This paper presents a novel approach to bridging the gap between clinical expertise and artificial intelligence in healthcare. While AI is frequently used for medical risk assessment, it is often relegated to a "black box" assistant role, providing predictions without clear context or a way for clinicians to influence the underlying logic.
The authors propose a two-phase visual analytics framework designed to facilitate continuous collaboration between physicians and AI, specifically for analyzing sequelae—the long-term pathological conditions resulting from a prior disease or treatment (in this case, hormone-induced bone death).
TL;DR
The researchers developed a two-phase visualization system that allows clinicians to lead the discovery of medical risk factors and then iteratively refine AI predictive models. By combining retrospective data exploration with interpretable machine learning (SHAP), the system transforms AI from a static assistant into a transparent, collaborative partner for analyzing complex medical side effects.
What Problem Does This Paper Address?
Analyzing medical sequelae is notoriously difficult. Unlike acute diseases with clear, short-term symptoms, sequelae like hormone-related osteonecrosis often have subtle characteristics and complex, long-term development paths.
Clinicians face three main hurdles:
- Data Complexity: Electronic Health Records (EHRs) are noisy, unstructured, and difficult to navigate for long-term trends.
- Trust and Transparency: Standard AI models lack interpretability, making it hard for doctors to trust "black box" risk scores.
- Rigid Collaboration: Most systems don't allow clinicians to inject their domain knowledge—such as identifying specific high-risk patient cohorts or relevant clinical features—directly into the AI training loop.
Methods
The authors developed a framework divided into two distinct phases, as illustrated in their system overview:
- Phase I: Human-Led, AI-Assisted Retrospective Analysis. In this phase, clinicians use interactive visualizations to explore historical data. Key tools include a Cohort View (using Sankey diagrams to link demographics, diseases, and drugs), a Patient Projection View (using dimensionality reduction like t-SNE or UMAP to cluster similar patients), and a detailed Medical Event View (a timeline of medications and lab results).
- Phase II: AI-Mediated, Human-Reviewed Iterative Modeling. Here, the clinician takes the features and patient samples identified in Phase I to build a predictive model. The system uses a Random Forest classifier. Crucially, it employs SHAP (Shapley Additive Explanations) to visualize why the model made certain predictions, showing the relative importance of different clinical features.
Datasets
The system was built and evaluated using real-world, anonymized Electronic Health Records (EHR) provided by a local hospital. The data focused specifically on hormone-related medical records, including:
- Primary diagnoses and sequelae (e.g., osteonecrosis).
- Detailed medication orders (specifically glucocorticoids).
- Laboratory test results and examination reports.
- Demographic information (age, gender).
Key Insights and Ideas
- Role Transfer: The framework formalizes a "role transfer" between humans and AI. The human leads the initial discovery and hypothesis generation, while the AI takes over the heavy lifting of pattern recognition and prediction, with the human returning to the role of an expert reviewer.
- Interactive Sampling: Instead of relying on standard statistical matching (like PSM), which can be prone to bias in complex datasets, the system allows clinicians to use a "lasso" tool in the projection view to manually select and refine positive and negative patient samples for the model.
- Visualizing Feature Importance: By integrating SHAP values into a beeswarm plot (seen in Figure 1D), the system makes the model's logic transparent. A doctor can see, for instance, that a specific cumulative dose of a drug is the primary driver for a high-risk prediction.
- The Power of the Log: The Logs View (Figure 1E) acts as a "model archive," allowing clinicians to track how their choices (changing features or samples) affect model performance (Accuracy, F1 score, etc.) over multiple iterations.
Key Results
The system was evaluated through semi-structured interviews with seven medical experts (surgeons and data scientists).
- Qualitative Success: Experts found the visual metaphors (like the patient glyphs) intuitive and appreciated the ability to "see" the AI's reasoning. They noted that the system stimulated new perspectives on sequela analysis that traditional methods missed.
- Quantitative Performance: The prototype model achieved respectable metrics on the hormone-related dataset, such as an Accuracy of 0.73 and Precision of 0.86, while remaining fully interpretable to the users.
Limitations and Future Work
- EHR Bias: The authors acknowledge that inherent biases in hospital data entry can affect the accuracy and generalizability of the models.
- Abstraction Levels: The current system relies on binary "checks" and specific medication dosages; future versions could benefit from higher-level clinical abstractions.
- Scale of Evaluation: The study was limited to a small group of internal experts. The researchers plan to involve a larger pool of external users for more rigorous testing and to explore extending the framework to other domains like drug discovery and protein analysis.
Summary updated