TL;DR
This paper introduces a novel data-driven framework to identify temporal patterns in electronic health records (EHRs) that predict severe maternal morbidity (SMM). By analyzing sequences of diagnoses, they discovered that certain combinations of antepartum conditions and lifestyle factors significantly increase the risk of SMM, offering a more nuanced understanding beyond traditional risk factor analysis.
What Problem Does This Paper Address?
Severe maternal morbidity (SMM) is a critical public health issue, encompassing life-threatening complications during pregnancy and childbirth. While existing research has identified demographic and medical factors associated with SMM, these studies often treat risk factors as independent and static. This overlooks the dynamic nature of health, where conditions can evolve over time. The paper addresses the gap in understanding how SMM develops through a progression of health states, particularly during the antepartum period, aiming to identify temporal patterns that can serve as early warning signs for intervention.
Methods
The authors developed a three-component data-driven framework:
- Data Mining for Temporal Relations: A sliding window-based algorithm was used to build temporal relationships between diagnosis codes (ICD-9) extracted from patient encounters. This algorithm quantifies the strength of temporal association between pairs of codes based on their proximity within a code sequence.
- Statistical Model for Temporal Patterns: Statistical matching and Chi-square tests with Bonferroni correction were employed to identify significant temporal patterns from the learned temporal relations. A temporal pattern was defined as a significant sequential relationship between diagnoses.
- Regression Model for SMM Association: Logistic regression was used to measure the association between identified temporal patterns and the occurrence of SMM during delivery hospitalization. Age, race, and EHR length were included as confounders.
Datasets
The study utilized data from 6,184 obstetric patients who had delivery encounters at Vanderbilt University Medical Center between 2015 and 2017. The dataset included demographic information, antepartum and delivery encounter data, and ICD-9 codes. SMM cases were identified using 25 SMM indicators defined by the Centers for Disease Control and Prevention (CDC).
Key Insights and Ideas
- Beyond Static Risk Factors: The core idea is that SMM is not just a result of isolated risk factors but a progression along a path. Understanding these temporal sequences is crucial for early detection and intervention.
- EHRs as a Rich Source: Electronic health records, often underutilized for complex temporal analysis, can be mined to reveal these hidden progression paths.
- Antepartum Period is Key: Focusing on the antepartum period allows for the identification of risk factors and patterns before delivery, offering a window for preventative measures.
- Drug Dependence and Mental Health as Major Drivers: The study highlights drug dependence and mental disorders as particularly strong predictors of SMM, suggesting these areas require focused attention.
Key Results
- The framework identified 69 temporal patterns from the antepartum period.
- Of these, 12 patterns were confirmed to be significantly associated with SMM, representing 17.4% of the discovered patterns.
- Specific temporal patterns, such as those involving drug dependence and mental disorders, showed a significantly increased risk of SMM (e.g., greater than 8-fold risk).
- Other significant patterns included combinations of hypertension and diabetes, obesity, advanced maternal age, viral diseases, and tobacco use disorder.
- The study also categorized temporal patterns into subtypes like fetus abnormality and complicating pregnancy, providing detailed insights into specific condition progressions.
Limitations and Future Work
The authors acknowledge several limitations:
- Small Sample Size for SMM Instances: The relatively small number of SMM cases might have limited the discovery of all potential temporal patterns and could affect the generalizability of findings.
- Focus on Antepartum Period Only: The study did not consider health information prior to the antepartum period, which could also contribute to SMM risk.
- Exclusion of Other EHR Data: The analysis was limited to diagnosis codes, neglecting other valuable data like medications, vital signs, and lab results, which could offer richer insights.
- Single-Institute Study: The findings are based on data from one institution, suggesting the need for broader studies across multiple centers to establish more generalizable recommendations.
Future work could involve:
- Applying more advanced temporal pattern mining algorithms (e.g., Allen's algebra, graph neural networks).
- Expanding the analysis to include a wider range of EHR data.
- Conducting multi-institutional studies to validate and generalize the findings.
- Investigating temporal patterns beyond the antepartum period.
Summary updated