Application of Bayesian network and regression method in treatment cost prediction

Li-Li Tong, Jin-Bo Gu, Jing-jiao Li, Guang-Xuan Liu, Shuowei Jin, Aiyun YanBMC Medical Informatics Decis. Mak 2021 · DOI

TL;DR

This paper introduces a novel hybrid approach combining Bayesian networks and locally weighted LASSO regression to predict patient treatment costs, particularly effective in scenarios with limited data. By first classifying patients into treatment schemes using Bayesian networks and then applying regression within these categories, the method achieves significantly higher prediction accuracy than traditional methods.

What Problem Does This Paper Address?

Rising healthcare expenditures present a significant challenge, necessitating better methods for predicting treatment costs. Accurate cost prediction is crucial for reforming medical insurance mechanisms, optimizing resource allocation, and reducing patient financial burdens. Existing methods often struggle with small datasets or fail to capture the complex interplay between patient conditions, treatment choices, and cost drivers. Traditional regression models, while useful, can be overly sensitive to noise and may not fully account for the causal relationships inherent in disease progression and treatment. This paper aims to bridge this gap by developing a more robust and accurate prediction model for treatment costs, especially when dealing with the common issue of limited patient data.

Methods

The authors propose a two-stage approach:

  1. Bayesian Network for Classification: A Bayesian network is employed to classify patients into distinct treatment schemes. This leverages the network's ability to model probabilistic dependencies between various patient features (age, gender, medical history, symptoms, etc.) and their likely treatment paths. This classification step is crucial for segmenting the problem into more manageable sub-problems.
  2. Locally Weighted LASSO Regression for Prediction: Within each classified treatment scheme, a locally weighted LASSO regression model is applied. LASSO regression is chosen for its ability to perform feature selection and regularization, which helps in handling multicollinearity and reducing model complexity, especially with a limited number of variables. The "locally weighted" aspect allows the model to give more importance to data points closer to the test point, enhancing accuracy.

Data Preprocessing: Before modeling, significant effort is dedicated to data preprocessing:

  • Text Variable Digitization: Descriptive text data from electronic medical records (EMRs) is converted into numerical formats using a method that considers both qualitative and degree-based characteristics, guided by expert input.
  • Missing Data Imputation: A KNN-based proximity method with weighted distance is used to impute missing values, aiming to reduce interpolation errors compared to global average methods, especially for smaller datasets. Variables with over 20% missing data are excluded.

Datasets

The study utilizes a dataset of 240 colon cancer patients from the Department of Enterology at a hospital in Shenyang, China, collected in March 2016. This dataset was used for both training and validation of the proposed model.

Key Insights and Ideas

  • The Power of Hybridization: Combining a probabilistic graphical model (Bayesian network) for classification with a regression technique (locally weighted LASSO) for prediction proves more effective than using either method in isolation.
  • Classification as a Precursor to Regression: Segmenting patients into treatment categories before applying regression significantly improves prediction accuracy. This acknowledges that cost drivers can vary substantially across different treatment pathways.
  • Handling Small Data with Weighted Methods: The proposed KNN-based imputation and locally weighted regression are specifically designed to perform well even with limited data, a common challenge in healthcare analytics.
  • Feature Importance in Bayesian Networks: The Bayesian network structure reveals interesting dependencies, such as gender being an important influencing variable for subsequent treatment decisions, and hypertension showing a tendency towards diabetes.

Key Results

The proposed Bayesian network and locally weighted LASSO regression fusion model achieved a prediction accuracy of 89.14%. This represents a significant improvement over:

  • Traditional LASSO regression model: 23.36% higher accuracy.
  • LASSO regression model without prior classification: 3.49% higher accuracy.
  • Linear regression model: Low accuracy due to complexity and poor generalization.
  • Neural network model: Better than linear regression but lower than the proposed hybrid model.

The model also demonstrated a lower Mean Squared Error (MSE) and a higher R-squared value compared to other benchmark models.

Limitations and Future Work

  • Data Limitations: The study acknowledges that the amount of data used was relatively small, which is precisely where their method is designed to excel. However, larger and more diverse datasets could further validate and refine the model.
  • Classification Accuracy: While the classification step is beneficial, errors in classification can still impact the final prediction accuracy. Future work could focus on further improving the classification accuracy or developing methods to mitigate the impact of misclassification.
  • Generalizability: The model was validated on colon cancer patients. Further research is needed to assess its generalizability to other diseases and patient populations.
  • Interpretability: While the Bayesian network offers some interpretability regarding feature dependencies, further work could enhance the explainability of the entire hybrid model.

Summary updated

BibTeX
@article{tong2021application,
  author = {Li-Li Tong and Jin-Bo Gu and Jing-jiao Li and Guang-Xuan Liu and Shuowei Jin and Aiyun Yan},
  title = {{Application of Bayesian network and regression method in treatment cost prediction}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2021},
  doi = {10.1186/s12911-021-01647-y},
}
Figure