TL;DR
This paper introduces a novel framework using process mining and sequential rule mining to analyze complex pharmacotherapy treatment patterns for Major Depressive Disorder (MDD) in veterans. By uncovering common treatment pathways and their associations with outcomes, the research offers operational insights into clinical practice and highlights critical points for increased patient monitoring.
What Problem Does This Paper Address?
Major Depressive Disorder (MDD) is a significant global health burden, and while prescription drug costs have decreased, the overall economic and human cost of MDD remains high. Veterans, particularly those who served in Iraq and Afghanistan, experience higher rates of MDD. Existing reporting systems at the Department of Veterans Affairs (VA) lack the detailed insight needed to guide corrective actions and resource allocation for effective treatment. Understanding the actual, real-world pharmacotherapy treatment patterns—how medications are prescribed, doses adjusted, and sequences of actions taken—is crucial for optimizing care and improving patient outcomes. However, inferring these complex, multi-step treatment pathways from electronic health records (EHRs) is challenging due to data complexity, missing information, and the need to link discrete events into meaningful sequences.
Methods
The authors developed a four-step framework to address this challenge:
- Data Preprocessing: This crucial step normalizes all drug dosages to fluoxetine equivalents, smooths prescribed daily doses using median filtering, and abstracts these into discrete "clinical actions." This allows for the comparison of different antidepressant medications and dosages on a common scale.
- Trace Clustering and Process Inference: Each patient's treatment episode is represented as a "trace" of clinical actions. These traces are then clustered to group individuals with similar treatment pathways. Process mining techniques are applied to each cluster to infer a process model representing the common pathway for that group.
- Sequential Rule Mining: The clinical actions are also stored in a sequence database to discover statistically interesting rules that associate specific treatment sequences with adverse events (like emergency department visits or self-harm).
- Merge Rules with Clinical Context: The discovered association rules are then integrated with their clinical context, using a "bow-tie analysis" to visualize the events surrounding an adverse outcome within a specific timeframe. This helps to understand the temporal relationships and potential triggers.
Datasets
The study utilized pharmacy and administrative data from the Department of Veterans Affairs (VA) Corporate Data Warehouse (CDW) for a cohort of veterans diagnosed with Major Depressive Disorder (MDD) who served during Operation Enduring Freedom (OEF) or Operation Iraqi Freedom (OIF). The data spanned from January 1, 2006, to January 1, 2020. This dataset included outpatient antidepressant pharmacy fills, dispensed inpatient antidepressant medications, emergency department visits, self-harm incidents, and all-cause mortality.
Key Insights and Ideas
- Dominance of a Few Patterns: A surprisingly small number of pharmacotherapy patterns account for a large majority of MDD treatment sequences. This suggests that while there's variability, clinical practice tends to converge on a few common approaches.
- Suboptimal Dosing is Common: A significant portion of individuals start antidepressant treatment at suboptimal doses, which is a concern for treatment efficacy.
- Temporal Associations Matter: The framework highlights that the timing of dosage changes and the duration of treatment at specific dose ranges are strongly associated with adverse events. This emphasizes the importance of monitoring during critical treatment phases.
- Context is Key for Interpretation: Simply identifying associations between drug use and outcomes is insufficient. Visualizing these associations within their clinical context (e.g., what happened in the weeks leading up to an adverse event) is essential for generating actionable insights.
Key Results
- The study analyzed a cohort of 252,179 veterans with MDD.
- The top ten prescription patterns accounted for 69.3% of the data for individuals starting antidepressants at a fluoxetine equivalent of 20-39 mg.
- The framework achieved an average replayability score of 0.891, indicating that the discovered process models could represent nearly 90% of the patient treatment traces.
- Sequential rule mining identified associations between dosage levels and timeframes with outcomes like emergency department visits and self-harm. For instance, rules like "80+mg for 0-2 weeks ⇒ self-harm" with an odds ratio of 3.70 were found, suggesting a heightened risk.
- The study found that 55.0% of the OEF/OIF MDD cohort took SSRIs or SNRIs at some point during their depressive episode.
Limitations and Future Work
The authors acknowledge several limitations:
- Observational Data Bias: The study relies on observational data, which is prone to biases like "indication bias" (where outcomes are also indicators of the condition being treated or side effects).
- Unaccounted Factors: The study did not account for depression severity, pre-existing conditions, or actual medication consumption (e.g., stockpiling).
- Imprecise Calculations: The low incidence of adverse events, particularly mortality, made some calculations imprecise.
Despite these limitations, the authors propose that their framework is valuable for documenting and exploring clinical practice, providing a basis for future medication management efforts and informing drug monitoring. Future work could involve performing causal analysis to establish definitive links between prescription patterns and outcomes.
Summary updated