This study used a mixed methods approach consisting of 2 main analytical processes: qualitative factor extraction and relational analysis. First, episodes of medication-related errors involving nurses were qualitatively analyzed using root cause analysis (RCA) to identify the contributing factors [15]. Next, the data were categorized into 3 groups based on years of unit experience: low adaptation (>1 y), moderate adaptation (1 to >3 y), and high adaptation (3 years or more). For each group, BNM was conducted to evaluate the relationships between the factors and compare the structural differences in the flow of medication-related errors.

This study used data from the Medical Error Database published by the Japan Council for Quality Health Care (JCQHC) [27]. The JCQHC undertakes various initiatives, including the collection of medical errors and near-miss cases, to maintain public trust in health care services and improve their quality. The database includes reports from a total of 1899 medical institutions across Japan. The JCQHC requires participating medical institutions to report incidents that fall into the following categories (1–3). (1) Cases in which inappropriate medical care or management was clearly identified and resulted in patient death, permanent physical or psychological impairment, or the need for unexpected or more extensive treatment than anticipated. (2) Cases in which it could not be conclusively determined whether inappropriate medical care or management occurred, but which were otherwise comparable to category (1). (3) Cases other than (1) and (2) that provide useful information for preventing and reducing the recurrence of incidents within health care institutions. The reporting form of the JCQHC requires detailed documentation, including the patient’s condition and age group, the working status of the staff involved in the error, the circumstances leading to the incident, and measures for future prevention. From this database, 119 cases were selected for analysis based on the following criteria (): (1) a nurse was involved, (2) the error was medication-related, involving adverse events or equivalent medical errors, and (3) the case included sufficient contextual information to allow for RCA.

Although there is no universally accepted definition of nursing competency, it is commonly described as “a core ability required to fulfill nursing responsibilities” [28] or as “the knowledge, behaviors, judgment, skills, values, and attitudes necessary to provide safe, effective, and high-quality care” [29]. These definitions are grounded in experiential learning theory [20] and suggest that competency is a transferable ability developed through the integration of knowledge and skills. In many studies, total years of clinical experience and length of employment are frequently treated as variables that influence competency development [20,30-32]. In this study, informed by the Benner 5-stage model [20] and its application within clinical ladder systems in the world and Japan [21,33], total years of nursing experience were treated as categorical data (<5 y, 5–10 y, and ≥10 y).

In contrast to competency, performance, particularly work performance, is defined as “behaviors or actions that are relevant to the goals of the organization” [34] and as “scalable actions, behaviors, and outcomes” [35]. Work performance is categorized into 4 domains: task, contextual, counterproductive, and adaptive [36]. Fatigue and burnout are closely associated with work performance among nurses [17], which is a critical element influencing patient safety and holds substantial importance in clinical practice.

Adaptive performance refers to “the extent to which an individual adapts to changes in a work system or work roles” [16]. Although this concept has received limited attention in medical and nursing research, it is theoretically associated with experiential learning [21]. In this study, years of current unit experience were treated as categorical data (low adaptation: <1 y; moderate adaptation: 1 to <3 y; high adaptation: ≥3 y) as a pragmatic proxy for the degree of adaptation to a specific clinical environment, informed by the application of clinical ladder systems in Japan and internationally [21,33]. These 3 groups were intended to correspond to levels of clinical practice and functional independence in the workplace and were operationally defined as follows:

This study used data from the anonymized and publicly available database of the JCQHC [26]. Therefore, it did not violate any privacy policies and did not fall under the scope of an ethics review.

The distribution of the 119 incidents by clinical department was as follows. A total of 27 (22.7%) incidents occurred in surgical departments: neurosurgery (n=7 cases, 5.9%), orthopedic surgery (n=5 cases, 4.2%), gastrointestinal surgery (n=3 cases, 2.5%), cardiovascular surgery (n=1 case, 0.8%), plastic surgery (n=1 case, 0.8%), and oral and maxillofacial surgery (n=1 case, 0.8%). An additional 11 (9.2%) incidents were categorized as “other surgical departments,” referring to cases documented only as “surgery” or from mixed surgical wards (Figure 1).

A total of 47 (39.5%) incidents occurred in internal medicine departments: cardiology (n=7, 5.9%), hematology (n=6, 5.0%), gastroenterology (n=5, 4.2%), pulmonology (n=3, 2.5%), otorhinolaryngology (n=3, 2.5%), metabolism/endocrinology (n=3, 2.5%), neurology (n=2, 1.7%), dermatology (n=1, 0.8%), rheumatology (n=1, 0.8%), oncology (n=1, 0.8%), and general internal medicine (n=1, 0.8%). Another 14 (11.8%) incidents were categorized as “other internal medicine departments,” referring to cases recorded only as “internal medicine” or from mixed internal medicine wards.

In addition, 10 (8.4%) incidents occurred in the intensive care unit, 7 (5.9%) in the emergency room, and 4 (3.4%) in the neonatal intensive care unit. Other clinical areas included obstetrics and gynecology (11 incidents), pediatrics (9 incidents), psychiatry (1 incident), and palliative care (1 incident).

The distribution of incident locations was as follows. Patient rooms and wards accounted for the majority of cases (n=97 incidents, 81.5%), followed by outpatient clinics (n=12, 10.1%). Other locations were operating rooms (n=5, 4.2%), radiology examination rooms (n=3, 2.5%), dialysis units (n=1, 0.8%), cardiac catheterization laboratories (n=1, 0.8%), and patients’ homes (n=1, 0.8%).

Medication-related errors were classified into 10 event types, and 23 contributing factors were extracted (Table 1). Descriptive statistics by the unit experience group are summarized in Table 2. In the adaptation group (n=33), defined as nurses with less than 1 year of current unit experience, 8 (24.2%) had 5 to 9 years’ experience as registered nurses and 15 (45.5%) had ≥10 years. These findings indicated that this group was not composed solely of nurses with limited clinical experience.

The BNM results for the 3 adaptation groups (low, moderate, and high) are presented in Multimedia Appendices 1-3, and the comparison of their structural characteristics is summarized in Table 3. In each model, the incident event was positioned at the lowest level, while the most distal contributing factors were represented as level 1. The average entropy of the incident events ranged from 0.44 to 0.76 in the low-adaptation model, from 0.55 to 0.78 in the moderate-adaptation model, and from 0.37 to 0.75 in the high-adaptation model. In the low-adaptation model (Multimedia Appendix 1), the structure consisted of up to 5 layers. Four nodes were excluded from the model: 2 event nodes (“Incorrect storage” and “Wrong route”) and 2 factor nodes (“Patients who do not follow instructions” and “Miscalculation”). Three factor nodes were included in the model but were not connected to any event nodes: “Lack of explanation to patients,” “Failure of information communication,” and “Time pressure.”

The moderate-adaptation model (Multimedia Appendix 2) had a maximum of 4 layers, with 4 nodes excluded from the model: 2 event nodes (“Wrong timing” and “Wrong route”) and 2 factor nodes (“Years as registered nurse” and “Task interruption”). Seven factor nodes were included in the model but were not connected to any event nodes: “Lack of explanation to patients,” “Miscalculation,” “Failure of information communication,” “Failure to detect by the electronic system,” “Invalid double check,” “Assumptions and forgetfulness,” and “Non-reconfirming of inappropriate doctor’s orders.”

In the high-adaptation model (Multimedia Appendix 3), the structure consisted of up to 4 layers. Six nodes were excluded from the model: one event node (“Wrong timing”) and 5 factor nodes (“Time pressure,” “Assumptions and forgetfulness,” “Miscalculation,” “Lack of explanation to patients,” and “Failure of monitoring”). Five factor nodes were included in the model but were not connected to any event nodes: “Insufficient knowledge of medications,” “Failure of information communication,” “Work time,” “Failure of doctor’s orders,” and “Non-reconfirming of inappropriate doctor’s orders.”

The node “Years as registered nurse” was included in both the low- and high-adaptation models (Multimedia Appendices 1 and 3). In the low-adaptation model, all categories except nurses with 5 to 9 years were positively associated with “Failure to confirm the five rights,” and nurses with ≥10 years were positively associated with “Wrong or unnecessary administration.” In the high-adaptation model, nurses with 5 to 9 years were positively associated with “Failure of information communication” and “Invalid double check,” whereas all categories except nurses with less than 5 years were positively associated with “Wrong dose.”

This hypothesis-generating study examined whether years of experience of nurses in their current unit, used as an indicator of unit adaptation, might be related to medication-related errors. While prior qualitative studies have highlighted the importance of adaptation [40], our mixed methods approach using an AI-based BNM suggests the potential for a more objective and quantitative assessment.

Incident causation inherently involves uncertainty because incidents arise from complex interactions among contextual and behavioral factors rather than predetermined outcomes. The moderate to slightly high entropy values observed in this study reflect this uncertainty and should be interpreted as relative indicators. Future studies with larger samples and additional variables, such as work environment factors, may help improve model stability and reliability.

BNM integrates expert judgment while reducing subjectivity through probabilistic modeling, differing from approaches that rely solely on RCA. Consistent with the Reason Swiss Cheese Model [41], our analysis visualized a multilayered structure of medication-related errors. Especially in the low-adaptation model, unsafe conditions and behaviors appeared to precede subsequent unsafe actions, suggesting a potential chain of events. The findings provide preliminary insights into how unit adaptation may influence error structures although this study is limited by its sample size and the interpretive constraints inherent to probabilistic modeling.

International studies on medication incidents have mainly reported that wrong dose errors account for a large proportion of events and that adherence to the “five rights” is a critical factor [37]. The descriptive statistics of incident events and contributing factors in the present study were generally consistent with these prior findings. Although the existing literature is limited, no notable bias was observed in the dataset used in this study.

With respect to RCA, factor extraction was deliberately restricted to elements that could be objectively derived from narrative descriptions; subjective elements prone to inconsistent reporting—such as psychological states and perceptions—were excluded. Furthermore, each incident event and its contributing factors were coded using predefined operational definitions to safeguard objectivity in the extraction process.

For the BNM, relationships among factors were specified with reference to human factors theory, drawing on frameworks [37,39]. These specifications informed the blacklist constraints, thereby supporting the conceptual validity of the resulting network structures.

Regarding the hierarchical structure models, the intent was not to establish definitive causal relationships but to conduct a hypothesis-generating exploration. All nurse knowledge and behavioral factors identified through RCA were treated as potential contributors, and the BNM was used to examine shared structural patterns and differences by unit experience. A greedy algorithm guided by the Akaike information criterion was selected to detect weak associations and latent risks without excluding relevant links, allowing modeling even with a relatively small sample size. However, the limited sample size constrains model stability and robustness, and the results should be interpreted as provisional hypotheses rather than confirmed evidence. The models do not determine causal relationships among factors or between factors and incidents. Future studies should increase the sample size and conduct cross-regional analyses to empirically test the hypotheses generated here.

As the concept of adaptive performance suggests, nurses’ adaptation to their unit is expected to enhance their problem-solving abilities, predictive practice skills, behavioral flexibility, and adjustment to the organizational culture and physical environment in highly uncertain clinical settings [19]. The findings of this study appear to be consistent with these expectations.

As shown in Table 2, incident-related factors did not decline sharply with increasing adaptation. However, the comparison of the 3 model structures (Multimedia Appendices 1-3) suggests that the overall number of layers leading to incident occurrence becomes fewer as adaptation increases. One possible interpretation is that the low-adaptation group shares more homogeneous incident pathways, making certain unsafe conditions more likely to trigger errors. This trend underscores the importance of population-level patient safety education that supports early adaptation among novice and newcomer nurses.

Table 3 further shows that several factors—such as “Work time,” “Failure of information communication,” and “Non-reconfirming of inappropriate doctor’s orders”—were not directly connected to incidents in the high- and moderate-adaptation models. Other factors, including “Time pressure,” “Insufficient knowledge of medications,” and “Lack of explanation to patients,” were not included in the models at all. These patterns suggest two potential effects of unit adaptation: (1) improved adaptation to work processes, such as anticipatory multitasking, increased psychological resilience, acquisition of unit-specific pharmacological knowledge, and the ability to provide appropriate patient explanations [42]; and (2) enhanced adaptation to organizational culture, including smoother communication with colleagues and other health care professionals [43,44]. These interpretations remain tentative, and larger samples are needed to assess the stability and robustness of these structural differences.

Across all models, a consistent structure emerged involving “Wrong patient” and its associated factors—“Failure to confirm the 5 rights” and “Improper use of instruments and equipment.” Addressing these issues will likely require approaches beyond those focused solely on promoting unit adaptation.

Previous studies on clinical experiences have primarily focused on patient safety measures for novice nurses with limited experience [22-25]. However, the inclusion of the node “Years as registered nurse” in both the low- and high-adaptation models indicated that extensive clinical experiences did not necessarily contribute to reducing errors. Instead, positive associations were observed with behaviors such as “Failure to confirm the five rights” and “Invalid double check,” suggesting lapses in verification behaviors.

One possible explanation is the emergence of inappropriate adaptation, in which safety behaviors become ritualized and lose their intended effectiveness. Another possibility is the presence of a cognitive bias whereby experienced nurses are assumed by other staff to be less likely to make errors, leading to the invalidation of double-checking procedures [45]. Although the potential negative effects of extensive clinical experiences have rarely been considered, and the present findings do not allow us to conclude a causal relationship between greater years of experience and errors in verification behaviors, the consistent adherence to safety-verification practices should apply to all nurses, regardless of their level of experience.

The most significant contribution of this study lies in suggesting that, while previous research has emphasized the lack of clinical experiences among novice nurses as a major issue for patient safety [22-25], our analysis of real-world data indicates that challenges in adapting to clinical units may also underlie this issue. This represents a novel application of AI in nursing education [46]. These findings underscore the potential importance of supporting nurses’ adaptation to their assigned units—not only for newly graduated nurses but also for those who have recently transferred to new settings. Educational approaches that emphasize adaptation may also play a meaningful role in the training of other health care professionals. Further hypothesis-testing research will be required.

As possible practical interventions, intensive orientation programs focusing on unit-specific workflows during the initial assignment phase, along with educational initiatives aimed at improving medication knowledge and fostering safe practices, may be effective. In the field of safety behavior training, virtual reality programs have been developed and applied to surgical education [47]. Following this example, scenario-based simulation training can be designed using the results of the present models to reproduce situations in which unsafe nursing behaviors are likely to cascade.

Furthermore, the models presented in this study can be used as feedback tools for safety management and education, reflecting individual differences in clinical experiences and adaptation levels. Although adverse events and medical errors are often influenced by multiple latent factors [37], postevent instructions tend to focus primarily on direct causes due to hindsight bias. The multilayered structures revealed in the current models may help overcome this bias by encouraging more comprehensive and proactive safety-management strategies.

We conducted RCA based on narrative descriptions of reported medical errors. Therefore, situations or latent factors unrecognized by reporting nurses may not have been captured in the analysis. In particular, nurses with extensive clinical experiences often care for critically ill patients or manage heavier workloads, which may serve as underlying factors contributing to counterproductive behaviors, such as task shortcuts or neglect of verification behaviors [48]. However, this limitation is not unique to this study but rather reflects a structural issue common to existing incident reporting systems worldwide. Indeed, prior studies on nurse fatigue risk management have rarely incorporated variables such as patient load and workload in a systematic manner. In addition, even in routine clinical practice, systems for consistently recording indicators such as the number of patients per nurse or patient acuity for the purpose of workload analysis are often lacking. If quantitative approaches to analyzing incident reports—such as those explored in our study—become more established, they may contribute to future international discussions on revising reporting standards.

Another limitation concerns the balance between global and local optimization in the analysis; that is, the extent to which institutional characteristics are incorporated depends on the research purpose and may affect the generalizability of the results [2]. In this study, organizational factors specific to individual institutions could not be extracted from the report narratives. Future research should compare data across institutions and design unit-specific or institution-specific training programs, followed by an evaluation of their effectiveness and acceptability [49].

The findings from the hierarchical network models in this study should be regarded as exploratory and hypothesis-generating rather than definitive. Future work is needed to expand the sample and to evaluate the stability and robustness of the model structures through random-sampling approaches, such as analyzing pre- and post-COVID-19 datasets [50] and employing bootstrap methods [51].

The variable “Years of nursing experience” spans a broad continuous range, so it was converted into categorical data using 5-year intervals, following classifications commonly used in clinical ladder systems [21,33]. This categorization, however, is not definitive. Greater experiences may also reflect age-related factors, and although declines in work ability with age are well recognized in occupational health [52], they have rarely been examined in medical or nursing contexts. Further research is needed to clarify how nursing experience relates to work performance and to determine more appropriate ways of grouping experience levels.

This hypothesis-generating study applied an AI-based BNM to examine structural patterns underlying medication-related errors, demonstrating the value of a quantitative, probabilistic approach that goes beyond the traditional RCA. The use of BNM represents a novel methodological contribution in patient safety research by combining expert judgment with reduced subjectivity and enabling structural comparison across experience groups. The findings suggest that unit adaptation plays an important role in shaping error pathways: low-adaptation nurses showed more layered and homogeneous structures leading to unsafe actions, indicating a higher vulnerability to incident chains. In contrast, total years of nursing experience did not uniformly reduce risk and may even relate to lapses in verification behaviors. Together, these results highlight the need to support adaptation not only among new graduates but also among nurses transitioning to new units, while reinforcing verification practices across all experience levels. Further empirical studies are needed to validate these preliminary hypotheses.