Symbiosis in Health: The Powerful Alliance of AI and Propensity Score Matching in Real World Medical Data Analysis

1. Introduction

The rapid expansion of real-world data and evidence [1,2,3] has driven the adoption of advanced technologies and methods in medicine, including artificial intelligence (AI) and propensity score matching (PSM). AI is applied in various medical fields, such as diagnostics, prediction, imaging and pattern recognition, risk assessment, robotics, education, treatment planning and many others [4,5,6,7,7,8,9,10,11,12,13,14]. In contrast, PSM is primarily a statistical methodology with a more specific purpose. It is widely used in quasi-experimental studies, such as retrospective analyses of medical claims, data from disease or product registries, digital health technologies, routine healthcare datasets, observational studies, and electronic medical records. PSM leverages individual patient covariates to balance potential confounding factors when comparing different patient groups. Adjusting for baseline disparities enables analyses to approximate the reliability of randomized, prospective studies, which are often considered the gold standard in research [15,16,17,18,19,20,21].

The use of PSM and AI in data analysis has garnered significant research interest, evolving from traditional methods with baseline use of PSM to more advanced enhanced AI techniques and combined hybrid approaches and integrations [22,23]. Researchers typically begin with conventional propensity score matching to construct a methodological baseline, as this established technique provides fundamental insights into the data structure before applying alternative approaches [24,25]. However, there is growing interest in using AI and machine learning methods to enhance propensity score estimation (the use of neural networks, decision trees, using AI to automate processes of variable selection and model specifications relied on PSM), especially when dealing with high-dimensional data or complex relationships [26,27].

When thus combined, AI and PMS enhance medical research in a symbiotic way. First, PSM improves AI applications by increasing the reliability and efficiency of machine learning algorithms [28,29,30,31]. Second, integrating artificial intelligence (AI) with propensity score matching (PSM) presents a promising avenue to overcome limitations of traditional PSM and broaden its utility in medical and healthcare research. AI can augment multiple stages of the PSM workflow—including data preprocessing (e.g., natural language, image, and signal processing [32], propensity score estimation, covariate selection, and post-matching analysis. By mitigating biases, refining causal inference, and fostering more robust and interpretable models [33], AI-enhanced PSM holds substantial potential for advancing methodological rigor in observational studies [34].

The above symbiotic association between AI and PSM has not yet been thoroughly and holistically investigated. To close this gap, Synthetic thematic analysis (STA), derived from the synthetic knowledge synthesis [35,36], was used in this study to review for the first time the scope and content of the existing research literature and answer the main research question, namely How and in which context do AI and PSM complement each other? Furthermore, the main research question has been further decomposed into following more specific research question:

• What are the dynamic and spatial features of the research literature production of the AI and PSM use in medicine?
• How is the symbiosis association between AI and PSM reflected in most prolific source titles and most productive countries?
• What research themes emerge in studies combining AI and PSM for medical data analysis?
• What are the more prolific AI methods, medical applications, and diagnoses in combined AI and PSM analyses?
• What are the dominant research trends in the combined use of PSM and AI?

2. Methodology

Synthetic thematic analysis (STA) in this study was performed with the algorithm shown in Figure 1:

To obtain the necessary bibliometric data for extensive bibliometric study, Scopus was used due to its wider interdisciplinary coverage, inclusion of PubMed, advanced search options including AI, and the ability to export larger amounts of data for bibliometric analysis in one run. For the bibliometric analysis, two main bibliometric tools were used, namely Bibliometrix version 5.1.1 and VosViewer version 1.6.20. [38,39]. Bibliometrix was run in RStudio version 2025.09.1+401, which was utilising the R programming language version R-4.5.1 for Windows. VosViewer was run under the latest version of the programming language, Java 8 Update 461, using the Java Runtime Environment (JRE) 1.8.0_461. VosViewer was used to create the bibliometric and thematic mapping based on the author keywords found in the bibliometric data, while the Bibliometric library was used to perform the remaining part of the bibliometric analysis, including base bibliometric information gathering, publication overview, citation analysis, countries production, Lotka’s Law, and Bradford’s Law. Bibliometrix was run and utilized using the biblioshiny method, which allows for interactive web-based engagement.

Figure 2. The algorithm for determining niche, emerging, declining, motor, and basic themes.

Figure 2. The algorithm for determining niche, emerging, declining, motor, and basic themes.

3. Results

Dynamic and spatial features of the research literature production

The search based on the provided search string in Scopus was performed end of September 2025. The search yielded 433 documents from 283 sources that were authored by 3858 authors. Of the 433 documents, 415 were articles, one book, 11 conference papers, one conference review, and 4 reviews. The results encompass documents from 2011 until 2025 with no papers scheduled for 2026. The average citation was 10.73 per document, and the average document age was 2.02, indicating a relatively young and interesting field of study. The average number of co-authors per document is 21, and the international co-authorship is 18.94 indicating global interest and collaboration.

In Figure 3, there is a noticeable spike in the number of published papers over the last few years, indicating a heightened interest in the observed field from 2020 to 2024. There is a small dip in 2025. This is to be expected, as the results encompass data up to the end of September, which is still several months away from the end of the year. It's important to note that while the dip in the graph may suggest a larger disparity, the actual difference is just one paper. Given the remaining time for publication, the total number of publications is expected to far exceed those published in 2024. While the citation, as seen in Figure 4, declines over time, this is natural since citation is accumulated through time, and as the number of publications rises, the average number of citations decreases. However, there were 3 noticeable spikes in the average annual number of citations in the years 2011, 2014, and 2018. Those dates correspond to the top 3 most cited documents in the field [40,41,42].

Observing the most relevant authors through Lotka’s Law, there is a classical and clear power distribution with a very small number of authors present in the scientific space that contributed several papers. The top five contributors, Li C. (5 articles), Liu M. (4 articles), Wang Y. (4 articles), Qu J. (4 articles), and Zhang Y. (4 articles), all started publishing in this particular field of study in 2021 or later. All of them also contributed in 2025. When looking at the authors’ local impact based on the h-index and additionally taking into account their relevance, Li C. is the most prominent contributor, tightly followed by the previously mentioned remaining four contributors. Fujian Medical University (54 articles), Guangxi Medical University (54 articles), The Second Hospital of Xian Jiaotong University (46 articles), Capital Medical University (45 articles), Tongji Medical College of Huazhong University of Science and Technology (39 articles). The production over time for the outlined affiliation is starting to grow from 2021 and is noticeably growing each year.

Productive countries and source titles

Looking at Figure 5, it is clear that China and the United States of America are dominating in this field of study, with the former far outcompeting the other areas and the latter having a far more international presence in relation to cumulative production. Also noteworthy is that Germany has far more Multiple Country Publications (MCP) than Single Country Publications (SCP). Looking at the countries’ production over time, there was very little publication from 2011 until 2020. After 2020 the production in China and the United States of America increased exponentially. Looking at the citations based on the country The United States of America and China again dominate the research field and are followed by the United Kingdom and Canada.

Table 2 shows that the rank of countries' productivity in PSM-related research corresponds well with the rankings of countries in overall productivity across all disciplines, as well as in medicine, AI, medicine, and Statistics and probability. That might indicate that systemic national factors—like consistent R&D funding, shared infrastructure, high-quality education, research capacity and supportive policies—uniformly drive research productivity across academic disciplines/topics shown in Table 1. It might also signal the co-operation in symbiotic multidisciplinary research across medicine, AI and statistic focusing on PSM.

As seen in Figure 5, there are more than 30 core source titles publishing research on PSM and AI, according to Bradford's law [44]. Among them, Frontiers in Oncology, Frontiers in Cardiovascular Medicine, BMC Infectious Diseases, Frontiers in Public Health, Journal of Clinical Medicine, and JMIR Medical Informatics published five or more papers, and there were 14 journals publishing more than three papers. When considering the local impact of the sources in relation to the h-index, the above journals are complemented by Cancers, Frontiers in Pharmacology, and BMC Public Health. The publication rate for the top five journals has increased steadily over time, with Frontiers in Oncology and JMIR Medical Informatics at the forefront since 2018.

Figure 5. Presentation of core sources using Bradford’s Law.

Figure 5. Presentation of core sources using Bradford’s Law.

Table 2. shows the most prolific core journals. Most of them are ranked in the first quarter of their respective research areas indicating the high quality of combined research on AI and PSM. Most of the top core journals are from the medical research areas with the exception of JMIR which is categorized in the health informatics category. However, taking into account the large number of medical journals ,compared to the much smaller number of health informatics journals this combination of medical and informatics research areas might indicate the symbiotic nature of AI and PSM research.

Table 2 enumerates the most prolific core journals contributing to research on artificial intelligence (AI) and patient safety management (PSM). The majority of these journals are ranked into the first quartile within their respective subject categories, underscoring the high scholarly impact and methodological rigor characterizing this interdisciplinary domain. Predominantly, these journals are situated within medical research fields, with the notable exception of the Journal of Medical Internet Research (JMIR), which is classified under health informatics. Given the disproportionate representation of medical journals relative to the comparatively limited number of health informatics journals, this distribution might suggests a structural interdependence between medical and informatics research. highlighting the possible synergistic nature of AI and PSM scholarship, wherein Al methodologies and clinical applications converge to advanced and more innovative health care.

More prolific themes

The results of the STA are presented in Figure 6 and Table 3 and Table 4. Following Zipf’s Law [34], bibliometric mapping was applied to all keywords appearing in four or more publications, yielding a landscape of 45 author keywords grouped into four distinct clusters (Figure 1). Implementation of steps 2–6 in Algorithm 1 generated four thematic categories, 12 association sub-networks, and 28 high-impact publications (Table 2). A synthesis of these 28 publications is provided in Table 2.

The results of the deductive thematic analysis based on the authors' keywords are shown in Table 3. As we can see, the analysis did not reveal in detail which machine learning methods have actually been used because authors mainly used the generic keyword Machine learning. Consequently, we performed an additional analysis of terms found in publication abstracts. This analysis revealed that Boosting (n=30), Support vector machines (n=20), and Nearest neighbours (n=20). Random forests (n=18) and Decision trees (n=15) were the most popular methods. In addition, SHAP was used in 25 publications to make black box machine learning outputs explainable. The most popular applications were prognosis, prediction, decision support, and risk assessment. The most popular diagnoses were related to chronic diseases and cancer.

Table 3. Deductive thematic analysis - Ten most prolific AI algorithms and methods, medical applications, and diagnosis.

AI algorithms and services	n	Medical applications	n	Diagnosis	N
Machine learning	66	Prognosis	11	heart diseases	12
Artificial intelligence	23	Mortality	10	atrial fibrillation	6
Deep learning	11	Nomogram	5	covid-19	6
Decision tree	5	Prediction model	11	gastric cancer	6
Random forest	6	Diagnosis	4	hepatocellular carcinoma	6
Natural language processing	5	Decision support	6	Stroke	3
Big data	3	Risk assessment	14	Kidney diseases	11
SHAP/Explainable AI	2	epidemiology	5	Diabetes	9
Missing data imputation	1	Survival analysis	11	Cardiovascular diseases	18
Feature selection	2	Health services	6	Coronary diseases	7

Table 3. Deductive thematic analysis - Ten most prolific AI algorithms and methods, medical applications, and diagnosis.

AI algorithms and services	n	Medical applications	n	Diagnosis	N
Machine learning	66	Prognosis	11	heart diseases	12
Artificial intelligence	23	Mortality	10	atrial fibrillation	6
Deep learning	11	Nomogram	5	covid-19	6
Decision tree	5	Prediction model	11	gastric cancer	6
Random forest	6	Diagnosis	4	hepatocellular carcinoma	6
Natural language processing	5	Decision support	6	Stroke	3
Big data	3	Risk assessment	14	Kidney diseases	11
SHAP/Explainable AI	2	epidemiology	5	Diabetes	9
Missing data imputation	1	Survival analysis	11	Cardiovascular diseases	18
Feature selection	2	Health services	6	Coronary diseases	7

Prolific term and topic trend analysis

Focusing on the terms, there is a clear trend developing where AI, machine learning, and deep learning are being introduced and considered for application since 2022, as seen in Figure 7. The symbiosis seems to evolve first from association of PSM and machine learning, through introduction of AI and deep learning and finally of concrete applications of both approaches in concrete medical application (cancer and sepsis).

Additional insight and confirmation of the symbiotic association of PDA and AI, can be obtained from the thematic map presented in Figure 8, where symbiosis of AI and PSM in chronic diseases presents as a motor theme in the upper right quadrant, while also being present in the lower left quadrant, as an emerging approach in various medical subfields..

4. Discussion

This study systematically examined the evolving intersection between AI and PSM in medical research and identified a rapidly expanding methodological domain characterised by bidirectional methodological enhancement. Evidence from the thematic and bibliometric synthesis indicates that the combined use of AI and PSM has accelerated sharply since 2020, coinciding with the broader adoption of large-scale electronic health records, registry data, and digital health infrastructures. The findings suggest that the integration of AI into PSM workflows, and conversely the incorporation of PSM into AI-driven medical studies, is emerging as a critical methodological approach for improving causal inference, mitigating confounding bias, and advancing real-world digital medical evidence/big data analysis, aggregation and synthesisis.

A key observation of this knowledge synthesis study is that AI techniques are increasingly being adopted to augment the PSM pipeline. Across included studies, machine learning and deep learning algorithms were employed to support covariate selection, estimate propensity scores in high-dimensional settings, identify non-linear and interaction effects, and improve pre-processing of structured and unstructured clinical data. Natural language processing models, particularly domain BERT, ChatGPT and Gemini adoptions, were used to extract clinically meaningful covariates from free-text clinical notes, thereby enhancing the completeness and quality of matching sets. The results demonstrate that AI-enhanced PSM is particularly valuable in complex observational datasets, where classical regression-based estimation techniques may be limited by covariate sparsity, non-linearity, or multidimensional feature relationships.

Importantly, the inverse symbiotic relationship—PSM used to strengthen AI applications—was also evident. In a substantial proportion of studies, PSM was applied prior to machine learning model training to reduce baseline imbalances, control confounding, and ensure that subsequent AI-based prediction models were trained on balanced and comparable cohorts. This sequencing reflects a methodological recognition that AI systems, when trained on imbalanced observational data, risk amplifying bias and generating unreliable or unfair predictions. PSM therefore functions as a foundational causal-inference layer that improves the validity, interpretability, and clinical acceptability of AI-based models. Notably, several studies coupled PSM with fairness-aware machine learning frameworks, suggesting an emerging alignment between causal inference methodology and ethical AI development.

Comparison with the broader literature confirms alignment with current methodological transitions in medical data science, including the adoption of target trial emulation, doubly robust learning methods, and explainable AI techniques. The increasing appearance of SHAP-based model interpretability approaches in the analysed corpus reflects a growing acknowledgement that explainability is essential when AI is deployed in clinical decision-making environments. Furthermore, the results correspond with regulatory trends emphasising transparency, fairness, and accountability in medical AI systems, as seen in evolving FDA and EMA guidance on machine learning-based medical technologies.

Several practical implications arise from these findings. First, the observed symbiosis between AI and PSM underscores the value of hybrid analytical pipelines for generating robust evidence from real-world clinical data. Such approaches may support precision medicine, enable reliable risk stratification, and enhance translational research by strengthening causal assumptions in observational studies. Second, the combined use of AI and PSM holds promise for hospitals and healthcare systems seeking to leverage routinely collected data for predictive modelling, outcome evaluation, and clinical decision support. Finally, the results highlight a methodological trajectory that can inform future clinical research training and the development of multi-disciplinary analytical frameworks integrating biostatistics, machine learning, and clinical epidemiology.

Despite substantial momentum, several gaps and challenges remain. The reviewed literature indicates limited methodological standardisation; AI-enhanced PSM implementations vary considerably across studies, and benchmarking frameworks for assessing performance and bias reduction are not yet mature. Although causal machine learning techniques were identified, their use remained limited relative to classical PSM frameworks, signalling a need for broader adoption of causal representation learning, counterfactual inference, and heterogeneous treatment-effect modelling. Furthermore, dependency on cross-sectional or static datasets limits applicability to time-varying clinical processes. Dynamic PSM and online learning approaches capable of updating matches as new data accumulate remain under-developed. Ethical considerations—including privacy protection, bias mitigation, and responsible model governance—were acknowledged only sporadically, reflecting another key research frontier.

This study also has methodological limitations. Although Scopus provides broad multidisciplinary coverage, relevant studies indexed exclusively in other databases may have been excluded. Keyword-based bibliometric retrieval may not fully capture studies using alternative terminologies for matching or AI methods. Author keyword heterogeneity limited granular quantification of algorithm-specific trends, necessitating complementary abstract-level analysis. Additionally, synthetic thematic analysis provides interpretive depth, but does not replicate the exhaustive appraisal of systematic reviews. Nonetheless, the triangulation of bibliometric mapping with thematic synthesis strengthens the credibility of the findings.

Future research should prioritise automated and scalable AI frameworks for propensity score estimation, integration of causal machine learning systems, and development of shared benchmark datasets for evaluating hybrid AI-PSM pipelines. Increased adoption of longitudinal and multi-modal clinical data, federated architectures, and transparent model governance protocols will be essential. As AI-integrated PSM methods advance, rigorous validation in prospective clinical settings and alignment with ethical and regulatory frameworks will remain critical to ensuring robust, fair, and clinically meaningful implementation.

5. Conclusion

In conclusion, this study reviews and synthetically analyses the current scientific literature in order to present the current state of integration and usage of AI in the domain of propensity score matching. This synthetic knowledge review provided several clear research themes, which demonstrate the bi-directional and complementary nature of AI in propensity score matching and integration in condition prognosis and diagnostics. The integration of AI in propensity score matching proved integral in the improvement of diagnosis accuracy and the understanding of reasoning and helped overcome the limitations that are intrinsic to propensity score measurements. On the other hand, critical shortcomings were outlined, for instance, the need for advanced techniques for flawless integration of AI into propensity score matching, the need for clarity and interpretability of AI models, robust solutions for dealing with confounding unmeasured factors, and the need for standardization of evaluation. The other major problem, which is still one of the greatest obstacles when it comes to AI usage in medicine, is the ethical concern. Future research will have to focus on those critical points, the integration of AI, adding robustness, integrating, and formulating ethical concepts that will enable AI to be safe and acceptable. The implementation and verification of those new approaches will have to be performed on different longitudinal medical datasets since the synergistic usage of AI in propensity score matching provides new avenues and progress in the field of precise medicine. The main goal must be to improve decision-making in healthcare and the optimization of patient outcomes.