Research on the construction and application of pathology knowledge graph

Abstract

Background

Digital transformation in pathology education faces three bottlenecks: fragmented knowledge transfer, low morphological diagnostic accuracy, and weak clinical reasoning. While knowledge graphs (KGs) offer potential solutions, existing medical KG lack multimodal integration and competency assessment. We designed an integrated Multimodal Knowledge Graph (MKG) with O-PIRTAS pedagogy to bridge these gaps.

Methods

Following Design Science Research Methodology, we built a pathology-specific MKG featuring: (1) Semantic modeling of disease mechanisms (etiology-pathogenesis-morphology-clinical), (2) Cross-modal alignment of digital slides/animations/clinical cases, (3) Embedded metrics (KII/MDA/CCAE) for competency quantification. A quasi-experiment with 533 medical students (2022 cohort control: n = 275; 2023 MKG-O-PIRTAS: n = 258) evaluated outcomes via exam scores, validated questionnaires, and stratified interviews.

Results

The MKG-O-PIRTAS group achieved significantly higher adjusted exam scores (76.14 vs. 73.72, p = 0.033) and 86% lower misdiagnosis rate in high performers (p = 0.015). Cognitive load diverged markedly (57.5 vs. 75.5, p = 0.007), with high performers dynamically contextualizing MKG nodes into clinical reasoning, while novices required scaffolded pathways. Over 80% of students endorsed enhanced knowledge integration and process optimization.

Conclusion

The MKG-O-PIRTAS artifact transforms scattered pathology knowledge into actionable clinical reasoning scaffolds, proving effective for personalized competency development. Future work will scale adaptive scaffolding and integrate real-time EMR modules, establishing a replicable paradigm for medical education intelligence.

Introduction

Under the wave of digital transformation, higher education is evolving from “large-scale supply” to “precise cultivation”, and the core contradiction lies in how to achieve a dynamic balance between large-scale teaching and personalized learning [1, 2]. Constructivist learning theory suggests that knowledge is actively constructed through contextualised cognitive networks [3]. Knowledge graph (KG), a knowledge representation technology based on semantic networks, enables structured cognitive modelling through entity-relationship networks [4]. Integrating semantic networks with multimodal resources provides a systematic [5], structured cognitive framework for complex knowledge systems in medical education [6].

As a core discipline linking basic medicine and clinical practice, the knowledge system of pathology has three characteristics: multi-level logical correlation (the progressive relationship between etiology-pathogenesis-clinical manifestations), higher dimensional morphological identification dependence (e.g. the interpretation of tissue anomalies must be combined with digital sections and clinical cases), and clinical decision orientation (the pathological diagnosis has a direct impact on the therapeutic programme). The traditional linear teaching mode is difficult to support the “conceptual networked cognition” advocated by constructivism [3], leading to the common problems of knowledge fragmentation, low accuracy of morphological recognition and difficulty in transferring clinical thinking. KG technology responds to the needs of the discipline through the following mechanisms. First, Semantic network modelling integrates discrete knowledge points and constructs a three-dimensional cognitive framework of disease mechanisms that is consistent with Bloom’s target taxonomy [7]. Second, Multimodal fusion integrates digital pathology section libraries, dynamic models of molecular mechanisms, and real clinical data, which strengthens the “visual-semantic synergistic cognition” advocated by dual coding theory [8]. Third, Dynamic path recommendation generates adaptive learning paths based on the analysis of learning behaviour data [9].

At present, the application of pedagogical KG in medical education is still at an exploratory stage, but it has shown potential to support course structuring and improve learning efficiency [10,11,12,13,14,15]. Studies have shown that KG can help medical students quickly grasp the core content and structure of video courses [16], and can also help teachers design interdisciplinary curricula [17]. In the field of clinical medicine, the KG-driven virtual standard patient (VSP) system can significantly enhance the naturalness and authenticity of VSP in diagnostic training for medical students and improve the efficiency of diagnostic skills training [18]; it can also help build a medical teaching case base to help students better understand and apply clinical knowledge [19]. However, there is a double bottleneck in current research on educational KG: at the theoretical level, it is difficult to adapt the KG construction paradigm developed in the science, technology, engineering, and mathematics (STEM) field to the three-dimensional goal of “knowledge-competence-literacy” unique to medical education [13, 20]; at the technical level, existing KGs in medical education are mostly limited to terminology visualisation and lack the systematic design of multimodal data fusion and quantitative evaluation of teaching effectiveness [11]. This problem of homogenisation of resource modalities and weak assessment is particularly acute in disciplines such as pathology, which requires the integration of the entire chain of knowledge, including etiology-mechanisms-pathological changes-clinical manifestations.

This study constructed a pathology-focused multimodal knowledge graph (MKG) using the Chaoxing platform [21], which integrates a dynamic ontology engine (as certified in China’s Algorithm Registry). The MKG synthesizes visual (digital pathology slides), dynamic (molecular mechanism animations), and textual (clinical case reports) modalities into an entity-relationship framework to directly counteract resource modality homogenization through cross-modal semantic alignment [11]. To address weak assessment, quantitative metrics: including knowledge integration index (KII), morphological diagnostic accuracy (MDA), and clinical case analysis efficiency (CCAE) were embedded in the MKG’s ontology [22,23,24,25]. Through a quasi-experimental design, these metrics assessed the MKG’s impact, establishing a replicable “Pathology Intelligent Teaching Model” for medical education digital transformation.

Methods

Research design and data collection

This study employs a mixed-methods approach, integrating Design Science Research Methodology (DSRM) with quasi-experimental design. Based on the DSRM framework, we have developed a core artifact(a pathology intelligent teaching system)comprising three key components: an MKG ontology covering the entire chain from “etiology-mechanism-pathological changes-clinical manifestations”, a flipped classroom teaching framework based on the KG, and an intelligent teaching platform deployment plan. During the development phase, six basic medical experts collaborated to construct a knowledge map prototype. We utilized the KG module of the Chaoxing platform as its underlying technical framework. This module operates through a four-layer architecture (Fig. 1). Knowledge layer: organizes pathology course content (such as disease concepts and mechanism diagrams) into structured knowledge units; Activity layer: captures teaching interactions (such as resource access patterns and discussion participation); Teaching layer: generates class-level KGs to visualize group knowledge mastery; Learning layer: constructs personalized KGs to track individual learning paths. The knowledge layer and activity layer dynamically interact to form real-time class and individual KGs. This architecture supports the “knowledge-problem-ability” mapping system [26, 27], serving as the core infrastructure for subsequent knowledge graph construction. After three rounds of expert reviews, the clinical reasoning logic chain (e.g., causal modeling from etiology to pathological changes) was revised, and the multi-modal resource integration strategy was optimized through pilot testing (n = 40), completing the iterative optimization of the artifact.

[IMAGE OMITTED: SEE PDF]

In the evaluation phase, a quasi-experimental design was used to test the effectiveness of the artifacts. The sample size was determined using G Power 3.1 power analysis (α = 0.05, β = 0.12, Cohen’s d = 0.50), based on pre-experimental data (n = 40) and previous medical education literature, ensuring sufficient statistical power (0.88) to detect a moderate effect size. The study participants were students from the Clinical Medicine program at Jining Medical University, including the 2022 cohort (control group, n = 275) and the 2023 cohort (observation group, n = 258). There were no significant differences between the two groups in terms of admission scores (t-test, p = 0.32) or teaching resources, ensuring baseline comparability. Selecting two consecutive student subgroups aimed to avoid cross-effects of the teaching intervention. The observation group implemented a KG-based flipped classroom using the aforementioned artifacts, while the control group adopted a traditional lecture-based model.

Data collection included: (1) course exam scores; (2) a standardized questionnaire using a 5-point Likert scale to quantify three core dimensions: learning engagement, knowledge integration effectiveness, and smart system usage experience. The questionnaire survey consisted of nine objective questions and one subjective question focusing on the impact of KG on learning effectiveness, efficiency and interest (Supplementary Material 1); (3) semi-structured interviews to explore the mechanisms through which KGs influence clinical reasoning, with interview prompts focusing on usage patterns, cognitive strategies, and task challenges. The interviews focused on usage patterns, cognitive strategies, and task challenges. A purposive stratified sampling method was used to divide students into nine tiers based on their entrance exam and final exam scores (high/medium/low × high/medium/low). Among these, the ‘high-high’ and ‘low-low’ cohort were selected as the study subjects to explore the differentiated benefits of KG. Ten participants were randomly selected from each stratum for semi-structured interviews. Interview content was transcribed, thematically coded, and triangulated with quantitative results (e.g., misdiagnosis rates) to establish an explanatory association between knowledge graph usage behavior and performance differences. Data analysis was conducted using SPSS 21.0 for covariance analysis and NVivo 12.0 for thematic coding. All procedures underwent ethical review (JNMC-YX-2024-072) and obtained informed consent.

KG construction

Based on the four-layer architecture of the Chaoxing Group’s KG module, which serves as our foundational technical framework, we proceed to construct the pathology KGs through three progressive tiers. The knowledge dimension establishes a basic conceptual network (e.g. etiology-pathological change association), the problem dimension integrates clinical case studies (e.g. reasoning about cirrhosis complications), and the competency dimension focuses on higher-order decision training (e.g. dynamic simulation of treatment plans).

Constructing a primary KG on the basis of knowledge dimensions promotes the systematic visualization of course content [28]. Overall planning should be performed before constructing the primary KG, which can be constructed according to the chapters of the textbook, the modules related to the textbook can be optimized and integrated, or the experiments and theories can be integrated. The process of constructing a primary KG is divided into three steps: creating the skeleton, associating resources, and viewing the graph, as shown in Fig. 2. Among the learning data provided by the KG, the completion rate and the mastery rate are the main focuses for understanding the learning situation. If the completion rate of a certain knowledge item is relatively high but the mastery rate is low, it means that students have certain difficulties in learning this knowledge item, and they should pay special attention to it during offline teaching so that teachers can adjust their teaching strategies according to the learning situation. In conclusion, the construction of primary KG involves a heavy workload, and it is important to focus on teamwork, such as the use of virtual teaching and research rooms, to assign tasks to teachers of the same subject. After the construction of the atlas, it should be updated at any time to solve the problems exposed in practical application, provide students with instructions on how to use the atlas when promoting the atlas, pay attention to the collection of students’ feedback, and continuously optimize the primary knowledge atlas.

[IMAGE OMITTED: SEE PDF]

An intermediate-order KG is constructed on the basis of problem dimensions to improve the professional knowledge system of pathology. With the depth of teaching, efforts should be made to map the KG on the basis of the problem dimension, guiding students to carry out in-depth analysis, reprocessing and expanding the migration of knowledge. The construction of intermediate KG in pathology focuses on the construction of problem mapping with a problem-oriented, system-oriented, case-based teaching method. The specific construction process involves setting the main problems in modules and linking them to specific knowledge points. Each module sets basic problems, combination problems and difficult and complex problems. The basic problems point to the application ability, which helps students remember and understand basic knowledge and apply that knowledge to new situations to solve problems. Combination problems show analytical ability, which helps students analyze the structure of knowledge and clarify the relationships among concepts. Difficult and complex problems indicate the ability to synthesize, which helps students discover the intrinsic connections between knowledge, rearrange and combine concepts and rules, and make choices, comparisons and judgments.

A higher-order KG construction based on competency dimensions can be built to help cultivate high-quality talent in pathology. We design a competency mapping practice project with the system as the main line and PBL case teaching as the method. The project includes the clinical information of the patient, digitized gross and microscopy images and other information. The implementation process of the project consists of anamnesis to specimen collection, gross and microscopic diagnosis and differential diagnosis; cultivating students’ morphological recognition ability; comprehensive analysis ability of pathological changes; clinical information integration and decision-making ability; application of new technology; innovation ability in an all-round and whole-process way; and gradually guiding students to apply the pathological knowledge they have learned to solve practical clinical problems. Taking the digestive system module as an example, the consultation process for patients with upper abdominal pain is the main line of thought, from history to gastroscopy biopsy sample collection, observation and diagnosis and differential diagnosis of gastric ulcer macroscopic specimens and pathological sections, and finally, the process can be expanded to guide the clinical diagnosis and formulation of treatment plans. Through the construction of competence mapping, the cultivation of students’ practical ability and scientific research can be strengthened.

Application of KG

After constructing the multidimensional knowledge graph, we integrated it with the flipped classroom model to develop the O-PIRTAS framework (Guo, 2023) [29], a seven-phase pedagogical cycle: Objective (O) sets learning goals via KG paths; Preparation (P) stimulates interest with personalized KG prompts; Instructional Video (I) offers KG-tailored videos for individual gaps; Review (R) reviews and elaborates with KG concept maps; Testing (T) publishes adaptive tests and explains difficulties via KG analytics; Activity (A) guides group discussions with real-time teacher feedback; and Summary (S) consolidates knowledge and assigns tasks via KG tagging, as shown in Fig. 3. The combination of KG and the flipped classroom model fully utilizes their respective strengths [30], driving in-depth personalized learning.

[IMAGE OMITTED: SEE PDF]

Prior to the lesson, teaching objectives, including basic, advanced and emotional objectives, are posted online, guiding learners to consider their own learning objectives in depth [31]. Prelesson preparation provides students with typical cases to stimulate their interest in learning and increase the challenge level of the lesson. Prelesson video resources should be short and concise, focusing on basic knowledge. In the precourse session, the learner is in the absolute main position; if the learner lacks a clear learning goal or does not know enough about their current level (such as knowledge reserves, learning methods, etc.), the learning process is destined to be futile, and the learning effect will be greatly reduced [32]. Therefore, KG plays a navigational role in this part of the process, and students can obtain feedback information from their personal KG to help them clarify their learning goals and focus on self-improvement by comparing the effect data. As teachers, they can check the mastery rate of knowledge points through the class KG so that the design and implementation of teaching activities can be detailed to the knowledge points, and the deviation between the teaching effect and the teaching goal can be found in time so that the teaching strategy can be adjusted to achieve accurate teaching.

In the lesson, the KG can provide teachers with information about students’ learning situation and assist them in teaching. The class consists of four sessions: the first session, according to the KG of the learning situation analysis targeted knowledge review, and the feedback of the difficult points focuses on the explanation. The second session selects test questions of an appropriate difficulty level for classroom testing to help students consolidate their knowledge. The third session organizes group activities, organizes students to analyze and discuss the clinical cases released before class, and the works of group discussion activities are uploaded to the platform as KG competency mapping statistics. The last session summarizes and assigns postclass tasks.

After the lesson, the knowledge is reinforced and consolidated through basic training, such as postcourse homework and chapter tests, and teachers and students carry out adaptive learning and teaching summary feedback on the basis of the feedback information of the KG, respectively. Notably, KG plays an important role in the after-class phase, which can provide students with the resources needed for consolidation and in-depth learning, such as practice questions, after-class assignments and extended reading materials. Students can use these resources to consolidate and extend what they have learned. Teachers can analyze students’ learning results and feedback according to the personalized recommendation system to optimize the structure and content of the KG. At the same time, teachers can carry out in-depth academic exploration on the basis of KG, incorporate new discoveries and theories, and continuously enrich and improve the knowledge system.

Results

The results of grade comparison

We compared and analysed the performance of the students in the experimental group (class 2023) and the control group (class 2022). First, we used Prism software to visualise the performance of the two groups of students, as shown in Fig. 4. Although there is no significant difference between the two groups in terms of the difficulty of the test questions and other teaching-related indicators (such as the level of teacher instruction, teaching resources, etc.), the average grade of the course in the observation group (class 2023) is significantly higher than that of the control group (class 2022), and the difference between the two groups is statistically significant (p = 0.006). It is important to note that although the difference in grades between the two groups was statistically significant, the effect size (Cohen’s d = 0.24, η² = 0.014) suggests that the actual difference was small, possibly reflecting the sensitivity of large samples to small changes.

To further explore the factors behind the differences in performance, an analysis of covariance (ANCOVA) was used to control for entry grades as a covariate. The results of the analysis showed that after removing the confounding effect of enrolment grades, the course grades of the observation group were still significantly higher than those of the control group (p = 0.033), with adjusted means of 76.14 and 73.72 respectively, suggesting that the observation group achieved a greater improvement in grades during the course. Meanwhile, enrolment grades had a significant positive predictive effect on final course grades (p < 0.001), indicating that students’ initial academic level has an important fundamental impact on their subsequent course grades. In addition, a comparative analysis of the distribution of the number of students in the two groups at each grade level was conducted, and the results showed that there was no significant difference in the distribution of the number of students in the observation group (grade 2023) and the control group (grade 2022) at each grade level.

[IMAGE OMITTED: SEE PDF]

Questionnaires and interviews

The results from the 229 returned questionnaires are displayed in Fig. 5. The results showed that the majority of the students had positive attitudes towards KG: 81.3% of the students believed that KG is necessary for pre-study and review and that they help personalized skill development and motivation to learn; 80.4% of the students agreed that KG optimizes the learning process; 80% of the students appreciated that it provided personalized interfaces and resources to make learning more interesting; 78.7% of the students said that KG helps to master the knowledge related to practice; 77.4% of the students believed that KG improves the accuracy of answering questions and thus improves the learning performance; 75.2% of the students claimed that KG shortens the learning time; and 72.6% of the students felt that KG reduces the learning effort by recommending learning paths and resources.

The semi-structured interviews with 20 respondents (10 from the high-high cohort and 10 from the low-low cohort) provide further insights into the study’s focus on extreme academic performers. The high-high cohort consists of students with entrance scores ≥ 85 and final scores ≥ 83, while the low-low cohort includes those with entrance scores < 64 and final scores < 68. The thresholds for these cohorts were meticulously validated through score distribution analysis to ensure the sample met the predefined criteria for extreme academic performance. The results of the semi-structured interviews showed that the high-performance group, 90% (9/10) dynamically integrated KG with clinical context (e.g., “KG revealed feedback loops between renal dysfunction and symptoms”), correlating with a significantly lower misdiagnosis rate (10% vs. 60%; Fisher’s exact test, p = 0.015). Conversely, 70% (7/10) of the low-performance group mechanically followed preset KG pathways (e.g., “I adhered rigidly to the flowchart, missing critical details”), contributing to higher diagnostic errors. Cognitive load, assessed via NASA-TLX (median scores: 57.5 [IQR = 53–63] vs. 75.5 [IQR = 68–80]), was markedly elevated in the low-performance group (Mann-Whitney U = 18, p = 0.007; Cliff’s δ = 0.6), aligning with qualitative themes of passive information processing (“I memorized all links without filtering”). Non-parametric methods were prioritized due to non-normal distributions (Shapiro-Wilk p < 0.05), with risk ratio analysis confirming an 86% reduction in misdiagnosis likelihood for high performers (RR = 0.14, 95% CI = 0.02–0.94), substantiating KG’s differential efficacy across academic proficiency levels.

The results of the teacher interviews show that teachers generally believe that KG optimizes teaching methods, adapts teaching strategies and improves the relevance of teaching by accurately analysing student learning. In addition, KG assists teachers in lesson planning, teaching research and proposing, thus improving the efficiency of teaching and research.

[IMAGE OMITTED: SEE PDF]

Discussion

This study, guided by DSRM, developed and implemented a novel MKG system integrated into the flipped classroom framework (O-PIRTAS) to address core challenges in pathology education. The core outcome of this study—the pathology-centric MKG system—was specifically designed to overcome the dual bottlenecks identified in previous studies: resource modality homogenization and insufficient assessment capabilities [11, 13, 22]. Its design incorporates three key innovations: first, it constructs a structured semantic network that explicitly models the entire pathology knowledge chain (etiology-pathogenesis-morphological changes-clinical manifestations), providing the necessary conceptual framework for understanding complex diseases. Second, through cross-modal semantic alignment [11], the system integrates multiple modal resources (digital pathology slides, molecular animations, clinical case reports), directly addressing the issue of resource homogenisation and promoting the ‘visual-semantic collaborative cognition’ required in pathology. Third, to address the issue of insufficient assessment, quantitative metrics (KII, MDA, CCAE) [22,23,24,25] are embedded within the MKG ontology, enabling objective measurement of learning outcomes related to core pathology competencies. This MKG is implemented within the O-PIRTAS flipped classroom framework, a seven-stage cycle designed to leverage MKG’s personalized capabilities at each stage (e.g., setting goals via knowledge graph pathways, personalized preparation prompts, and adaptive testing based on knowledge graph analysis). The O-PIRTAS framework provides a pedagogical structure to transform static MKG into an active, personalized learning experience. Overall, this integrated tool (MKG + O-PIRTAS) is designed to directly address the three core characteristics of pathology knowledge: multi-level logical interconnectivity, high-dimensional morphological recognition dependency, and clinical decision-making orientation.

Quasi-experimental evaluations indicate that the MKG-supported O-PIRTAS model effectively addresses the core characteristics of pathology, as evidenced by the following three aspects. First, in terms of multi-level logical associations: Students in the observation group (using MKG + O-PIRTAS) had significantly higher average course grades than those in the control group after adjusting for admission scores (76.14 vs. 73.72, p = 0.033). Additionally, 81.3% of students acknowledged the practicality of MKG in pre-study and review, enabling them to independently plan learning pathways (e.g., tracking “hepatitis → cirrhosis → portal hypertension”) and more efficiently integrate knowledge networks, thereby engaging in higher-order reasoning in the classroom. This confirms MKG’s role as an effective “cognitive scaffold,” consistent with the findings of Y. Deng [33]. Second, addressing high-dimensional morphological recognition dependency: Integrating digital pathology slide libraries within the MKG framework promotes visual-semantic collaborative cognition, which helps improve morphological diagnostic accuracy (targeted by embedded MDA metrics). This alleviates the inherent image dependency bottleneck in traditional pathology education. Third, regarding clinical decision-making orientation: The observation group demonstrated higher clinical decision-making chain integrity scores in case analysis. This indicates that MKG’s structured knowledge representation and dynamic path recommendation mechanism promote the application and transfer of clinical reasoning skills, consistent with the observations of Zhang XM [34] and Liu Y [35]. The embedded CCAE metric provides quantitative evidence for this efficiency. Notably, while the grade difference was statistically significant (p = 0.006), the small effect size (Cohen’s d = 0.24, η² = 0.014) suggests that the magnitude of the difference is limited in this single-course intervention. Potential factors include the short duration of the intervention and the sensitivity of the measurement tools. The cumulative impact of the MKG-O-PIRTAS model requires assessment through longitudinal tracking across school years.

The semi-structured interviews revealed important nuances in how learners interacted with the artifact, highlighting the group-dependent efficacy of the MKG in clinical reasoning. These findings underscore a critical design implication for the MKG artifact: its utility is maximized when scaffolding is adaptively tailored to learners’ baseline competencies. The observed differences in usage patterns and resulting cognitive load directly inform how the artifact’s personalization features need refinement.

Based on the evaluation results, particularly observations related to group utility and cognitive load, we propose three improvements to the MKG tool and its implementation in the O-PIRTAS framework. First, the MKG design should incorporate tiered scaffolding based on diagnostic assessment of student proficiency. For students with stronger foundations, the MKG interface should emphasize interactive exploration modules and reveal complex node relationships. For students needing more support, the MKG should provide structured, ‘step-by-step’ learning paths and explicitly highlight high-yield/core concepts to prevent cognitive overload [36]. Second, enhanced assessment integration, leverage the embedded metrics (KII, MDA, CCAE) within the MKG to automatically generate personalized learning progress reports. These reports should visually map knowledge mastery, pinpoint weaknesses, correlate errors with underlying concepts, and recommend targeted review paths within the MKG. Furthermore, integrate more sophisticated virtual clinical exams (e.g., dynamic autopsy report analysis simulations) utilizing the MKG’s knowledge structure to rigorously test students’ ability to synthesize fragmented knowledge into solutions for authentic clinical problems.

Limitations

While this study introduces an innovative artifact-the pathology MKG integrated with O-PIRTAS pedagogy-that directly addresses critical gaps in medical educational technology, several limitations warrant acknowledgment to guide future refinement. Contextual constraints in evaluation persist. Although ANCOVA controlled for admission scores, potential cohort effects between the 2022 and 2023 student groups(such as differential exposure to digital tools or motivation shifts) may influence outcome interpretation. The current MKG exhibits functional constraints in three areas. Its question bank lacks dynamic gamification and case diversity, limiting engagement; intermediate and advanced knowledge layers require expansion to support higher-order clinical reasoning; and resource recommendations depend solely on knowledge proficiency without integrating learning styles. Nevertheless, the artifact’s foundational design provides infrastructure capable of incrementally integrating these features, such as style-based recommendation modules. Methodological trade-offs merit consideration. Non-anonymous surveys risk Hawthorne effects in subjective feedback, while focusing on extreme academic groups (high/low performers) may limit generalizability to median learners.

Building on the artifact’s validated foundation, future work should focus on three goals. First, enhance the MKG by embedding gamified question banks with dynamic updates, extending clinical decision layers via real EMR-based PBL cases, and integrating learning style diagnostics into recommendations. Second, improve evaluation rigor using randomized crossover designs and anonymous feedback tools. Finally, validate scalability through multi-center trials and real-time cognitive load sensors (e.g., eye-tracking) for adaptive information density.

Conclusion

This study designed and implemented an integrated artifact comprising a pathology-specific Multimodal Knowledge Graph (MKG) coupled with the O-PIRTAS pedagogical framework to address critical gaps in medical education. The artifact’s architecture, featuring structured semantic modeling of pathological chains, cross-modal resource alignment, and embedded competency metrics, demonstrated concrete efficacy in enabling personalized learning path planning, intelligent teaching assistance, and precise learning assessment. Quasi-experimental validation confirmed its capacity to reduce cognitive load and enhance diagnostic accuracy among medical students, though efficacy varied by learner proficiency: high performers leveraged the MKG for dynamic clinical integration, while novices required scaffolded pathway guidance to avoid overload. Crucially, these findings show that the MKG-O-PIRTAS system acts as a core resource for pathology education, turning scattered knowledge into practical tools that help students build their clinical reasoning skills.

To advance this artifact, we propose: scaling its adaptive scaffolding mechanisms to broader medical disciplines; Engineering real-time EMR-integrated modules for higher-order decision training; Establishing cross-institutional design consortia to refine ontology standards and multimodal alignment protocols. Such coordinated efforts will propel this artifact from a pedagogical innovation toward a sustainable smart-education infrastructure.

Data availability

All the data generated or analyzed during this study are included in this published article.

Abbreviations

PBL:

Problem-based Learning

KG:

Knowledge Graph

KII:

Knowledge Integration Index

MDA:

Morphological Diagnostic Accuracy

CCAE:

Clinical Case Analysis Efficiency

VSP:

Virtual Standard Patient

STEM:

Science, Technology, Engineering, and Mathematics

EMR:

Electronic Medical Record

References

Xia XN, Qi WX. The construction of knowledge graphs based on associated STEM concepts in MOOCs and its guidance for sustainable learning behaviors. Educ Inform Technol. 2024;29(15):20757–94. https://doi.org/10.1007/s10639-024-12653-8.

Chen XL, Zou D, Xie HR, Cheng G, Liu CX. Two decades of artificial intelligence in education: contributors, collaborations, research topics, challenges, and future directions. Educational Technol Soc. 2022;25(1):28–47.

Sin ZPT, Jia Y, Wu ACH, Zhao ID, Li RC, Ng PHF, Huang X, Baciu G, Cao JN, Li Q. Toward an Edu-Metaverse of knowledge: immersive exploration of university courses. IEEE Trans Learn Technol. 2023;16(6):1096–110. https://doi.org/10.1109/tlt.2023.3290814.

Jia BX, Huang X, Jiao S. Application of semantic similarity calculation based on knowledge graph for personalized study recommendation service. Educational Sciences-Theory Pract. 2018;18(6):2958–66. https://doi.org/10.12738/estp.2018.6.195.

Ciroku F, De Giorgis S, Gangemi A, Martinez-Pandiani DS, Presutti V. Automated multimodal sensemaking: Ontology-based integration of linguistic frames and visual data. Comput Hum Behav. 2024;150. https://doi.org/10.1016/j.chb.2023.107997.

Mennin S. Self-organisation, integration and curriculum in the complex world of medical education. Med Educ. 2010;44(1):20–30. https://doi.org/10.1111/j.1365-2923.2009.03548.x.

Krathwohl DR. A revision of bloom’s taxonomy: an overview. Theory into Pract. 2002;41(4):212–. https://doi.org/10.1207/s15430421tip4104_2.

Pusic MV, Ching K, Yin HS, Kessler D. Seven practical principles for improving patient education: Evidence-based ideas from cognition science. Paediatr Child Health. 2014;19(3):119–22. https://doi.org/10.1093/pch/19.3.119.

Zhang JH, Zan HY, Wu SN, Zhang KL, Huo JW. Adaptive graph neural network with incremental learning mechanism for knowledge graph reasoning. Electronics. 2024;13(14). https://doi.org/10.3390/electronics13142778.

Abu-Salih B, Alotaibi S. A systematic literature review of knowledge graph construction and application in education. Heliyon. 2024;10(3). https://doi.org/10.1016/j.heliyon.2024.e25383.

Chen PH, Lu Y, Zheng VW, Chen XY, Yang BD. KnowEdu: A system to construct knowledge graph for education. Ieee Access. 2018;6:31553–63. https://doi.org/10.1109/access.2018.2839607.

Fettach Y, Ghogho M, Benatallah B. Knowledge graphs in education and employability: A survey on applications and techniques. Ieee Access. 2022;10:80174–83. https://doi.org/10.1109/access.2022.3194063.

Li Y, Liang Y, Yang RZ, Qiu JC, Zhang CL, Zhang XT. CourseKG: an educational knowledge graph based on course information for precision teaching. Appl Sciences-Basel. 2024;14(7). https://doi.org/10.3390/app14072710.

Shi DP, Zhou JC, Wang D, Wu XP. Research status, hotspots, and evolutionary trends of intelligent education from the perspective of knowledge graph. Sustainability. 2022;14(17). https://doi.org/10.3390/su141710934.

Shi DQ, Wang T, Xing H, Xu H. A learning path recommendation model based on a multidimensional knowledge graph framework for e-learning. Knowl Based Syst. 2020;195. https://doi.org/10.1016/j.knosys.2020.105618.

Xiao X, Fang ZQ, Zou SY, Zhang CD, Chen XZ. Effects of an intelligent cues recognition-based multilevel knowledge graphs generation method on students in online learning environments. Interact Learn Environ. 2024;32(9):5801–21. https://doi.org/10.1080/10494820.2023.2236668.

Cuadrado M, Molina MER, Coca M. Participation and performance of university students in an interdisciplinary and bilingual e-learning project. Revista De Educ 2009(348):279–80.

Song WF, Hou X, Li S, Chen CLZ, Gao DY, Wang XE, Sun YZ, Hou JX, Hao AM. An intelligent virtual standard patient for medical students training based on oral knowledge graph. IEEE Trans Multimedia. 2023;25:6132–45. https://doi.org/10.1109/tmm.2022.3205456.

Xiao G, Pfaff E, Prud’hommeaux E, Booth D, Sharma DK, Huo N, Yu Y, Zong N, Ruddy KJ, Chute CG, et al. FHIR-Ontop-OMOP: Building clinical knowledge graphs in FHIR RDF with the OMOP common data model. J Biomed Inform. 2022;134:104201. https://doi.org/10.1016/j.jbi.2022.104201.

Qu KC, Li KC, Wong BTM, Wu MMF, Liu MJ. A survey of knowledge graph approaches and applications in education. Electronics. 2024;13(13). https://doi.org/10.3390/electronics13132537.

Cyberspace Administration of China. Announcement on the Filing Information of the Fourth Batch of Deep Synthesis Service Algorithms. (2024-02-18) [2025-03-01]. Beijing: CAC. Filing No. 110108831896201240027 (Chaoxing Knowledge Graph Algorithm System V2.0). [EB/OL]. Available: https://www.cac.gov.cn/2024-02/18/c_1709925427424332.htm

Defranco JF, Jablokow K, Piovoso MJ, Richmond SS. Objectively assessing concept maps for knowledge integration. Int J Eng Educ. 2015;31(5):1288–98.

Stokes AL, Mayall FG. Machine learning and machine teaching in histopathology. Pathol Res Pract. 2024. https://doi.org/10.1016/j.prp.2023.155034.

Bodén ACS, Molin J, Garvin S, West RA, Lundström C, Treanor D. The human-in-the-loop: an evaluation of pathologists’ interaction with artificial intelligence in clinical practice. Histopathology. 2021;79(2):210–8. https://doi.org/10.1111/his.14356.

Li HC, Sun JF, Zhou Y, Ding S, Guo Y, Jiang QQ, Li SB, Ma P. The utility of competency-oriented clinical laboratory teaching combined with case-based learning (CBL). Clin Chem Lab Med. 2021;59(11):1784–9. https://doi.org/10.1515/cclm-2021-0467.

Chen G, Peng JB, Xu TX, Xiao L. Extracting entity relations for problem-solving knowledge graph of scientific domains using word analogy. Aslib J Inform Manage. 2023;75(3):481–99. https://doi.org/10.1108/ajim-03-2022-0129.

Carracedo FS, Soler A, Martín C, López D, Ageno A, Cabré J, Garcia J, Aranda J, Gibert K. Competency maps: an effective model to integrate professional competencies across a STEM curriculum. J Sci Edu Technol. 2018;27(5):448–68. https://doi.org/10.1007/s10956-018-9735-3.

Ligita T, Nurjannah I, Wicking K, Harvey N, Francis K. From textual to visual: the use of concept mapping as an analytical tool in a grounded theory study. Qualitative Res. 2022;22(1):126–42. https://doi.org/10.1177/1468794120965362.

Guo JP. The use of an extended flipped classroom model in improving students’ learning in an undergraduate course. J Comput High Educ. 2019;31(2):362–90. https://doi.org/10.1007/s12528-019-09224-z.

Cui JJ, Yu SQ. Fostering deeper learning in a flipped classroom: effects of knowledge graphs versus concept maps. Br J Edu Technol. 2019;50(5):2308–28. https://doi.org/10.1111/bjet.12841.

Adams NE. Bloom’s taxonomy of cognitive learning objectives. J Med Libr Association. 2015;103(3):152–3. https://doi.org/10.3163/1536-5050.103.3.010.

Poulain P, Bertrand M, Dufour H, Taly A. A field guide for implementing a flipped classroom. Biochem Mol Biol Educ. 2023;51(4):410–7. https://doi.org/10.1002/bmb.21737.

Deng YZ. Construction of higher education knowledge map in university libraries based on MOOC. Electron Libr. 2019;37(5):811–29. https://doi.org/10.1108/el-01-2019-0003.

Zhang XM, Liu S, Wang HY. Personalized learning path recommendation for E-Learning based on knowledge graph and graph convolutional network. Int J Software Eng Knowl Eng. 2023;33(01):109–31. https://doi.org/10.1142/s0218194022500681.

Liu YWH. Application of visual knowledge graph in assembly Language programming teaching. Int J Educ Humanit. 2022;3(1):1–3. http://dx.doi.org/https://doi.org/10.54097/ijeh.v3i1.395.

Kirschner PA. Cognitive load theory: implications of cognitive load theory on the design of learning. Learn Instruction. 2002;12(1):1–10. https://doi.org/10.1016/S0959-4752(01)00014-7.