TL;DR
This paper introduces a novel approach to mining clinical pathways by enhancing Latent Dirichlet Allocation (LDA) with topic assignment constraints and correlation limitations. This improves the quality and interpretability of discovered "clinical goals" (topics), leading to more accurate and comprehensive clinical pathway models.
What Problem Does This Paper Address?
Expert-designed clinical pathways are rigid and struggle to adapt to the variability in patient conditions and hospital practices. While data-driven approaches like Clinical Pathway Mining (CPM) offer more dynamic solutions, existing methods often produce overly complex and difficult-to-interpret models. A promising avenue is topic modeling (specifically LDA) to identify latent "clinical goals" that represent stages or objectives within a patient's treatment. However, raw LDA struggles with clinical data, often assigning the same activity to multiple unrelated topics or assigning a single activity to many different topics, hindering the discovery of meaningful clinical goals. This paper aims to address these limitations by improving the quality of LDA-derived topics for CPM.
Methods
The core of this work lies in enhancing the standard LDA algorithm with two key constraints:
- Topic Assignment Constraint: This constraint ensures that identical clinical activities occurring within the same "clinical day" (a set of activities within a single day of hospitalization) are assigned to the same topic. This is crucial because, in clinical practice, a specific action (e.g., a particular lab test) performed on the same day usually serves the same immediate purpose. This is implemented by modifying the graphical model and the Gibbs sampling procedure.
- Topic Correlation Limitation: This addresses the issue of clinical activities being assigned to too many topics. The method introduces an "informativeness" measure (using Inverse Document Frequency - IDF) for each clinical activity. Activities with higher informativeness are encouraged to correlate with fewer topics. This is achieved by adaptively limiting the set of relevant topics for each activity during the inference process, effectively reducing redundancy and improving topic distinctiveness.
After obtaining these high-quality topics, the paper reuses a previously proposed process mining framework to construct a concise and interpretable topic-based clinical pathway model.
Datasets
The authors evaluated their proposed method on two real-world billing datasets from a city-center hospital. These datasets represent two different diseases:
- Intracerebral Hemorrhage (ICH): An internal disease.
- Inguinal Hernia (IH): A surgical disease.
The datasets consist of trace ID, item name, amount, and occurrence time, which are then processed into a word-count format suitable for LDA.
Key Insights and Ideas
- Clinical Goals as Topics: The fundamental idea that latent topics discovered by LDA can effectively represent "clinical goals" or stages within a patient's treatment journey is a powerful abstraction.
- LDA's Shortcomings in Clinical Data: The paper clearly illustrates how standard LDA can misinterpret clinical data by assigning activities inconsistently or too broadly, leading to low-quality topics that don't reflect clinical reality.
- The Power of Constraints: The introduction of topic assignment and correlation constraints significantly improves the coherence and relevance of discovered topics, making them more interpretable as clinical goals.
- Topic Quality Drives Pathway Quality: The research emphasizes that the quality of the discovered topics directly impacts the quality and interpretability of the final clinical pathway model. Better topics lead to better process models.
- Billing Data as a Proxy: Using billing data, despite its inherent noise and low granularity, is a practical choice due to its widespread availability and ease of extraction, making the approach applicable in real-world hospital settings.
Key Results
The proposed method, CDG-LDA, consistently outperformed the original LDA in several key aspects:
- Topic Quality:
- Coherence: CDG-LDA topics were more coherent, meaning activities within a topic were more closely related to a single clinical goal. For example, in ICH, CDG-LDA topics were more focused on surgery, whereas LDA mixed surgery and nursing care.
- Redundancy: CDG-LDA showed significantly lower redundancy, meaning topics were more distinct and less overlapping, leading to better differentiation of clinical goals.
- Coverage: CDG-LDA demonstrated better coverage by identifying more critical clinical activities, such as "brain CT" for ICH and "IH repair" for inguinal hernia, which LDA missed or ranked lower.
- Process Model: The resulting topic-based clinical pathway models were more concise and interpretable, aligning well with established clinical guidelines for ICH and IH.
Limitations and Future Work
- Informativeness Measure: The authors suggest that finding a more suitable measure for "informativeness" could further improve the ranking adjustment strategy.
- Topic Number Selection: While a trade-off strategy was used to select the optimal topic number (K), this remains a common challenge in topic modeling.
- Generalizability: While tested on two datasets, further validation on a wider range of diseases and hospital settings would be beneficial.
- Manual Labeling: The evaluation of topic quality involved manual labeling by doctors, which, while valuable, is a subjective and time-consuming process. Future work could explore more automated or objective evaluation metrics.
Summary updated