💻 computer science

Mapping the Evolution of Artificial Intelligence in Healthcare, 2017–2025: An Integrated Bibliometric and Latent Dirichlet Allocation Topic Modelling Analysis

This study integrates standard bibliometric analysis with Latent Dirichlet Allocation (LDA) topic modelling on a curated corpus of 2,631 healthcare AI documents from 2017 to 2025 to overcome keyword limitations, revealing a maturation trajectory from deployment-ready applications to governance-dependent frontiers like large language models and explainable AI.

Original authors: Piyusha Ambradkar, Vandana Tandon Khanna.

Published 2026-09-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Piyusha Ambradkar, Vandana Tandon Khanna.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The landscape of modern medicine is being quietly reshaped by a powerful new force: artificial intelligence. For decades, computers have helped doctors organize records and run basic calculations, but today's systems are learning to see patterns in medical images, predict patient outcomes, and even draft clinical notes. This shift is happening so fast that it is difficult for hospital leaders and researchers to keep a clear picture of where the field is going. They need a way to map the vast ocean of new research, distinguishing between ideas that are ready for real-world use and those that are still just experiments. To do this, they must look at thousands of scientific papers, but simply counting words or relying on the labels authors choose to describe their work often misses the deeper meaning. Different scientists might use different words for the same idea, or the same words for different things, creating a fog that hides the true trends.

A team of researchers at the K J Somaiya Institute of Management in Mumbai decided to clear that fog by combining two different ways of looking at the data. They gathered a massive collection of 2,631 scientific articles published between 2017 and 2025, focusing specifically on how artificial intelligence is applied to healthcare. Instead of just reading the titles and keywords, they used a computer method that reads the titles and abstracts of the articles to find hidden themes, much like a librarian who listens to the content of a conversation rather than just reading the subject line on a letter. By pairing this deep reading with a standard analysis of who is citing whom and which journals are most active, they created a detailed map of the field's evolution. Their work reveals not only which technologies are maturing but also which emerging ideas are flying under the radar of traditional research tracking.

The researchers found that the field has grown explosively, with the number of publications rising from about 87 in 2017 to 731 in 2025. This surge was particularly sharp around 2020, driven by the urgent need for tools to help diagnose and track the COVID-19 pandemic. When they analyzed the most cited papers, they saw that the early focus was heavily on using computers to read medical images, such as X-rays and scans, and on connecting wearable devices to hospital systems. These areas have now reached a point of maturity, meaning the core technical problems are largely solved, and current research is focused on refining these tools for specific diseases like heart conditions or diabetes.

However, the study's most significant discovery came from the deep reading method, which uncovered themes that standard keyword searches completely missed. While traditional analysis showed a clear picture of imaging and wearable devices, it failed to spot two rapidly growing areas: the use of large language models to write clinical notes and the urgent need for rules to explain how artificial intelligence makes its decisions. The computer analysis identified that these topics are exploding in popularity, yet they were invisible in the standard keyword maps because scientists were using a wide variety of terms to describe them. This suggests that the field is moving beyond just building smart tools to worrying about how to govern them and how to integrate them into the daily paperwork of hospitals.

The researchers organized these findings into a clear timeline of readiness. They determined that technologies for reading medical images and monitoring patients with sensors are ready to be deployed in hospitals right now. Other areas, such as using artificial intelligence to help doctors write patient records or to protect patient privacy while sharing data across different hospitals, are still in a phase of rapid growth. These newer fields are not yet ready for widespread use because they require new laws and safety frameworks to be put in place first. The study concludes that for hospital leaders and investors, the most valuable insight is not just what is working today, but recognizing which emerging technologies are on the horizon and require preparation before they can be safely adopted. By using this combined approach of counting citations and reading for hidden themes, the team provided a reliable way to see the future of medical technology before it becomes obvious to everyone else.

Technical Summary: Mapping the Evolution of Artificial Intelligence in Healthcare (2017–2025)

Problem Statement
The rapid transformation of artificial intelligence (AI) in healthcare—from rule-based systems to deep learning and large language models (LLMs)—has created a challenge for researchers and decision-makers in accurately mapping the field's current state and trajectory. While standard bibliometric analyses offer macro-level insights, they rely heavily on keyword co-occurrence. This approach is limited by author keyword heterogeneity, where different terms describe the same concept or identical terms describe different concepts, introducing noise and failing to capture semantic relationships across diverse research traditions. Consequently, emerging semantic themes, particularly those utilizing novel terminology (e.g., "explainable AI" or specific LLM applications), may be under-represented or invisible in traditional keyword-based maps.

Methodology
This study employs an integrated methodology combining a standard seven-phase bibliometric pipeline with computational Latent Dirichlet Allocation (LDA) topic modelling.

Corpus Construction: A dual-database corpus was extracted from OpenAlex and PubMed, covering a parent "cloud-healthcare" dataset from 2010–2025. After filtering for AI-related terms (e.g., "machine learning," "federated learning," "LLM") and restricting the timeframe to 2017–2025, a sub-corpus of 2,631 documents was finalized following PRISMA 2020 screening (deduplication, language, and abstract availability filters).
Bibliometric Analysis: Standard methods were applied using the bibliometrix R package, including Bradford's and Lotka's Laws, citation analysis, co-authorship/co-citation networks, bibliographic coupling, and thematic mapping across three periods (2017–2019, 2020–2022, 2023–2025). Multiple Correspondence Analysis (MCA) was used to visualize conceptual structures.
LDA Topic Modelling: The topicmodels R package was used to model the combined title and abstract text. Pre-processing involved lowercasing, stop-word removal (including domain-specific high-frequency non-discriminative terms), and stemming. The optimal number of topics ( $k=8$ ) was determined by minimizing held-out perplexity. Gibbs sampling (2,000 iterations) was employed. Topic labels were assigned via qualitative review of the top 20 terms per topic by two independent coders, achieving substantial inter-rater agreement (Cohen's kappa = 0.84).

Key Results

Growth and Trends: The sub-corpus exhibited explosive growth, with annual publications rising from ~87 in 2017 to 731 in 2025 (CAGR ~30%). A distinct inflection point occurred in 2020–2021, driven by AI-based COVID-19 diagnostics and federated surveillance models.
Bibliometric Findings: Bradford's Law identified IEEE Access, IEEE Internet of Things Journal, and Future Generation Computer Systems as core journals. Lotka's Law indicated a slightly more concentrated core researcher community ( $\beta=2.58$ ) compared to the parent dataset.
LDA Topic Discovery: Eight coherent latent topics were identified:
1. Deep Learning: Medical Imaging (18.4%, Maturing)
2. NLP for Clinical Text (13.2%, Growing)
3. Federated Learning & Privacy (11.8%, Rapid Growth)
4. IoT-integrated CDSS (16.7%, Maturing)
5. Drug Discovery & Genomics (8.9%, Stable)
6. LLMs for Clinical Documentation (9.3%, Explosive growth post-2023)
7. Explainable AI (XAI) Governance (10.6%, Fast Growth)
8. AI Remote Patient Monitoring (11.1%, Growing)
Integration and Divergence: Cross-referencing LDA with keyword co-occurrence revealed that while topics like Federated Learning aligned well, LLMs for Clinical Documentation and XAI Governance were almost entirely absent from standard keyword clusters. This confirms that LDA successfully detected nascent research fronts that keyword analysis missed.
Trajectory Analysis: Topics were categorized by maturity:
- Deployment-Ready: Deep learning imaging and IoT-based CDSS (plateauing growth, focus on validation).
- Integration-Ready: NLP for clinical text, Federated learning, and Remote monitoring (proven tech, requiring infrastructure integration).
- Investment-Watch Frontiers: LLMs and XAI governance (rapidly accelerating, requiring regulatory and governance frameworks before scale).

Significance and Contributions
The paper claims three primary contributions:

Methodological Advance: It demonstrates that integrating LDA with standard bibliometric pipelines offers superior semantic discovery compared to keyword analysis alone, specifically in identifying emerging themes (like LLMs and XAI) that are not yet standardized in author-supplied keywords.
Technology-Readiness Taxonomy: The study provides a replicable framework for classifying AI applications in healthcare based on their developmental phase (maturing, growing, or frontier), directly applicable to health technology investment and horizon-scanning decisions.
Governance Dimension: The analysis highlights the emergence of a distinct "governance and ethics" dimension (encompassing XAI, bias, fairness, and privacy) as a structuring force in the literature since 2020, a nuance only partially visible in traditional keyword maps but clearly defined through the integrated approach.

Limitations
The authors note that LDA results are sensitive to preprocessing and topic count selection. The study is restricted to English-language publications, potentially under-representing significant Chinese-language AI healthcare research. Citation counts for recent years are incomplete due to accumulation lags. Finally, the technology-readiness taxonomy is derived from publication patterns rather than direct evaluation of deployed clinical systems.

Technical Summary: Mapping the Evolution of Artificial Intelligence in Healthcare (2017–2025)

More like this