An Improved Bi-LSTM-Based Missing Value Imputation Approach for Pregnancy Examination Data

Xinxi Lu, Lijuan Yuan, Ruifeng Li, Zhihuan Xing, Ning Yao, Yichun YuAlgorithms 2023 · DOI

TL;DR

This paper tackles the critical issue of missing data in pregnancy examination records, which hinders accurate prediction of Hypertensive Disorder in Pregnancy (HDP). They propose an improved Bi-LSTM model that effectively imputes these missing values by leveraging both forward and backward temporal information, leading to significantly improved HDP prediction accuracy.

What Problem Does This Paper Address?

Hypertensive Disorder in Pregnancy (HDP) is a serious obstetric complication that poses risks to both mother and fetus. Early detection and prevention are crucial, and pregnancy examination data offers valuable insights. However, this data is often high-dimensional, collected over time (time-series), and suffers from a significant rate of missing values. Traditional methods struggle to effectively handle these complexities, leading to suboptimal HDP prediction models. The paper aims to address this gap by developing a robust missing value imputation strategy specifically tailored for pregnancy examination data.

Methods

The proposed approach combines traditional machine learning techniques with a deep learning model. The process begins with a series of data preprocessing steps:

  • Feature Relationship Analysis: A Random Forest (RF) model is used to identify the most important features, revealing that pregnogram data is more critical than biochemical data for HDP prediction.
  • Outlier Detection: Statistical methods are employed to identify and handle abnormal data points that could skew the analysis.
  • Data Filtering: Based on missing data patterns and pregnancy week distribution, data from weeks 11-40 are selected, and samples with insufficient examination records are removed.
  • Feature Extraction: Principal Component Analysis (PCA) is used for dimension reduction, reducing the feature set to 32 items while preserving essential information.
  • Dataset Balancing: Due to an imbalance between positive (HDP cases) and negative samples, random undersampling of negative samples is performed to create a balanced dataset for training.

The core of the imputation strategy is an improved Bidirectional Long Short-Term Memory (Bi-LSTM) model. Unlike traditional LSTMs that process data sequentially in one direction, Bi-LSTMs process the data in both forward and backward directions. This allows the model to capture temporal dependencies more comprehensively, which is crucial for time-series data like pregnancy examinations. The paper compares this approach against several other imputation methods, including cubic spline interpolation, KNN filling, and a standard LSTM model.

Datasets

The study utilizes a large dataset of pregnancy examination records from a hospital in China, spanning from 2008 to 2018. This dataset includes records from 120,396 pregnant women, with 7518 diagnosed with HDP (approximately 6.24%). The data covers the entire pregnancy cycle, from week 1 to week 70, and encompasses various types of examinations, including pregnogram data, routine urine data, routine blood data, and biochemical D data. No new benchmark datasets were introduced; the focus was on improving imputation for existing complex medical time-series data.

Key Insights and Ideas

  • The critical role of data preprocessing: The paper strongly emphasizes that effective missing value imputation is not a standalone task but is heavily reliant on thorough data preprocessing, including feature selection, outlier handling, and dataset balancing.
  • Bi-directional temporal context is key: The superiority of Bi-LSTM over unidirectional LSTM highlights the importance of considering both past and future temporal information when imputing missing values in time-series medical data. This is particularly relevant for dynamic health indicators during pregnancy.
  • Pregnogram data holds significant predictive power: The feature importance analysis clearly indicates that direct measurements related to the pregnancy itself (e.g., weight, fundal height, blood pressure) are more indicative of HDP risk than biochemical markers.
  • Imputation directly impacts prediction accuracy: The study demonstrates a tangible improvement in HDP prediction accuracy after applying the proposed imputation method, underscoring the practical significance of addressing missing data.

Key Results

The proposed improved Bi-LSTM-based missing value imputation approach achieved the best performance among the evaluated methods, as measured by the Symmetric Mean Absolute Percentage Error (SMAPE). The SMAPE for the improved Bi-LSTM was 6.569%, which was lower than Cubic Spline Interpolation (8.745%), KNN Filling (6.746%), LSTM Model (8.796%), and ST-MVL (6.734%).

Furthermore, when using the lightgBM disease prediction algorithm, the study showed that using data imputed by the Bi-LSTM model resulted in higher prediction accuracy compared to using the original, unfilled data. The AUC for HDP prediction reached over 0.75 around 25-26 pregnancy weeks and continued to improve, with data from week 25 to week 40 being more critical for model predictions than earlier data.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Inclusion of demographic factors: The current data processing does not consider factors like region and age, which could influence HDP risk and imputation.
  • Combining imputation algorithms: The current imputation method is relatively simple. Future work could explore combining multiple algorithms to handle different types of features more effectively.
  • Automation of data preprocessing: The current data preprocessing pipeline is largely manual. Developing automated ETL (Extract, Transform, Load) tools would significantly improve efficiency.
  • Integration with treatment programs: Beyond prediction, screening for appropriate drug treatment programs is crucial. Building comprehensive intelligent medical systems that integrate prediction and treatment recommendations is a meaningful direction.
  • Standardization of medical data: The chaotic and non-uniform nature of medical data across different institutions poses a significant challenge for research. Creating standardized medical research databases is essential for advancing disease prevention and drug development.

Summary updated

BibTeX
@article{lu2023improved,
  author = {Xinxi Lu and Lijuan Yuan and Ruifeng Li and Zhihuan Xing and Ning Yao and Yichun Yu},
  title = {{An Improved Bi-LSTM-Based Missing Value Imputation Approach for Pregnancy Examination Data}},
  journal = {Algorithms},
  year = {2023},
  doi = {10.3390/a16010012},
}
Figure