Implementierung von künstlicher Intelligenz (KI) im Gesundheitswesen: Historische Entwicklung, aktuelle Technologien und Herausforderungen

Jill von Conta, Merlin Engelke, Fin Hendrik Bahnsen, Amin Dada, Elisabeth Liebert, Felix Nensa, Hyeonhoon Lee, Anke DiehlBundesgesundheitsblatt 2025 · DOI

Abstract

The historical development of artificial intelligence (AI) in healthcare since the 1960s shows a transformation from simple rule-based systems to complex, data-driven approaches. Early applications focused on decision support, while innovative systems use neural networks and machine learning to recognise patterns in large datasets. The integration of AI technologies in medicine has given rise to diverse areas of application, which can be categorized into preventive AI, diagnostic AI, AI-assisted therapeutic, and administrative AI. Preventive AI analyses risk factors to enable early interventions, while diagnostic AI contributes to faster and more accurate diagnoses. AI-assisted therapy supports individualized treatments, such as personalized medication. Administrative AI optimizes processes such as appointment scheduling, resource management and billing. Despite their potential, AI systems face challenges. These include the fragmentation of health data, a lack of standardisation, data protection concerns and algorithmic biases. The establishment of interoperable data infrastructures and the development of ethical guidelines are crucial to overcoming these hurdles. Future trends include the further development of foundation models, the integration of structured and unstructured data and greater personalisation in medicine. In the long term, AI can improve the quality and efficiency of healthcare. However, this requires close co-operation between research, industry and politics in order to ensure safe and sustainable implementation.

Introduction: The Implementation Gap in Healthcare AI

Despite the rapid advancements in artificial intelligence, particularly with the emergence of Large Language Models (LLMs), a significant "implementation gap" persists in clinical settings. While AI has demonstrated transformative potential in medical research, its transition into everyday clinical practice remains sluggish. This paper explores the systemic hurdles—ranging from fragmented data silos to complex regulatory frameworks—and proposes architectural solutions to bridge this gap.

Historical Evolution: From Expert Systems to Foundation Models

The journey of AI in medicine began in the mid-20th century. Early systems were primarily rule-based, designed to support specific diagnostic tasks like the MYCIN system for infectious diseases. Over the decades, the field shifted toward data-driven approaches, fueled by the standardization of medical imaging (DICOM) and the rise of deep learning.

Timeline of AI development in medicine, from early rule-based systems to modern Foundation Models.

Key milestones include:

  • 1950s-70s: The birth of AI as a research field and early NLP programs like ELIZA (1966) and CASNET (1972).
  • 1980s: The introduction of the DICOM standard (1985), which revolutionized medical imaging analysis by enabling uniform processing.
  • 1990s-2000s: The development of Convolutional Neural Networks (CNNs) like LeNet-5 and the introduction of Deep Learning in medicine.
  • 2017-Present: The rise of Transformer models and Foundation Models like GPT-3, which enable multi-modal data synthesis and in-context learning without extensive retraining.

Categorizing AI in Healthcare: The Four Pillars

Modern medical AI applications can be categorized into four distinct domains, each addressing different aspects of the patient journey and hospital operations:

  1. Preventive AI: Focuses on risk stratification, screening, and early disease detection through the analysis of risk factors and longitudinal data.
  2. Diagnostic AI: Enhances pattern recognition in radiology and pathology, providing faster and more accurate diagnoses and survival prognoses.
  3. Therapeutic AI: Supports personalized medicine, including individualized medication planning and therapy optimization through data analysis.
  4. Administrative AI: Optimizes hospital workflows, resource management, and automated documentation (e.g., generating discharge summaries).

Overcoming the Bottleneck: The Smart Hospital Architecture

The primary obstacle to clinical AI is not algorithmic but infrastructural. To address this, the University Medicine Essen developed the "Smart Hospital Information Platform" (SHIP). This architecture utilizes an ETL (Extract, Transform, Load) pipeline to aggregate data from disparate clinical systems—such as PACS (imaging), RIS (radiology info), and LIS (lab info)—and standardizes it using the FHIR (Fast Healthcare Interoperability Resources) protocol.

The technical architecture of the Smart Hospital Information Platform (SHIP), highlighting the integration of clinical primary systems via FHIR and MLOps.

By early 2025, the SHIP platform had successfully aggregated over 1.6 billion data resources for more than 1.5 million patients, ingesting 600,000 new resources daily. This unified data layer enables robust MLOps (Machine Learning Operations), allowing models to be trained, deployed, and monitored directly at the point of care.

Key Challenges and Future Outlook

Despite the success of platforms like SHIP, several challenges remain:

  • Interoperability: The lack of standardized data formats across different institutions creates "monocentric" silos, preventing the generalizability of AI models.
  • Regulatory Gridlock: In Europe, developers must navigate the overlapping requirements of the EU AI Act, the Medical Device Regulation (MDR), and the GDPR.
  • Explainability and Bias: As models become more complex ("black boxes"), ensuring transparency and mitigating algorithmic bias is crucial to maintain clinical trust and safety.

The future of healthcare AI lies in Foundation Models—large-scale, multi-modal systems capable of processing text, images, and genomic data simultaneously. National initiatives like the European Health Data Space (EHDS) and the implementation of standardized electronic patient records (ePA) will be vital in breaking down institutional walls and fostering a truly data-driven healthcare ecosystem.

Conclusion

The integration of AI into healthcare is a transformative shift that requires more than just advanced algorithms. It demands a fundamental overhaul of hospital IT infrastructures, a commitment to data standardization, and a clear regulatory pathway. By bridging the implementation gap through platforms like SHIP, the medical community can move toward a future of higher quality, more efficient, and deeply personalized patient care.

Summary updated

Abstract

The historical development of artificial intelligence (AI) in healthcare since the 1960s shows a transformation from simple rule-based systems to complex, data-driven approaches. Early applications focused on decision support, while innovative systems use neural networks and machine learning to recognise patterns in large datasets. The integration of AI technologies in medicine has given rise to diverse areas of application, which can be categorized into preventive AI, diagnostic AI, AI-assisted therapeutic, and administrative AI. Preventive AI analyses risk factors to enable early interventions, while diagnostic AI contributes to faster and more accurate diagnoses. AI-assisted therapy supports individualized treatments, such as personalized medication. Administrative AI optimizes processes such as appointment scheduling, resource management and billing.Despite their potential, AI systems face challenges. These include the fragmentation of health data, a lack of standardisation, data protection concerns and algorithmic biases. The establishment of interoperable data infrastructures and the development of ethical guidelines are crucial to overcoming these hurdles. Future trends include the further development of foundation models (large AI models that are based on broad datasets and can be used in a variety of ways), the integration of structured and unstructured data and greater personalisation in medicine. In the long term, AI can improve the quality and efficiency of healthcare. However, this requires close co-operation between research, industry and politics in order to ensure safe and sustainable implementation.
BibTeX
@inproceedings{conta2025implementierung,
  author = {Jill von Conta and Merlin Engelke and Fin Hendrik Bahnsen and Amin Dada and Elisabeth Liebert and Felix Nensa and Hyeonhoon Lee and Anke Diehl},
  title = {{Implementierung von künstlicher Intelligenz (KI) im Gesundheitswesen: Historische Entwicklung, aktuelle Technologien und Herausforderungen}},
  booktitle = {Bundesgesundheitsblatt},
  year = {2025},
  doi = {10.1007/s00103-025-04086-6},
}
Figure