgency triage for patients with AAP, though the study’s small sample size of 215 patients limits its significance [12]. Similarly, Lee et al. utilized basic information and laboratory values to predict an admission class for patients in the emergency department, achieving an AUROC of 0.97 for intensive care admission prediction [13]. A recent review by Fernandes et al. further concluded that intelligent CDSS could improve patient prioritization in emergency settings. However, limited real-world AI applications in healthcare reveal a lack of validation and implementation [14]. Notably, a gap exists between our participants’ perspectives and current evidence regarding AI’s role in CDM. On the one hand, our participants expressed moderate confidence in AI’s ability to predict postoperative mortality and adverse events (Fig. 7), which is in line with Cobianchi et al.’s insight that international emergency surgeons favor traditional CDM tools [8]. On the other hand, high-quality studies by Bihorac et al. and Mahajan et al. demonstrated that ML algorithms can predict postoperative mortality with AUROC values ranging from 0.77 to 0.83 and 0.95 to 0.97, respectively. Their algorithms also achieved AUROC values between 0.82 and 0.94 for predicting postoperative complications, including neurologic and cardiovascular events [15, 16]. For emergency surgery, Bertsimas et al.’s ML algorithm achieved an AUROC of 0.92 for mortality prediction and 0.84 for morbidity, outperforming traditional scoring systems [17]. Additionally, two recent studies indicate that ML algorithms outperform surgeons in predicting postoperative outcomes [5, 18], and the ACS NSQIP risk calculator has transitioned from a regression-based approach to a ML model to improve accuracy [19]. Based on this evidence and guidelines recommending the use of risk calculators, we encourage surgeons to familiarize themselves with these ML tools for optimized patient care [20]. Both risk calculators and organizational support applications demonstrate desirable roles for AI in CDM. While our findings represent a current snapshot, and opinions may evolve, participants’ perspectives suggest that further research should focus on increased AI implementation and stronger evidence to support its clinical utility.

To better understand the causes of the abovementioned discrepancies, we focus on the characteristics of our participants. Like the findings of Cobianchi et al., we observed a broad spectrum of opinions on AI for CDM. The authors concluded that both enthusiastic early adopters and those who favor traditional methods play a role in the surgical community [8]. Our data suggest that attitudes toward AI may vary by age and hospital type: younger male participants working in tertiary (academic) hospitals appear to view AI’s potential in CDM more favorably and express fewer concerns. This pattern aligns with studies from the USA, Germany, and the Netherlands, where younger, male, and academic individuals of the general population tend to hold a more favorable view of AI’s implementation in medicine [11, 21, 22]. Conversely, Cobianchi et al.’s results show that residents do not report higher use of AI in CDM compared to more experienced surgeons [8]. The reasons behind this and the reluctance observed among other demographic groups remain unclear and warrant further investigation. In addition to demographic factors, personal knowledge and understanding of AI may also influence these attitudes, although our survey did not address this aspect. Interestingly, previous studies have shown that both surgeons and medical students have an equally limited understanding of AI’s technical principles [8, 23], potentially due to minimal exposure to AI topics in medical curricula [24]. The survey by Pinto Dos Santos et al. found that over two-thirds of medical students agree on the necessity of incorporating AI training into medical education [23]. Consequently, we see a strong need for educational initiatives on AI in the surgical community, beginning in medical training. Professional societies should also develop appropriate curricula, and guidelines are needed to establish a clear framework for AI’s use in CDM [24]. Ultimately, combining structured AI training with the development of robust algorithms is essential to foster trust in this new technology. While our study may be subject to some selection bias (e.g., limited representativeness of the German surgical demographic), our findings suggest that skeptics and their concerns should be actively included in this process.

Overall, our participants largely accepted the use of AI in CDM, as only 19% indicated that AI should not be involved. However, respondents agreed on various practical and ethical concerns regarding AI’s implementation in CDM. It should be noted that we predefined the items, and participants could only indicate their level of agreement, not describe their concerns in detail. The most agreed on concern (Fig. 4), was the potential loss of physicians’ clinical judgment. This issue, that the availability of technical aids may negatively affect doctors’ clinical judgment, was documented by Ersoydan et al., who observed a significant decline between 2005 and 2019 in the accordance between suspected diagnoses on abdominal computed tomography scan requests and final diagnoses, alongside a concurrent increase in radiological examinations [25]. Thus, the concerns highlighted by our participants appear to be valid and cannot be disregarded. Additionally, most of our participants indicated concerns about the transparency of AI, as it could become challenging to understand how AI-generated outputs are reached, and that the use of AI could decrease patient trust in surgical CDM (Fig. 4). A Deloitte survey similarly found that the German public is strongly concerned about nontransparent decision-making [11]. Furthermore, Lennartz et al. found that German patients are uncomfortable with the idea of AI making therapy decisions without review by a physician [26]. In the USA, 60% of the public are uneasy with the idea of their doctors relying on AI, and 57% believe that the doctor-patient relationship could deteriorate with AI’s use [21]. Consequently, a review by Tucci et al. concluded that establishing trust in medical AI applications requires a careful balance between algorithmic complexity, comprehensibility, and transparency [27]. The appropriateness of trust in the context of medical AI is currently debated. DeCamp and Tilburt deny this and define trust in the narrower sense as a patient placing their fate in the hands of a doctor in a vulnerable situation, thereby creating an interpersonal connection [28]. In contrast, reliance may be a more appropriate term for discussing human-AI relationships. Kerasidou et al. therefore propose a shift in the AI debate from trust to reliance [29]. While our study did not empirically explore these ethical dimensions, our findings suggest that reliance is a crucial factor for the effective implementation of AI in CDM. To foster this reliance, it is essential for surgical scientists to critically assess AI tools through applied research, generating validated experience and identifying potential biases and limitations. In clinical practice, the physician–patient relationship remains paramount, as it forms the basis of interpersonal trust [30]. Ultimately, we view AI tools, such as intelligent CDSS, as supplements rather than replacements for the human element in medicine.

Building on the preceding discussion, we explore how AI might be effectively implemented in CDM for AAP. It should be noted that we did not fully assess practical challenges associated with the use of AI, such as integration into clinical workflows and friendly user interface designs. The application of guidelines and the latest scientific evidence remains essential for optimal emergency care. Participants in our study, however, rated CDM for AAP as highly dependent on individual experience, poorly standardized, and less evidence-based (Fig. 5). Similarly, the survey by Cobianchi et al. of the international surgical community identified challenges in emergency CDM, including incomplete data and a dependence on personal beliefs [8]. In contrast, German patients surveyed by Lennartz et al. viewed AI as superior to human doctors in applying the latest scientific evidence [26]. Based on this, we believe that the implementation of AI in CDM can address the following problems in traditional CDM: firstly, ML algorithms could ensure structured and standardized data collection by identifying missing variables and prompting physicians to gather additional findings as needed. Secondly, from a clinical perspective, AI could help surgeons to compare collected data with the latest evidence to enhance CDM. This view aligns with conclusions from Cobianchi et al., who saw potential in AI for enabling surgeons to validate their own decision with AI’s recommendations [8]. In fact, a recent review by Fernandes et al. showed that intelligent CDSS, validated in the emergency department, improved consistency and reliability in physicians’ CDM. In contrast, specific research is yet needed to validate this assumption [14]. In addition to effective data collection and evidence-based recommendations, formulating an accurate initial diagnosis is central to the clinical management of AAP [2]. In our study, 46% of participants expressed confidence in AI’s ability to make a correct diagnosis for AAP (Fig. 7). Like all our results, this should be seen in relation to the other items and therefore indicates a rather mediocre level of confidence. Correspondingly, a double-blind, prospective study by Faqar-Uz-Zaman et al. showed that traditional doctor-patient interactions outperformed AI in diagnostic accuracy for AAP (81% vs. 52%, P < 0.001). Our findings similarly emphasize the importance of physical examination, which participants rated as the most critical component of CDM in AAP (Fig. 6). Yet, the combined accuracy of doctors and AI, at 87%, and the reduction in complications through earlier diagnosis exemplify the potential advantages of AI [31]. To us, future AI developments should prioritize human factors and interpersonal interaction, while data privacy and environmental impact were of less concern to our participants (Fig. 4). Provided that essential factors such as safety and usability are addressed, we anticipate that AI applications for AAP could enhance medical care by enabling faster, more reliable treatments.