Artificial Intelligence in Healthcare: 2025 Year in Review

The 2025 review of healthcare AI research reveals a near-doubling of publication volume and a significant maturation of the field, characterized by a shift from classical machine learning and text-only models toward the rapid adoption of multimodal foundation models that better reflect the complexity of real-world clinical practice.

The Gist

Imagine the world of healthcare as a massive, bustling library. For years, researchers have been trying to build a super-smart librarian (Artificial Intelligence) who can help doctors find the right book, diagnose a patient, or organize the shelves.

This paper is a year-end report card for that library, specifically looking at what happened in 2025. Here is the story of that year, told simply:

1. The Library is Exploding in Size

In 2025, the number of new research papers about AI in healthcare doubled compared to the previous year.

• The Analogy: If 2024 was like a small town library adding a few new books, 2025 was like the library suddenly building a massive new wing and filling it with thousands of new volumes overnight.
• The Takeaway: Everyone is excited and working hard. The field is growing faster than ever.

2. From "Toy Models" to "Real Tools"

The authors didn't just count every paper; they filtered them to find the "mature" ones.

• The Analogy: Think of early AI research as kids playing with toy cars in the sandbox. They were fun, but they couldn't actually drive you to the store. In 2025, we started seeing real, working cars. These are studies where the AI actually helps a doctor make a decision or improves patient care in a real hospital, not just in a computer simulation.
• The Takeaway: The field is maturing. We are moving from "what if?" to "this works."

3. The Shift: From "One-Track Minds" to "Super-Brains"

This is the biggest story of 2025.

• The Old Way (Text-Only): For a while, the smartest AI tools were like super-fast readers. They could read medical notes and write summaries, but they couldn't look at an X-ray or listen to a heart sound. They were "text-only" specialists.
• The New Way (Multimodal): In 2025, we saw a huge surge in Multimodal Foundation Models.
  • The Analogy: Imagine upgrading your librarian from someone who only reads books to a super-genius who can read a book, look at a painting, listen to a song, and understand a chart all at the same time.
  • Why it matters: Real life is messy. A doctor doesn't just read a chart; they look at the patient's face, listen to their breathing, and review their X-rays. These new "Super-Brains" can do all of that together, just like a human doctor does.

4. Who is Leading the Race?

• The Image Champions: Radiology and Imaging (X-rays, MRIs, CT scans) are still the kings of the hill. They have the most research because pictures are easier for computers to analyze than messy human language.
• The New Challengers: However, the "Super-Brains" (Multimodal models) are quickly taking over other departments like Surgery, Oncology (Cancer), and Eye Care. They are learning to combine images with text to make better diagnoses.
• The Bureaucrats: A surprising amount of research is also happening in Administration and Education. AI is being used to write patient letters, create exam questions for students, and organize hospital paperwork.

5. The "Old Guard" is Stepping Back

For the first time, the use of older, simpler AI models (called "Classical Machine Learning") actually dropped.

• The Analogy: It's like the invention of the smartphone. Once you have a device that can do everything (take photos, play music, navigate, and browse the web), you stop using the old flip phone, the MP3 player, and the paper map separately.
• The Takeaway: Researchers are ditching the simple tools for the powerful, all-in-one "Super-Brains" because they can handle complex medical problems better.

6. What's Next?

The paper concludes that we are at a tipping point.

• The Future: We are moving toward a future where AI doesn't just sit in a computer waiting for a question. It's becoming an active partner that can see, hear, read, and think across different types of data.
• The Catch: While the technology is amazing, we still need to make sure these tools are safe, tested, and actually helpful in real hospitals before they become the standard.

Summary in One Sentence

In 2025, healthcare AI stopped playing with simple text-based toys and started building multilingual, multi-sensory super-brains that can look at an X-ray, read a patient's history, and listen to a heartbeat all at once, bringing us one giant step closer to AI that truly works like a human doctor.

Technical Summary

1. Problem Statement

The field of Artificial Intelligence (AI) in healthcare is experiencing exponential growth, driven by breakthroughs in model architecture and data availability. However, the landscape is shifting rapidly from exploratory research to real-world application.

• The Challenge: There is a need to track and quantify the maturation of the field, specifically the transition from traditional Machine Learning (ML) and text-only Large Language Models (LLMs) to Multimodal Foundation Models (FMs) that reflect the diverse nature of clinical data (images, text, audio, tabular).
• The Gap: Previous reviews focused heavily on text-based models. There is a lack of comprehensive, year-over-year analysis regarding the adoption of foundation models across different clinical specialties, data modalities, and the "maturity" of these studies (i.e., whether they offer actionable clinical impact).

2. Methodology

The authors conducted a systematic review of literature published in 2025, employing a hybrid approach of automated classification and manual curation.

• Data Source: PubMed database search restricted to English-language, human-subject research.
• Search Query: Terms "machine learning" OR "artificial intelligence" AND "2025".
• Inclusion/Exclusion Criteria:
  • Excluded: Pre-prints, animal studies, robotic surgery studies (without AI), systematic reviews, and non-English publications.
  • Maturity Filter: A BERT-based deep learning classifier (pre-trained on manually labeled data) was used to screen publications. A study was classified as "Mature" if its output had a direct, actionable impact on patient care (e.g., providing information to providers or automated systems). Systematic reviews were excluded from the "mature" count.
• Manual Annotation: Five reviewers manually annotated the remaining "mature" publications (n=2,966) for:
  • Healthcare Specialty: 25 categories (22 clinical specialties + General, Administrative, Education).
  • Data Type: Image, Text, Tabular, Voice/Audio.
  • Model Type: Traditional Deep Learning (DL), Classical ML, AI General, Statistical, NLP, LLM, Large Vision Model (FM-LVM), and Multimodal Model (FM-MM).
• Temporal Scope: The study covers the year 2025, with comparative data from 2019–2024 to identify trends.

3. Key Contributions

• Quantitative Shift Analysis: The study provides the first comprehensive data showing the decline of classical ML and the surge of Multimodal Foundation Models in 2025.
• Maturity Classification: It introduces a rigorous, reproducible method for distinguishing between exploratory AI research and "mature" research with direct clinical applicability.
• Specialty-Specific Insights: It breaks down AI adoption by specialty, highlighting that while Imaging leads, other fields (Surgery, Oncology, Head & Neck) are rapidly adopting foundation models.
• Multimodal Trend Identification: It documents the critical inflection point where research is moving beyond text-only LLMs to models capable of processing images, text, and audio simultaneously.

4. Key Results

Publication Volume & Maturity

• Total Publications: 49,394 (nearly double the 28,180 from 2024).
• Mature Publications: 3,366 identified as mature; 2,966 included in final analysis after exclusions (up from 1,946 in 2024).

Model Architecture Trends

• Dominance of Foundation Models: Traditional text-based LLMs remain the most used model type (1,019 publications), but Multimodal Foundation Models (FM-MM) saw a massive surge from 25 (2024) to 144 (2025).
• Deep Learning: Traditional DL models increased substantially to 910 publications.
• Decline of Classical ML: For the first time, publications using classical Machine Learning models declined (173 publications).
• Proprietary "Black Boxes": A significant portion of studies fell into the "AI General" category, often involving proprietary models where specific architectures were not disclosed.

Data Modalities

• Image: Remains the dominant data type (53.9%).
• Text: Second most common (38.2%).
• Audio: Notable increase to 1.2%, correlating with the rise of multimodal models.
• Multimodality: 94 publications (3.1%) combined data types (e.g., Image + Text), driven by the adoption of foundation models.

Specialty Distribution (Mature Publications)

• Top Specialties:
  1. Imaging/Radiology: 976 publications (Dominant leader).
  2. Administrative: 277.
  3. General: 251.
  4. Head and Neck: 185.
  5. Education: 181.
• Foundation Model Adoption: Imaging (110), Head and Neck (92), Surgery (64), Oncology (55), and Ophthalmology (49) were the leading clinical specialties utilizing foundation models. Administrative and Education tasks were heavily driven by LLMs.

5. Significance and Conclusion

• Field Maturation: The doubling of mature publications and the shift away from classical ML indicate the field is moving from "proof-of-concept" to "real-world evaluation."
• Inflection Point: The surge in Multimodal Foundation Models suggests AI systems are beginning to mirror the complexity of real-world clinical practice, which rarely relies on a single data type.
• Clinical Integration: While Imaging remains the primary driver, the expansion of foundation models into Surgery, Oncology, and Pathology signals a broader clinical integration.
• Future Outlook: The authors predict continued growth in the evaluation of proprietary AI tools, the implementation of Agentic AI (autonomous agents), and the expansion of multimodal applications. The study highlights a need for resources to foster research in specialties currently lagging behind Imaging.

Limitations: The study is limited to the PubMed database (potentially missing non-indexed research) and relies on manual annotation which, while more accurate than LLMs for this task, introduces potential subjectivity.