The earliest article regarding use of AI in biomedical sciences was published in 1996 [45]. The extracted articles were classified by date into three major phases: pioneering phase (1996–2005), expansion phase (2006–2015), and prosperous phase (2016–2023) (Figure 2). In the pioneering phase, four articles attempted to explore AI in biomedical sciences, though none addressed the educational perspective. Two focused on clinical perspective (Liu 2001 and Prank 2005) [46, 47] while the other two examined the research perspective (Udelhoven 2000 and Giacomini 1996) [45, 48]. Liu 2001 aimed to develop a system capable of understanding and processing complex biomedical terms in written formats [46], while Prank 2005 predicted genotype from complex biochemical data, comparing linear and non-linear analytical methods to the performance of experienced clinicians [47]. Udelhoven 2000 established a hierarchical classification system for identifying bacteria using infrared spectra [48], and Giacomini 1996 employed AI to classify patients with HIV-1 using neural networks for improved clinical follow-up [45]. The main limitation of the models and algorithms during this decade was their complexity. Despite this, all studies reported high efficiency and accuracy with universal applicability [45]. All the studies in the pioneering phase were conducted in high-income countries including USA, Germany, Italy, followed by Taiwan and others.

In the expansion phase, 11 articles were published, two of them addressed the educational perspective (Munshi 2006 and Zheng 2015) [49, 50]. Munshi 2006 utilized simulation and visualization to focus on computational approaches for understanding molecular and cellular systems, reporting high reliability but limited applicability [49]. Zheng 2015 explored linking entities from unstructured full-texts of biomedical literature to 300 ontologies for automatic knowledge enrichment for scientific literature. Their approach had high accuracy and reliability, although the complexity of the algorithm was a limitation [50]. Five of the remaining articles discussed the research perspective, covering various aspects such as problem-solving, linguistics, handling narrative texts, and analytical schemes (Supplementary Tables S1, S2). The remaining four articles discussed the clinical perspective including clinical data assessment and diagnosis of infectious diseases (Supplementary Tables S1, S2). The majority of the opportunities reported in this decade highlighted high accuracy, reliability, and efficiency. However, the complexity of algorithms and limited model robustness were noted as significant limitations.

In the prosperous phase, 177 articles were published, a significant increase compared to the 11 published in the expansion phase and four in the pioneering phase. The majority, 122 articles, focused on the clinical perspective across various disciplines, with microbiology being the most discussed, followed by histotechnology, general laboratory sciences, haematology, and others (Supplementary Tables S1, S2). These articles primarily addressed the testing and diagnostic aspects of existing models. Fifty-one articles explored the research perspective, discussing the development and training of models to enhance research findings in biomedical sciences (Supplementary Tables S1, S2). Of all the articles in the prosperous phase, only four discussed the educational perspective Ambite 2019, Aparicio 2018, Henderson 2020, and Tota 2021 [51–54]. Ambite 2019 explored a tool that enabled computers to learn from data without programming, designing the ERuDIte framework to foster open-source educational resources on the web. Aparicio 2018 investigated how educators teaching biomedical subjects at the university level used intelligent information access systems such as BioAnnote, CLEiM, and MedCMap. Henderson 2020 explored the use of AI to improve the learning environment by supporting student engagement and activities. Tota 2021 proposed a telepresence robot kit with modules for teaching and conducting remote laboratory experiments.

Articles published during this phase highlighted enhanced efficiency, universal applicability, and real-world clinical relevance. However, algorithm complexity, limited model robustness, and moderate accuracy and reliability were reported as limitations.

One-hundred and thirty-four articles were from high-income countries (Supplementary Tables S1, S2), particularly USA (44 articles), followed by Germany (13 articles), Italy (10 articles), United Kingdome and Taiwan (9 articles each), and Japan (5 articles). Five of the AI models exploring the education perspective were conducted in high-income countries [49–53]. The majority of studies in these countries focused on the clinical perspective (84 articles), with a particular emphasis on microbiology and histotechnology. The opportunities reported in these studies included enhanced efficiency, universal applicability, and real-world clinical relevance. However, limitations such as algorithm complexity, limited model robustness, and moderate accuracy and reliability were also noted.

Fifty-four articles were from middle-income countries, notably China (25 articles), followed by India and Türkiye (7 articles each), Romania (3 articles), Thailand (2 articles), and others. Only one article from this group investigated AI from an educational perspective [54]. Forty articles explored AI from a clinical perspective and 13 articles from a research perspective. The majority explored AI in microbiology (23 articles) and general laboratory sciences (14 articles), with one in forensic medicine [55] and seven in haematology. Similar to studies from high-income countries, those from middle-income countries reported multiple benefits, including enhanced efficiency, universal applicability, and real-world clinical relevance. However, limitations such as algorithm complexity, limited model robustness, and moderate accuracy and reliability were also noted.

Only four articles were published from low-income countries: one article from each of Iraq [10], Bangladesh [56], Ethiopia [57], and Pakistan [58]. Three of the articles investigated AI from a research perspective, while one focused on a clinical perspective. No articles addressed the educational perspective. Three were in the field of microbiology and one in haematology. The articles reported opportunities such as enhanced efficiency, real-world clinical applicability, and high accuracy; however, the robustness of these AI models was limited (Supplementary Tables S1, S2).

Forty-seven articles explored AI in biomedical sciences using deep learning models. The majority (34 articles) examined AI from a clinical perspective, primarily in microbiology as well as laboratory pathology, while 13 articles focused on AI from a research perspective, mainly in general laboratory sciences. None of the deep learning models addressed the educational perspective.

In addition, 81 articles studied AI in biomedical sciences through machine learning models. Of these, 53 articles focused on clinical applications, primarily in microbiology, general laboratory sciences, and clinical chemistry, followed by histotechnology. Twenty-four articles investigated AI from a research perspective, with the general laboratory sciences being the dominant discipline, followed by microbiology and haematology. Three articles, Ambite 2019, Henderson 2020, and Tota 2021 examined the educational perspective around AI in laboratory sciences [51, 53, 54].

Twenty-five articles reported on hybrid AI applications in biomedical sciences, combining machine learning with deep learning, natural language processing, convolutional neural networks, or deep neural networks. The majority of this research focused on general laboratory applications, followed by microbiology, histotechnology, haematology, clinical chemistry, and forensic medicine. While most studies explored the clinical implications of hybrid AI models, none addressed their potential educational applications.

Natural language processing models were reported in seven articles, with six focusing on general laboratory disciplines and one on microbiology. Of these, four studied AI from a research perspective, one from a clinical perspective, and two from an educational perspective (Aparicio 2018 and Zheng 2015) [50, 52]. Neural networks were discussed in four articles, three from clinical perspective and one research, each addressing microbiology, chemistry, haematology, and serology.

Artificial neural networks and automated AI-based diagnosis method were each reported in two articles. Knowledge discovery, semantic analysis, and simulation and visualization were each discussed in one article, with the latter examining AI in biomedical sciences from an educational perspective (Munshi 2006) [49] (Figure 3).

Sixty-three articles focused on microbiology and infectious diseases. The main aims of these studies included pathogen identification, antimicrobial resistance prediction, automated image analysis and digital pathology, vaccine development, predictive modeling for disease outcomes and treatment, and biomedical text mining.

Twenty-four articles focused on laboratory pathology, histotechnology, and cytogynecology. The main aims of these studies included laboratory pathology and molecular analysis for cancer detection whether in body fluids samples or biopsies. Additionally, improving the diagnostic accuracy and consistency in cancer detection and therapy.

Ten articles focused on clinical chemistry, with the main aims including automation in urinalysis, biochemical and lipid analysis, clinical decision support systems, advanced applications in medical imaging, and automated scoring of laboratory assays.

Six articles focused on laboratory management, primarily addressing data management and clinical decision support systems.

Twenty articles examined haematology and blood bank applications, with key aims including automated blood smear analysis, abnormal blood cell identification, blood type identification, bone marrow cell classification, assessment of blood cell integrity in stored samples, screening and risk assessment for hematologic malignancies, blood clot detection, development of hematologic point-of-care testing systems, prediction of hemoglobin variants, and optimization of spectroscopic data analysis.

Fifty-eight articles explored general laboratory applications, those AI applications that are not limited to specific discipline and utilized across all laboratory sciences, as indicated in the extracted articles. The primary aims of those articles covered aspects including; conducting in-depth biomedical literature analyses, enhancing decision-making in intensive care unit laboratories, quality control applications, improving diagnostic accuracy through methods such as enhanced imaging, evaluating the effectiveness of AI use in biomedical laboratories, and proposing machine learning tools for non-expert biomedical technologists.

Four articles addressed serology, focusing on antigen identification and pathogen neutralization, improved immune cell treatment outcomes, classification of immune system diseases, and enhancing the diagnostic capabilities of serological tests through clinical cytometry (Figure 4).

Of the extracted (192 articles), six reported the use of AI in biomedical sciences education [49–54].

The earliest paper investigating AI in education was a population-based, cross-sectional observational study by Munshi 2006. This study was conducted in the USA, employed simulation and visualization AI techniques within the general laboratory sciences discipline. The primary goal was to leverage computational and mathematical approaches to unravel the complexities of molecular and cellular systems. By simulating and visualizing biological processes at multiple scales, the study authors aimed to comprehend the complex biomedical processes, enable the observation and analysis of interactions at these scales, and facilitate hypothesis testing and model validation. While the study demonstrated high reliability, its applicability was limited [49].

A Romanian study by Tota 2021 proposed a computational algorithmic design for developing Telepresence Robotics, a technology to enable remote participation in courses and laboratories, allowing students to interact physically with teaching and laboratory equipment through robotic avatars. The robotic system incorporated AI to facilitate real-time measurements and error correction. The study advocated for a modular Telepresence Robot Kit, easy to assemble and adaptable to various educational settings. Machine learning AI was employed as a key tool in this innovative approach [54].

Another population-based study by Aparicio 2018, conducted in Spain using a qualitative mixed-method design, aimed to gather expert opinions on the use of intelligent systems for processing text to enhance learning activities in biomedical sciences. Interviews were conducted with educators using a questionnaire containing 66 predefined closed- and open-ended questions. The results revealed that teachers highly valued the integration of reliable information sources, bilingualism, and selective annotation of concepts [52]. This study focused on general biomedical laboratory disciplines and utilized a natural language processing AI tool.

A study by Zheng 2015, in the USA, explored a computational supervised entity-linking design. This unsupervised collective inference approach linked entities from unstructured full texts of biomedical literature to ontologies. The method leveraged the rich semantic information and structure in ontologies for similarity computation and entity ranking. Despite using no labeled data, the unsupervised approach outperformed a state-of-the-art supervised method that was trained on a large amount of manually labeled data. The study covered general biomedical laboratory disciplines and utilized a natural language processing AI tool. This method offered the potential to save scientists an enormous amount of time by providing easy access to specific data within a vast pool of information [50].

Another US-based study by Henderson 2020 employed a computational model-based approach to compare various machine learning-based affective models. The research aimed to recognize students' affective states within a game-based learning environment, utilizing competing feature-level and decision-level multimodal data fusion approaches. The results indicated that multimodal affect detectors, driven by posture- and interaction-based data, were effective in identifying emotional states. The AI tool examined in this study was a machine learning AI model [53].

A computational entity-linking study, conducted in the USA by Ambite in 2019, proposed ERuDITe, a platform that organized over 11,000 data science training resources, including courses, tutorials, and talks, specifically for biomedical researchers. Using machine learning, the platform tagged resources with relevant concepts from a newly created data science education ontology and linked people and organizations to public databases like DBpedia and ORCID. This approach provided access to personalized training and connected biomedical researchers to linked educational resources in data science [51].

Ambite in 2019 reported high accuracy and reliability for AI software like ERuDIte, with its comprehensive collection of over 11,000 data science training resources, including courses, video tutorials, and conference talks. The metadata for these resources was uniformly described using Schema.org, enhancing consistency and discoverability. However, the study discussed limitations related to algorithm complexity. Despite ERuDIte’s extensive data collection, the task of automatically identifying and organizing such diverse training resources is inherently complex, and the model’s robustness remains limited [51].

Aparicio 2018 highlighted the potential of AI in higher education, particularly in the biomedical field. Their research explored the benefits and challenges of integrating intelligent information access systems into active learning environments. They identified opportunities to enhance teaching practices, boost student engagement, and gain insights into educators' attitudes towards technological integration in biomedical education. While AI offers significant potential, the study also acknowledged limitations such as algorithm complexity, limited data sources, and technical constraints [52].

Tota 2021 proposed a telepresence robot kit containing easily assembled modules adapted for various teaching situations, specifically for the development of remote laboratory experiments. Users could create their own designs and adapt the electronic components and software to their needs through 3D printing and reprogrammable development boards. With components tailored to each situation, both simple laboratory tasks and complex operations, such as transporting and monitoring laboratory samples, could be performed remotely in real time, offering universal applicability. While this study advocated for telepresence in remote laboratory experiments, it acknowledged the complexity of the algorithm and the limited applicability of the AI software. The use of AI for text detection and recognition is challenged by the need for diverse requirements across different scenarios e.g., text in street scenes for robot navigation vs. receipts for optical character recognition (OCR) in financial departments. Also, difficulties in handling lower-quality or degraded data e.g., scanned legacy books in Google Books [54].

The approach presented by Zheng 2015 significantly outperformed state-of-the-art entity-linking methods, even without using labeled data. This would save scientists, particularly those focused on keeping informed about research developments, an enormous amount of time. All methods were highly accurate and efficient. However, the disambiguation algorithm assumed that phrases within the same sentence or paragraph were related, potentially undermining the entity linking (EL) performance [50].

Henderson 2020 demonstrated that combining different types of data creates a more accurate model for understanding and predicting the unique signals of emotions. By merging these data types, the model developed a comprehensive view of learners, leading to more accurate predictions than using just one data type. However, this study proposed an AI model that focused solely on specific datasets, neglecting some important affective states due to limited available observations. Additionally, the model was designed for specific learning environments, limiting its generalizability and robustness [53].

Simulation and visualization, while powerful tools for understanding complex biological processes, have limitations in real-time clinical practice. Implementing interdisciplinary programs that integrate simulation and visualization can also present challenges, such as group heterogeneity and time constraints. However, Munshi 2006 demonstrated that simulation and visualization can offer significant benefits, such as enhanced comprehension of complex biological processes, the ability to observe and analyze interactions at multiple scales, facilitation of hypothesis testing and model validation, and insights into the emergent properties of biological systems [49].