MD-Manifold: A Medical-Distance-Based Representation Learning Approach for Medical Concept and Patient Representation

Shaodong Wang, Qing Li, Wenli Zhang

Abstract

Effectively representing medical concepts and patients is important for healthcare analytical applications. Representing medical concepts for healthcare analytical tasks requires incorporating medical domain knowledge and prior information from patient description data. Current methods, such as feature engineering and mapping medical concepts to standardized terminologies, have limitations in capturing the dynamic patterns from patient description data. Other embedding-based methods have difficulties in incorporating important medical domain knowledge and often require a large amount of training data, which may not be feasible for most healthcare systems. Our proposed framework, MD-Manifold, introduces a novel approach to medical concept and patient representation. It includes a new data augmentation approach, concept distance metric, and patient-patient network to incorporate crucial medical domain knowledge and prior data information. It then adapts manifold learning methods to generate medical concept-level representations that accurately reflect medical knowledge and patient-level representations that clearly identify heterogeneous patient cohorts. MD-Manifold also outperforms other state-of-the-art techniques in various downstream healthcare analytical tasks. Our work has significant implications in information systems research in representation learning, knowledge-driven machine learning, and using design science as middle-ground frameworks for downstream explorative and predictive analyses. Practically, MD-Manifold has the potential to create effective and generalizable representations of medical concepts and patients by incorporating medical domain knowledge and prior data information. It enables deeper insights into medical data and facilitates the development of new analytical applications for better healthcare outcomes.

Introduction

In healthcare analytics, the representation of data—transforming clinical codes into mathematical vectors—is a critical foundation for downstream tasks. Historically, researchers have relied on two primary paradigms: manual feature engineering (e.g., the Charlson Comorbidity Index) and deep learning embeddings (e.g., Word2Vec, Med2Vec). While feature engineering is interpretable, it often fails to capture complex, non-linear disease interactions. Conversely, deep learning models require massive datasets and frequently ignore established medical hierarchies like ICD-9.

Comparison of MD-Manifold with existing medical representation methods.

MD-Manifold addresses these limitations by bridging the gap between expert medical hierarchies and raw patient data. As shown in Figure 1, existing methods often struggle with generalizability or the inability to incorporate domain knowledge effectively. MD-Manifold is designed to respect medical logic while learning from real-world patient co-occurrence patterns, requiring significantly less training data than supervised deep learning approaches.

Methodology

The MD-Manifold framework utilizes a multi-step pipeline to generate robust representations:

  • Knowledge-Driven Augmentation: Patient records are augmented by adding "ancestor" codes from medical hierarchies. For instance, a specific pneumonia code is supplemented with broader categories like "respiratory system," allowing the model to "borrow" statistical strength for rare conditions.
  • Co-occurrence Mapping and Random Projection: A high-dimensional matrix of augmented concept co-occurrences is constructed. To maintain computational efficiency, Random Projection (based on the Johnson-Lindenstrauss lemma) is used to reduce dimensionality while preserving inter-concept distances.
  • Medical-Distance Metric ($CD_{new}$): A novel distance formula is introduced to normalize co-occurrence counts. This prevents highly prevalent conditions (e.g., hypertension) from overshadowing rare but clinically significant relationships.
  • Manifold Learning: A patient-patient network is constructed where nodes are connected based on their $k$-nearest neighbors in the medical distance space. Manifold learning techniques, such as Isomap, are then applied to uncover the underlying topology of patient health states.

Experimental Evaluation

The framework was evaluated using the MIMIC-III database, focusing on general performance, common diagnosis cohorts, and rare disease subsets (e.g., Lymphoma and Hepatic Failure).

Key Results

  • Medical Logic Alignment: In code referencing tasks, MD-Manifold achieved a 6% improvement in NDCG@100 over Skip-gram and Med2Vec, indicating better alignment with clinical knowledge.
  • Enhanced Patient Clustering: Visualizations demonstrated that MD-Manifold produces distinct, tight clusters for patient cohorts (e.g., Sepsis vs. Kidney Failure), whereas traditional embeddings often resulted in overlapping data points.
  • Predictive Performance: For ICU mortality prediction, the framework yielded a 9% increase in AUC over feature engineering and outperformed state-of-the-art embedding methods by up to 1.8%.
  • Data Efficiency: MD-Manifold showed a significant "jumpstart" in transfer learning for rare diseases, achieving high accuracy with far fewer records than competing models.

Discussion and Conclusion

MD-Manifold represents a significant advancement in "knowledge-driven" machine learning. By integrating medical hierarchies with manifold learning, it captures both the "gold standard" of clinical logic and the nuances of real-world data.

A key insight from this research is the necessity of normalization; simple co-occurrence counts are insufficient for capturing specific clinical patterns. Furthermore, the generative nature of manifold learning allows the model to understand data distributions more effectively than discriminative deep learning models when data is scarce.

Future work will focus on expanding the model to longitudinal data to track patient health trajectories over time and incorporating unstructured clinical notes via NLP to further enrich medical concept representations.

Summary updated

Abstract

Effectively representing medical concepts and patients is important for healthcare analytical applications. Representing medical concepts for healthcare analytical tasks requires incorporating medical domain knowledge and prior information from patient description data. Current methods, such as feature engineering and mapping medical concepts to standardized terminologies, have limitations in capturing the dynamic patterns from patient description data. Other embedding-based methods have difficulties in incorporating important medical domain knowledge and often require a large amount of training data, which may not be feasible for most healthcare systems. Our proposed framework, MD-Manifold, introduces a novel approach to medical concept and patient representation. It includes a new data augmentation approach, concept distance metric, and patient-patient network to incorporate crucial medical domain knowledge and prior data information. It then adapts manifold learning methods to generate medical concept-level representations that accurately reflect medical knowledge and patient-level representations that clearly identify heterogeneous patient cohorts. MD-Manifold also outperforms other state-of-the-art techniques in various downstream healthcare analytical tasks. Our work has significant implications in information systems research in representation learning, knowledge-driven machine learning, and using design science as middle-ground frameworks for downstream explorative and predictive analyses. Practically, MD-Manifold has the potential to create effective and generalizable representations of medical concepts and patients by incorporating medical domain knowledge and prior data information. It enables deeper insights into medical data and facilitates the development of new analytical applications for better healthcare outcomes.
BibTeX
@misc{wang2023md,
  author = {Shaodong Wang and Qing Li and Wenli Zhang},
  title = {{MD-Manifold: A Medical-Distance-Based Representation Learning Approach for Medical Concept and Patient Representation}},
  year = {2023},
  eprint = {2305.00553},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.00553},
}
Figure