Introduction
Trajectory modeling—the ability to predict and understand a patient's health journey over time—is a cornerstone of personalized medicine. However, biomedical data is inherently complex, often capturing a patient's status across disparate timescales, from real-time physiological waveforms in an ICU to longitudinal electronic health records (EHR) spanning decades. Traditional statistical methods, such as ARIMA, and standard machine learning approaches frequently struggle to capture the non-linear dependencies and multimodal nature of these medical histories.
The emergence of generative artificial intelligence (GenAI) offers a transformative approach. Unlike traditional "discriminative" models that learn to draw boundaries between data points, GenAI learns the underlying distribution of the data itself. This allows for more sophisticated modeling of complex health patterns. Despite this potential, a significant "translation gap" exists between rapid AI development and practical clinical application. This scoping review of 155 peer-reviewed articles aims to bridge this gap, providing a roadmap for clinicians and researchers to navigate the evolving landscape of GenAI in healthcare.
Methodology
Following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines, the authors screened over 3,500 papers to identify 155 high-quality studies published since 2000. These studies were categorized into four primary data modalities:
- Structured Text and Tabular Data: Including EHR records, laboratory results, and vital signs.
- Unstructured Text: Such as clinical narratives, radiology reports, and discharge summaries.
- Medical Imaging and Waveforms: Encompassing longitudinal MRI/CT series, EEG, and ECG data.
- Genetics and Multi-omics: Covering DNA sequences and single-cell transcriptomics.
The review evaluates these applications through the lens of four core generative architectures: Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Diffusion Models, and Foundation Models (including Large Language Models or LLMs).
Key Insights: Beyond Prediction
The application of GenAI in medicine extends far beyond simple forecasting. Several key themes emerged from the review:
Synthetic Data and Privacy
One of the most immediate impacts of GenAI is data augmentation. By generating "synthetic" patient records that maintain the statistical properties of real data without containing identifiable information, researchers can bypass the strict privacy hurdles of HIPAA and GDPR. This is particularly vital for research into rare diseases, where data scarcity often halts progress.
Architecture Trade-offs
The review highlights a clear hierarchy of model performance. While GANs and Transformers often provide the highest accuracy, they are computationally intensive and lack interpretability. In contrast, VAEs offer a "latent space" that is more accessible for clinicians to use in downstream linear modeling, providing a better balance between performance and usability.
Counterfactual Estimation
By learning the underlying data distribution, GenAI enables "what-if" scenarios. This allows clinicians to estimate counterfactuals—predicting how a patient’s trajectory might have changed under different treatment paths that were never actually taken.
The Rise of Specialized Models
The field is shifting toward specialized foundation models. The review distinguishes between Clinical Language Models (CLaMs), which generate text, and Foundation models for EMRs (FEMRs), which produce "patient embeddings"—mathematical representations of a patient's entire medical history that can be used for various predictive tasks.
Results and Clinical Performance
The review found that deep learning and generative methods consistently outperform traditional statistical benchmarks for medium- and long-range forecasting. For instance, medical LLMs like GatorTron, trained on over 90 billion words, have shown superior performance in clinical concept extraction compared to smaller, general-purpose models.
Furthermore, the paper introduces a systematic framework to help researchers identify the most suitable AI methods based on their specific data type and clinical task. This decision-making tool is designed to make these complex algorithms more accessible to non-developers.
Limitations and Future Directions
Despite the promise of GenAI, several hurdles remain:
- The "Black Box" Problem: The lack of transparency in GenAI models is a significant barrier to clinical adoption. Doctors require interpretable results to make informed life-and-death decisions.
- Hallucinations: LLMs can generate factually incorrect but plausible-sounding medical advice. In healthcare, the tolerance for such errors is near zero.
- The Outpatient Gap: Most current research focuses on high-fidelity, high-frequency inpatient data. Future models must adapt to the sparse and irregular nature of outpatient data.
- Advanced Imaging Trajectories: The authors suggest that emerging video-generation technologies could be adapted to forecast the physical progression of diseases, such as tumor growth or neurodegeneration, in longitudinal imaging.
Conclusion
Generative AI represents a paradigm shift in how we model and understand time-varying biomedical data. By moving from simple classification to understanding the generative processes of health and disease, these models offer unprecedented opportunities for personalized care. However, realizing this potential requires a concerted effort to improve model interpretability, ensure data quality, and bridge the gap between computational innovation and bedside application.
Summary updated